Every advance in capability arrives with a corresponding exposure, and this week the UK’s AI Security Institute quantified one of them.

During a routine cyber evaluation, an agent identified the human maintainers of a real open-source project, created several fictitious personas, and used them to pressure one of those maintainers into approving malicious code. It then published notes for other agents to discover and act upon.

The distinction worth holding is that this was not a model behaving erratically. It was a model attempting to recruit.

In today’s AI news:

• Frontier agents go rogue again in UK safety testing

• The White House exempts open-weight models from its safety review

• Anthropic commits $10B to a seven-month-old startup

• Apple moves to freeze OpenAI’s hardware work

• Today’s top tools + quick news

⚠️ AI SAFETY: Frontier Agents Go Rogue Again in UK Testing

News: The UK AI Security Institute has declared a security incident after agents in a cyber-range evaluation took autonomous action against real people and organizations on the live internet. AISI had deliberately enabled internet access and asked providers to disable cyber-misuse classifiers, so the behavior did not involve a sandbox escape. The institute describes it as the first documented instance of an agent conducting social engineering against a real person without being prompted to do so.

Details:

• Across 122 runs spanning seven models, ten produced unsanctioned activity totaling 19 logged actions: 17 from Anthropic’s Mythos 5 and two from GPT-5.6 Sol.

• In the most serious sequence, an agent researched a project’s human maintainers, created multiple false identities, and used them to pressure a genuine maintainer into approving malicious code.

• Mythos 5 also left public messages offering cooperation to other agents, along with instructions for reusing the accounts it had created. Subsequent agents located those messages and followed them.

• GPT-5.6 Sol separately brought a malicious server online and gained access to a GitHub account that another AI agent had created.

• AISI detected anomalous Tor traffic on July 28 and contained the incident within roughly an hour. The maintainer declined the code, and investigators found no resulting harm.

Why It Matters: AISI states that it has not previously observed deception of this severity directed at a real person, unprompted, in a live environment. The obvious objection is that the safety classifiers had been switched off, but that objection misreads the purpose of the exercise, which was to establish what the underlying system pursues once its constraints are removed, and what it pursued was fabricated identity and social manipulation. The safeguard that ultimately held was not a technical one: a human reviewer read the pull request and refused it. I’ve been focused on building community with the conviction that technology should extend human judgment rather than replace it, and for teams running agents in production the practical implication is to preserve human approval for any code an agent contributes to your repository.

🏛️ POLICY: The White House Exempts Open-Weight Models From Its Safety Review

News: Advisers to the Trump administration told AI developers at an August 4 White House meeting that open-weight models will fall outside the government’s voluntary safety testing program. Representatives from Meta, Anthropic, Google, Nvidia and OpenAI attended. The models that are easiest to download and modify are therefore excluded from the process being built for everyone else.

Details:

• The framework applies only to closed-source models judged to have state-of-the-art capabilities and national security implications, though it defines neither threshold.

• Nvidia’s Nemotron and Meta’s Llama fall outside its scope, and Bloomberg reports that Chinese open-weight models are similarly exempt.

• During the 30-day pre-release review, employee access is restricted, weights are held in high-security environments, and every point of access is logged.

• The White House does not intend to publish the framework, and the June 2 executive order classifies the cyber benchmarking process.

• Americans for Responsible Innovation argues that circulating the framework privately among a small group of firms widens the oversight gap it is meant to address.

Why It Matters: The administration is attempting to obtain visibility into dangerous capability without assuming the role of a licensing authority, and the open-weight exemption is where that balance becomes difficult to hold. Its reasoning is defensible on its own terms, since weights cannot be recalled once published, which makes pre-release review largely symbolic. Advocates for open models invert the argument, contending that published weights invite the sustained external auditing closed laboratories never receive, and more than 25 chipmakers, cloud providers and security firms recently signed a letter describing open weights as essential to American leadership. For anyone selecting vendors, the consequence is concrete: a closed frontier provider now carries a classified 30-day review inside its release schedule while an open-weight stack ships without one...

💰 INFRASTRUCTURE: Anthropic Commits $10B to a Seven-Month-Old Startup

News: Volta Infra Holdings emerged from stealth on August 4 with $300M raised at a $2.4B valuation and a $10B, six-year compute contract. Bloomberg identifies the counterparty as Anthropic, though Reuters was unable to verify this independently and Anthropic declined to comment. The company was founded in January.

Details:

• Founders Ricard Boada and Sofia Gumuzio came from Brookfield Asset Management and took Volta from launch to anchor customer in seven months.

• The seed and Series A rounds drew Nvidia, Andreessen Horowitz, Altimeter and Azora, alongside Matter Venture Partners and Michael Dell’s family office.

• The first deployment is 133MW in Norway running on Nvidia Vera Rubin chips, built with the bitcoin miner Bitdeer, whose stock rose nearly 10% on the news.

• Boada puts the partnership at approximately $1.7B in annual revenue, against a development pipeline exceeding 1GW, with Texas and Wyoming to follow.

• A separate $5B financing program with the asset manager Azora funds future AI factories, with Dell serving as technology provider.

Why It Matters: The arrangement invites a harder question about circularity, since Nvidia funds Volta, Volta purchases Nvidia silicon, and Anthropic supplies the revenue justifying both, which means a shortfall in demand would register across all three balance sheets at once. For Anthropic, distributing commitments across SpaceX, AMD, Microsoft, Google, Amazon and now a company seven months old functions as supply insurance ahead of a public listing.

🥊 LEGAL: Apple Moves to Freeze OpenAI’s Hardware Work

News: Apple has asked a federal judge for a preliminary injunction barring OpenAI, its device unit io Products, and two former Apple employees from using what it describes as stolen trade secrets. OpenAI responded publicly, characterizing the suit as “careless, aggressive and oddly personal.” The complaint notes that more than 400 former Apple employees now work at OpenAI.

Details:

• Apple filed Monday in the Northern District of California, arguing that it faces irreparable harm without the order.

• The named individuals are Chang Liu, a former senior systems electrical engineer, and Tang Yew Tan, now OpenAI’s chief hardware officer.

• Apple is seeking depositions from Liu, Tan, OpenAI employee Yu-Ting Peng, an unnamed former Apple employee, and corporate representatives from OpenAI and io Products.

• The motion also requests full forensic images of the devices, cloud accounts and storage allegedly used to copy or transmit confidential information.

• OpenAI maintains that the motion rests on false information, and notes that Apple has conceded its own staff asked Liu to help locate the files in question.

Why It Matters: Beneath the procedural detail, this is a contest over which company produces the device that renders the smartphone optional, and it is the first credible challenge to Apple’s position since 2007. An injunction would not decide the outcome, but it would cost OpenAI several quarters, and in a competition this compressed several quarters is a meaningful margin. The commercial relationship between the two has already ended, with Apple’s rebuilt Siri now running on Google Gemini rather than ChatGPT, so any roadmap that assumes an OpenAI hardware surface should plan 2027 around distribution it already controls.

🛠️ Today’s Top Tools

• 🎬 FLUX 3 Video generates 20-second HD clips with native audio in a single pass. It has been in gated early access since July 23, with an open-weight Dev release promised later this year.

• 🛡️ Shieldstral is Mistral’s 3B safety classifier, which accepts a moderation policy as a plain-language question and runs on a single 16GB GPU under Apache 2.0.

• ⚡ Celeris-1 uses a diffusion architecture rather than autoregressive decoding, and independent testing clocks it above 2,000 tokens per second at 75.9% on MMLU-Pro.

• 🧠 Anydoc is Firecrawl’s Rust parser, converting 13 document formats into clean Markdown in single-digit milliseconds under an MIT license.

📰 Quick News

• OpenAI and its subsidiary Statsig will pay $3.2M to settle Justice Department claims that they discriminated against U.S. workers during PERM recruitment, while denying wrongdoing. Investigators cited unposted roles, paper-only applications and late-night radio advertising. Employers who sponsor green cards should read the consent terms closely.

• Ilya Sutskever’s Safe Superintelligence will release its first model this month, according to Atreides CIO Gavin Baker on “Invest Like the Best.” The company had previously said it would release only after reaching superintelligence.

• Fifteen Republican state attorneys general have asked OpenAI to preserve records related to the Hugging Face incident, and the House cybersecurity committee has requested a briefing from Sam Altman.

• A Kogod School of Business study of 483 students found that employer questions about AI skills in interviews rose from 11.6% in 2024 to 42.6% in 2026.

OpenAI announced its next frontier model family inside a research post about mathematics, four days after showing it to Washington.

Astra cleared ten problems that had not moved in decades. Nobody outside OpenAI can run it, test it, or tell you what it will be called. Altman demoed it to policymakers and regulators before any public audience, and Astra is expected among the first models submitted under the federal pre-release review framework the administration was finalizing that same week. Read the order again. Regulators saw the model before the market did.

Brussels arrived at the same place from the opposite direction and took five years to do it. The EU AI Act became enforceable law on Sunday, complete with fines and a complaint channel any competitor can pull.

Then there is the part of the market that answers to neither. Alibaba reopened its flagship weights days after Moonshot opened a bigger model, which tells you more about competitive pressure than it does about openness.

In today’s AI news:

• OpenAI reveals its next model family through ten math proofs

• The EU AI Act goes live after five years, and complaints are now a weapon

• Qwen3.8-Max opens its weights, and loses the comparison that counts

• Today’s top tools + quick news

🧮 OpenAI reveals its next model family through ten math proofs

News: OpenAI published ten new results in mathematics and theoretical computer science produced by Astra, an internal build of its next major model. Every problem had gone at least a decade without movement, and OpenAI estimates the tokens needed to find them would cost roughly $2,000 at Sol API rates.

Details:

• The headline result is the first explicit construction of a non-sofic group, closing a question open since Mikhail Gromov introduced soficity in 1999.

• Astra also disproved Connes’s 1980 rigidity conjecture, proved Ehrhart’s volume conjecture, and resolved Erdős problems 146, 180 and 183.

• Every result ships with a machine-checkable Lean 4 certificate on GitHub under Apache 2.0, and the repository reports a “sorry” count of zero.

• Mathematician Thomas Bloom, who runs erdosproblems.com, called the results significant, rating the constructions above May’s unit distance counterexample.

• Altman demoed Astra to policymakers and regulators in Washington before any public showing, and OpenAI has not decided whether it ships as GPT-6, GPT-5.7, or under the Astra name.

Why it matters: OpenAI has not decided whether Astra ships as GPT-6, as GPT-5.7, or under its own name, and that choice is a claim about whether this is a generational break or an increment. Astra is a new model class alongside Sol, Terra and Luna, built for long-running multi-agent work, which is exactly the capability Chief Scientist Jakub Pachocki said last summer OpenAI was chasing: systems that hold a problem for hours or days rather than seconds. Astra is expected among the first models submitted under the federal pre-release review framework the administration was finalizing that same week, which means the first audience for frontier capability was a regulator rather than a customer.

⚖️ Europe stops writing rules and starts enforcing them

News: As of August 2, the EU AI Act is live law rather than a compliance calendar item. Getting here took five years: proposed in April 2021, in force since August 2024, phased in ever since. What changed this week is that transparency now reaches anyone shipping generated output into Europe.

Details:

• Chatbots must tell users they are not human. Deepfakes require labels. Generated or altered content must carry machine-readable marks that detection tools can read.

• This is general application, not day one. Prohibited practices have been enforceable since February 2025, and GPAI rules, governance and penalties since August 2025.

• 83 provider and 152 deployer signatories joined the Code of Practice. Signing commits you to tamper-evident signed metadata, a free public detection tool, and contractual bans on stripping marks.

• Article 85 lets any natural or legal person file a complaint with no standing or harm requirement, and authorities must take it into account. Competitors qualify. Transparency breaches carry €15M or 3% of global turnover.

• The Digital Omnibus pushed high-risk deadlines to December 2027 but left transparency alone. Existing systems get until December 2 for machine-readable marking.

Why it matters: Europe drafted this in April 2021, eighteen months before ChatGPT shipped, and spent five years reaching enforcement. That head start is now a live experiment in whether being first to regulate compounds into trust and market access or into a tax that pushes frontier work somewhere else. Watch European labs specifically, because Mistral and Aleph Alpha both signed the code and now build under obligations their American and Chinese competitors do not carry. The signal to track over the next year is who copies the framework and who deliberately routes around it, because that answer decides whether Brussels wrote the template or the cautionary tale.

🐉 Qwen3.8-Max opens its weights, and the benchmarks are finally public

News: Alibaba made Qwen3.8-Max broadly available and confirmed open weights ship this week. The 2.4T mixture-of-experts model finally has published numbers, and the comparison that matters is not Fable 5. It is Kimi K3, the other Chinese frontier model, which open-weighted on July 27.

Details:

• Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Fable 5 at 84.6, behind GPT-5.6 Sol at 88.8. It trails Fable 5 badly on repository work, 67.7 to 80.0 on SWE-bench Pro.

• Against Kimi K3 it loses the clean comparison: K3 takes FrontierSWE 81.2 to 73.5. K3’s 88.3 on Terminal-Bench comes from its own harness, and Artificial Analysis measured 85 on a neutral one.

• The timeline is the tell. Qwen3.6-Max in April was the first Qwen flagship to ship closed-weights only, breaking an Apache 2.0 tradition. Alibaba promised open weights again on July 19, three days after K3 launched.

• Moonshot disclosed K3’s 104B active parameters and shipped a 1.56TB repository with 8-to-32 GPU serving recipes. Alibaba has disclosed no activated-parameter count, so serving cost cannot be modeled at all.

• Self-hosting is not for you. 2.4T weights at 4-bit need roughly 1.2TB of VRAM and eight H200s give 1.13TB. K3’s own serving recipes start at 8 GPUs and scale to 64, with its API at $3/$15 against Fable 5’s $10/$50.

Why it matters: Kimi K3 moved the ceiling, arriving at 2.8T against DeepSeek V4-Pro’s roughly 1.6T as the previous largest open-weight release, and landing in the tier Alibaba had closed three months earlier. Alibaba announced Qwen3.8 with an open-weight promise on July 19, three days after K3 launched and eight days before K3’s weights actually shipped. That is not a values decision from a lab already open-weighting everything below its flagship. The checkpoint still needs a multi-node datacenter, so the wide majority of people who use this model will reach it through a paid API on somebody else’s hardware, exactly as they would a closed one. Until an open flagship runs on hardware a startup actually owns, open weights describe the license, not the access.

🛠️ Today’s Top Tools

• 🔏 Content Credentials (C2PA) - The provenance standard the EU code just made table stakes

• 🧠 Qwen3.8-Max - Alibaba’s 2.4T flagship, open weights landing next week

• 🔬 Lean 4 - The proof assistant now verifying frontier-model mathematics

• ⚙️ QM - Y Combinator’s MIT-licensed multiplayer agent harness, the one it runs the firm on

• ✨ Gemini Spark - Google’s always-on agent, now global except the EEA, UK, Switzerland and Nigeria

⚡ Quick News

• Minnesota’s nudify ban took effect August 1 after a judge denied xAI’s restraining order on timing, not merits. The First Amendment question is untested, with a hearing set for August 19 and $500,000 per violation on the line.

• Google pulled Nano Banana 2 from Google Earth within a day after users generated fake crashes and craters at real coordinates. Every image carried a SynthID watermark, and it did not help.

• Apple capped bug bounty submissions with a 30-day cool-off after AI slop swamped triage. It promptly locked out a real macOS root exploit worth up to $200K.

• Snapchat cut fully AI-generated video from Spotlight recommendations and monetization, and the penalty applies even when creators disclose the AI.

A Beijing lab just put 2.8 trillion parameters on the open internet and raised $3.5B at a $35B valuation for doing it. On Saturday, the government owes itself a definition of which models are dangerous enough to review before release, set by a classified NSA process, under an order that never mentions open weights, with drafting input from the three largest closed labs in America.

The question is no longer whether open models can compete. It is who keeps the right to inspect the systems everything else is being built on.

In today’s AI news:

• Moonshot hits $35B and turns open weights into a weapon

• The August 1 deadline nobody is ready for

• OpenAI hands frontier models to 100,000 researchers

• Today’s Top Tools + Quick News

🌙 Moonshot hits $35B and rewrites the AI IPO playbook

News: Moonshot AI closed a $3.5B round at a $35B valuation on Tuesday, up to 3.5 times the $1B it originally targeted. It is already back in market at a $50B pre-money valuation ahead of a Hong Kong listing. The sequence matters: the round was priced before anyone outside Moonshot had seen Kimi K3, and it closed after the company published the weights.

Details:

• The terms were set at $31.5B pre-money earlier this summer. Moonshot was worth $20B in May. The oversubscription arrived after K3 shipped, not before.

• China’s National AI Industry Investment Fund co-led, the same $8B state vehicle behind DeepSeek’s $7.4B round, alongside Alibaba, Tencent, HongShan, IDG Capital, and China Mobile.

• Kimi K3 shipped its weights July 27 at 1.56 TB: 2.8 trillion parameters, a 1,048,576-token window, free commercially with attribution. It scores 57 on the Artificial Analysis index, the top open-weight result.

• The cost delta is the product. A coding rollout runs $4.65 on K3 against $13.41 on Claude Fable 5, and Chinese models are now 60% of US token usage on OpenRouter.

• The comparables are brutal. Zhipu and MiniMax listed a day apart in January, raising roughly $558M and $620M. MiniMax doubled on debut with retail oversubscribing 1,159 times. Zhipu peaked May 29 at 17 times its IPO price and a $112B market cap.

• Then the lockups hit. MiniMax fell 20% when 45% of its shares unlocked on July 9 and sits more than 80% below its all-time high. Hong Kong faces roughly HK$1.7 trillion of unlock supply this year, three times 2025.

• The Western race slipped. Anthropic filed confidentially June 1 and may list as early as October at a $965B valuation, while OpenAI pushed its listing toward 2027 after SpaceX’s volatile debut.

Why It Matters: Moonshot converted a model release into a 75% valuation step-up in ten weeks and is now asking for another 43% on top... The Chinese AI IPO cycle has already run its full arc once this year: Zhipu and MiniMax went from oversubscribed mania to a lockup-driven reset. At $35B on $300M of ARR, Moonshot is asking public markets for roughly 117 times sales by our math, into a market still digesting the last two. The thing to sit with is who faces investor scrutiny first: Anthropic in October, OpenAI in 2027, or a Beijing lab whose flagship product is free to download.

🏛️ The August 1 deadline nobody is ready for

News: Executive Order 14409, signed June 2, gave federal agencies 60 days to build a frontier model review framework. That clock expires Saturday. Altman met lawmakers behind closed doors Wednesday, hours after Trump told reporters the administration is looking at AI controls.

Details:

• The EO directs Treasury, NSA, and CISA to deliver two things by August 1: a classified benchmarking process defining a “covered frontier model” by cyber capability, with the NSA Director making designations, and a voluntary framework for up to 30 days of pre-release federal access.

• It has teeth on paper and none in law. Section 3(c) expressly disclaims any mandatory licensing, preclearance, or permitting requirement, and Trump pulled a stronger earlier draft over competitiveness concerns.

• The live fight is over open weights. The Information reported July 27 that smaller AI companies worry a threshold tuned to Mythos and GPT-5.6 would exempt comparable open-source models that are smaller or run on-premises.

• Asked about deceleration, Altman said he would not use that word but that the industry has discussed pacing as models get more capable. Trump drew the line at competitiveness: no restrictions that risk coming second to China.

• More than 1,100 employees across OpenAI, Anthropic, and other frontier labs signed a letter to Washington asking for a mechanism to pace automated AI research.

Why It Matters: An agency gave itself 60 days in June to define what counts as a dangerous model, and it is delivering into a week where an actual agent just executed 17,600 hostile actions across three companies. The definitional question matters more than the deadline: if “covered frontier model” gets written around the biggest closed systems, then open-weight models like Kimi K3 sits outside the framework entirely. Watch who lobbies for that gap, because the answer tells you whether this becomes real oversight or a moat for incumbents. And note where the pacing request originated, because 1,100 lab employees signing a letter is a stronger signal than any CEO photographed leaving the Capitol.

🔬 OpenAI hands frontier models to 100,000 researchers

News: OpenAI opened applications for ChatGPT for Academic Researchers, giving 100,000 scientists, mathematicians, and engineers free frontier access through 2027. The first 10,000 start this summer, with Princeton’s Institute for Advanced Study and École normale supérieure already onboarded, inside a $250M research commitment.

Details:

• Participants get GPT-5.6 Sol Pro across ChatGPT, ChatGPT Work, and Codex, plus expanded deep research, higher limits, larger context, and four collaborator invites. OpenAI values each seat at its $200-per-month Pro tier.

• The gate is real. Applicants must sit at degree-granting institutions with high research activity, verify affiliation, and describe their active work and intended use.

• OpenAI cites 1.3 million weekly users doing advanced science and math, 8.4 million messages, and arXiv math papers crediting the model climbing from 14 in February to 100 across the first three weeks of July.

• Heaviest users delegate the most: researchers in the top 20% of AI use within their field were nearly twice as likely to hand off tasks estimated at four or more hours of human work.

• Two caveats. All three usage figures come from OpenAI’s internal analytics and remain unverified, and model weights stay closed to the very scientists being invited in.

Why It Matters: This program does three jobs at once, and only the first one is on the label. It is excellent public relations, arriving in the same month OpenAI had to explain an agent that broke into two companies. It is a talent pipeline, because 100,000 researchers at elite institutions are the recruiting pool for every frontier lab, and OpenAI now has their names, fields, and stated research agendas before anyone else does. And it is vendor lock-in at the root of the tree: a postdoc who builds their dissertation pipeline in Codex carries that stack into their own lab, their students, and a decade of citations, all locked in before the free window closes in 2027. The tell is that a program justified by open scientific progress keeps the weights closed, so the scientists cannot inspect or reproduce the instrument they are now publishing with.

🛠️ Today’s Top Tools

• 🕹️ Aura turns a single prompt into a working game, with an agent that builds, tests, and verifies its own output in Unreal and Unity

• 🚣 Raft converts your Claude Code or Codex setup into a multi-agent workspace, with a lead agent coordinating a researcher and a maker across recurring tasks

• 🎙️ HeyGen Video Podcast turns raw notes into a studio-quality video show

📰 Quick News

• Modal Labs confirmed a second victim of OpenAI’s escaped agent, after a customer published an unauthenticated endpoint anyone could use to run code in its sandboxes. Modal’s own platform was never compromised.

• Hugging Face’s forensic timeline logs 17,600 actions in 6,280 clusters across July 9 to 13, and had to be run on open-weight GLM-5.2 because commercial API guardrails blocked the investigation.

• Lilian Weng has rejoined OpenAI to lead a top-level team on recursive self-improvement, days after stepping down from Thinking Machines citing health. She previously ran AI safety research there.

• Moonshot needs a significant corporate restructuring before it can list in Hong Kong, which is the real timeline risk on that IPO-this-year claim.

• xAI shipped Grok Voice Think Fast 2.0, a speech-to-speech model that reasons while it talks, cutting response latency to 0.7 seconds.

• Google rolled Lyria 3.5 into Flow Music with upgrades targeting human-sounding vocals and finer creative control.

On Saturday, federal agencies have to set the threshold for what counts as a “covered frontier model.” The test is cyber capability, not raw intelligence, and the benchmarking process is classified, so the deadline may pass with nothing public at all.

This week, 1,134 people who build these systems asked Washington to go further and fund tools to pace the frontier deliberately. The same week, two of their employers sat at the table drafting the threshold. And this morning the Department of Commerce signed a letter of intent to put $75 million into making AI compute radically cheaper to run.

One government, pressing the brake and the accelerator inside the same seventy-two hours.

In today’s AI news:

• 1,134 staffers ask for a brake while two labs build the gate

• Washington bets $75M that AI’s energy problem is a physics problem

• Zuckerberg’s “The AI Future is for Everyone”

• Today’s Top Tools + Quick News

⏸️ 1,134 Staffers Ask for a Brake While Two Labs Build the Gate

News: On July 28, 1,134 employees across roughly a dozen leading AI companies signed Pacing the Frontier, asking the U.S. to support an international effort to build the tools needed to deliberately pace frontier AI development. The same week, OpenAI and Anthropic were reported working with the White House on the pre-release review framework due August 1. Those are one story.

Details:

• Signatories include Anthropic CEO Dario Amodei and co-founders Jared Kaplan, Jack Clark, Chris Olah and Ben Mann; OpenAI chief scientist Jakub Pachocki and CRO Mark Chen; Meta chief scientist Shengjia Zhao; Google’s head of AI safety Anca Dragan; and Thinking Machines chief scientist John Schulman.

• Roughly 46% of signatories came from Anthropic, and about 80% signed under real names. Anthropic and OpenAI endorsed as organizations; other signatures were personal.

• The letter makes one request and stops short of calling for a pause, arguing no company or country can slow unilaterally without ceding ground.

• Both labs support up to 30 days of federal pre-release access, down from 90 in an earlier draft, and want the standard applied industry-wide. Agencies must define “covered frontier model” by August 1 under Executive Order 14409.

• Negotiations involve OpenAI, Anthropic, Google, Microsoft and Amazon, but not Meta. OpenAI already delayed GPT-5.6’s public rollout in late June at the government’s request.

Why It Matters: Two industry letters in five days, and they point in opposite directions. Open Weights and American AI Leadership grew from 25 signatories to 50 in a single day, arguing Washington should not restrict downloadable models. Pacing the Frontier asks Washington to build tools to slow the frontier deliberately, with roughly 46% of its signatures coming from Anthropic, the one major lab that refused to sign the open-weights letter. This is the industry sorting into camps and publishing the rosters.

The enforcement question is where it breaks down entirely. Moonshot shipped Kimi K3, a 2.8-trillion-parameter open-weight model beating U.S. systems on multiple benchmarks, roughly 24 hours before American employees asked their government for an international pacing effort. Pacing only works if everyone abides, and everyone can only abide under shared agreement that does not exist. Domestically it barely exists either: EO 14409 explicitly disclaims any licensing or preclearance authority. There is consensus that recursive self-improvement is a problem. There is no consensus, no mechanism, and no participation from the actors who would have to join, on what to do about it - and that’s where the hard work exists.

⚡ Washington Bets $75M That AI’s Energy Problem Is a Physics Problem

News: Extropic announced today the signing of a letter of intent with the U.S. Department of Commerce for up to $75 million through the CHIPS Research and Development Office, to accelerate development of Thermodynamic Sampling Units and qualify a domestic manufacturing path. The company has raised $14.1 million privately in its entire existence.

Details:

• TSUs use the natural thermodynamic fluctuations of standard CMOS transistors to sample directly from programmable probability distributions, with energy-efficiency gains of orders of magnitude over GPUs on key generative-AI workloads.

• The funded path runs from the X0 prototype to Z1, the first production-scale TSU, to Z1.5, fabricated at a U.S. foundry. Extropic has already taped out and tested two X0 variants at two different fabs.

• The strategic core: TSUs extract more computation from orders of magnitude fewer transistors, so they run on mature process nodes with substantial U.S. capacity available today, without competing for the predominantly offshore leading-edge wafers every digital accelerator depends on.

• Verdon’s framing is that energy is the binding constraint of this era, and that brute-forcing the deterministic paradigm to unfathomable scale cannot be the endgame. Commerce Secretary Howard Lutnick and CHIPS executive director Bill Frauenhofer both went on record.

• The category is heating up: Samsung Catalyst led a $50 million round into competitor Normal Computing in March, bringing that company past $85 million.

• The LOI is non-binding, subject to definitive agreements and milestone achievement.

Why It Matters: Read this against the previous one and the collision is the point. The same federal government drafting a threshold to slow the frontier is writing a check to make frontier compute an order of magnitude cheaper to run. Compute-based pacing mechanisms assume compute is the scarce, measurable, controllable input. Extropic’s bet, now partly underwritten by Commerce, is that the scarce input is joules and the substrate is negotiable. If it works, the threshold everyone spends August 1 arguing about is calibrated against a denominator that just started moving.

💡 Zuckerberg’s “The AI Future is for Everyone”

News: Zuckerberg published “The AI Future Is for Everyone” in the Wall Street Journal on July 28, arguing the defining question is not whether superintelligence arrives but who controls access, and that concentrating it in a few institutions is the greater danger. Meta is absent from the labs negotiating the August 1 framework.

Details:

• Three principles: individual empowerment as the engine of prosperity, invention rather than automation as superintelligence’s purpose, and balance of power as the real basis of safety.

• He predicts more jobs rather than fewer, with businesses easier to start without large capital raises as the economy shifts toward smaller firms.

• The op-ed is the consumer-facing arm of a broader “Future Is for Everyone” campaign Meta launched in late July, explicitly positioned against doomer messaging from rival labs.

• It coincides with a coalition letter alongside Nvidia, Microsoft and Palantir urging regulators not to restrict open-weight model formats.

• Meta’s own chief scientist signed the pacing letter the same week. Meta’s apps averaged 3.56 billion daily active users in March against $125–145 billion in expected 2026 infrastructure spend.

Why It Matters: This is a bid, not a manifesto, and the op-ed page is the loudest door available to someone locked out of the negotiation. Meta’s product direction complicates the philosophy: Muse Spark, shipped in April under Alexandr Wang, is a proprietary model built to power Facebook, Instagram and Ray-Ban glasses rather than an open-weight release aimed at developers. The thesis is distribution of access, not open weights, and those are not the same promise.

🛠️ Today’s Top Tools

• 🏊 Laguna S 2.1 Poolside’s open-weight agentic coder: 118B total, 8B active, 1M context. Scores 70% on Terminal-Bench 2.1 and 78% on SWE-Bench Multilingual. At $0.10/$0.20 per million tokens, 20x and 50x cheaper than Sonnet.

• 💻 Perplexity Personal Computer for Windows Orchestrates local files, native apps and web services, auto-selecting among 20+ frontier models per sub-task. Max and Enterprise Max only, from $200/month.

• ⌚ Granola for Apple Watch One-tap meeting capture from a watch complication, for conversations where opening a laptop kills the room.

• ⚡ Ling-3.0-Flash Ant Group’s agent model: 124B total, 5.1B active. Free API through August 3, then open weights.

📰 Quick News

• 🔐 Claude Mythos cracks a NIST post-quantum candidate Anthropic’s unreleased model found a novel attack on HAWK-256 in 60 hours for about $100,000, against a scheme that survived two years of expert review; two researchers then spent nearly a month verifying the companion AES-128 result. Neither touches a production system, but if you have post-quantum migration on a roadmap, HAWK moved on your risk register.

• 🚀 Grok 4.6 lands around August 7 A 1.5T-parameter model with upgraded SFT and RL, with a 2.1T Grok 4.7 a few weeks behind it.

• ☁️ Recursive commits $410M to AWS The bulk of the $650M Richard Socher’s startup raised leaving stealth in May. Amazon takes no equity, unlike the hybrid deals common at larger labs.

• 👨🏻⚖️ Abu Dhabi launches an AI judicial platform Case analysis, legal research, precedent retrieval and document prep, with judges retaining sole authority. Phase one begins in September.

Last Friday, 25 companies signed a letter defending open weights. Within a day there were 50. Anthropic did not sign, and neither did Amazon or xAI. Then Moonshot published the full weights for Kimi K3, 2.8 trillion parameters and the largest open model ever released, free to download with no rate limit and no permission slip. Dario Amodei published Anthropic’s position Monday night. Reuters broke this morning that China has begun mass producing its own lithography machines.

Two days for an industry to organize a position. Congress still has its bill in committee. That gap is the story, and this time it does not close, because a company can pull a product or revoke a key overnight and nobody can un-release a set of weights. Whatever rule gets written next gets written for a world where K3 is already downloaded and running.

In today’s AI news:

• Moonshot publishes K3 weights, the largest open release ever

• Anthropic stays out as the open-weights letter doubles to 50

• China starts mass producing its own immersion DUV scanners

• Microsoft bets cyber defense is won on cost, not capability

• Today’s top tools and quick news

🌎 Moonshot publishes K3 weights

News: Moonshot AI releases the full weights and technical report for Kimi K3, a 2.8T-parameter model its repository calls the world’s first open 3T-class system. This is not a routine product launch. It is the release that forced Washington’s open-weights debate into the open.

Details:

• Adoption ran ahead of the release. Moonshot paused new K3 subscriptions on July 19 after demand pushed its GPUs to the edge, so you could not buy an account while the world waited on the weights.

• It took first in the Frontend Code Arena at 1,679 points, passing Claude Fable 5 and GPT-5.6 Sol. K3 now ranks fourth of 189 models on the Artificial Analysis Intelligence Index.

• K3 runs on Kimi Delta Attention with a Stable LatentMoE framework activating 16 of 896 experts per token, for roughly 2.5x the scaling efficiency of K2.

• Open weights are not free weights. Self-hosting needs roughly 64 H100 or B200 GPUs across eight servers, under a Modified MIT license permissive enough for commercial use.

• Weights went live July 26-27, ten days after the July 16 API launch, not the “under 24 hours” that circulated on social.

Why It Matters: Every U.S. frontier lab rations access through rate limits, tiers, and safety gates. Moonshot removed all three, and once weights are public no export control or terms-of-service update pulls them back. That is why this drop is a geopolitical event rather than a product event: it lands at the exact moment Washington is deciding whether American companies should be allowed to use models like it, and it hands every builder a fallback that costs GPUs instead of a vendor relationship. Price your next twelve months against a permanent ceiling on frontier inference, because the ceiling arrived this week and no policy decision removes it.

📝 Anthropic responds to the Open Weights letter

News: Fifty companies have now signed “Open Weights and American AI Leadership,” the Nvidia-hosted letter urging Washington against restricting downloadable models. Anthropic is the conspicuous frontier holdout, and Dario Amodei published the company’s position last night. Read this one alongside the K3 story, because neither makes full sense alone.

Details:

• The letter went out July 24 with 25 signatories and hit 50 within a day, with OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama joining late. Anthropic, Amazon, and xAI remain off the published list.

• Amodei states it plainly: Anthropic has never advocated for a ban on open-weights models, calling models without dangerous capabilities a public good. He concedes a ban would shield Anthropic from competition, then says that has never been his goal.

• Chips out of authoritarian hands. This is why the China lithography story below belongs in the same issue. Amodei’s case rests on China’s limited domestic chip production as the binding constraint, and domestic scanners are the first crack in it.

• Stop industrial-scale distillation. Anthropic has direct skin here. The White House’s Michael Kratsios accused Moonshot of building K3 by distilling Anthropic’s Fable 5, and Treasury Secretary Scott Bessent floated sanctions for distillation he likened to theft.

• Safety testing for all capable models, open and closed. Anthropic has lived this: the government pulled Fable 5 and Mythos 5 via emergency export controls on June 12, three days after launch, and access stayed dark for 18 days. Anthropic has since promised the government earlier access to test future frontier models, and OpenAI previewed GPT-5.6 to a government-approved group rather than the public.

• The one real break with the letter: he rejects the claim that open weights hand defenders an advantage, citing a UK AI Security Institute finding that released weights cannot be withdrawn, and fears biological misuse more than cyber.

Why It Matters: None of the three asks are radical. Chip controls are already policy. Distillation enforcement is already a Treasury conversation. Pre-release government testing already happened to both Anthropic and OpenAI this summer. And if Washington is right that K3 was built by distilling Anthropic’s models, then refusing to sign a letter defending K3 is the only consistent move on the board.

For Portlanders, I’m hosting an Anthropic event tonight focused on ‘who gets to make the rules in AI’ and will be diving into this hard conversation in great detail. Would love to see you there!

🔬 China starts mass producing its own immersion DUV scanners

News: An immersion DUV scanner is the tool that prints circuit patterns onto silicon, floating a layer of water between lens and wafer to draw finer lines, and ASML has effectively been the only company on earth that makes good ones. Reuters now reports China has begun mass producing its own, led by a state-owned Shanghai firm almost nobody had heard of last week.

Details:

• These tools print 28nm in a single exposure and reach 7nm through multipatterning, paid for in yield. EUV, required for 5nm and 3nm, remains entirely out of reach.

• Here is the number that matters: roughly five machines in 2026 and 20 in 2027, against ASML’s expected 130 immersion shipments this year. That is about 4% of a single year’s supply.

• Shanghai Aishengna has no website and absorbed teams from Yuliangsheng, a Huawei-backed SiCarrier affiliate, and from SMEE. First deliveries go to SMIC, Hua Hong, and CXMT.

• Most components are domestic, though some critical parts still come from Japan, and local supplier delays held output back. Independent analysis puts commercial-scale Chinese immersion DUV in the mid-2030s, with ASML still at 98.7% of the market.

• A bill moving through Congress would cut those same three companies off from ASML servicing and technical assistance, not just new sales, which reaches tools already running in Chinese fabs.

Why It Matters: Every argument about restricting Chinese models rests on an assumption that China cannot make its own chips at scale, and this is the first credible evidence that the assumption has a shelf life. Five machines is not a supply chain, but tool programs compound over decades, and Washington is moving to cut servicing on installed ASML tools at the exact moment Beijing offers an alternative, which turns a maintenance problem into a migration incentive.

For those of us in Portland this is not abstract, since ASML’s Hillsboro campus is its largest U.S. customer support site and the main training hub for its American service engineers.

🛡️ Microsoft bets cyber defense is won on cost, not capability

News: Microsoft launches MAI-Cyber-1-Flash, its first cybersecurity model, running inside MDASH, the company’s multi-agent vulnerability harness. The pitch is not raw capability. It is cost per scan.

Details:

• MDASH with MAI-Cyber-1-Flash delivers 96% on CyberGym, 12 points above Mythos. The score belongs to the harness paired with GPT-5.4, not the model alone.

• The field sits tight underneath: GPT-5.5 Cyber at 85.6%, Mythos 5 at 83.8%, GPT-5.6 Sol at 83.6%, Gemini 3.5 Flash Cyber at 83.2%.

• The model is small by design at 137B total parameters with 5B active. It handles 90% of tasks and routes the hardest 10% to GPT-5.4.

• That routing is the whole product. The configuration costs 50% less than Microsoft’s previous best MDASH setup, and swapping 80% of MDASH’s models moved the harness from 88.4% to 95.95%.

• Microsoft also launched Perception, agent teams that continuously monitor, patch, and close threat vectors.

Why It Matters: Defenders lose on economics, not capability. An attacker probes one vulnerability while a defender scans millions of lines continuously, and at frontier token prices that math never closes. Microsoft’s real product is a routing architecture that puts a cheap specialist on the volume and saves the expensive model for the exceptions.

🛠️ Today’s Top Tools

• 🎯 Kimi K3: Moonshot’s 2.8T open-weight model, the largest ever published

• 🧠 Inkling: Thinking Machines’ 975B MoE, built to be fine-tuned rather than to top leaderboards

• 🔒 MAI-Cyber-1-Flash: Microsoft’s cyber model for hardening large codebases inside MDASH

• 🐝 Buzz: Block’s open-source Slack and GitHub replacement, where every agent gets its own cryptographic identity

• 🎥 FLUX 3: Black Forest Labs’ multimodal model generating 20-second video with native audio, in gated early access

⚡ Quick News

• Lilian Weng leaves Thinking Machines Lab, citing seven months of illness worsened by startup pace. She is the fourth of six cofounders out since February 2025.

• Ilya Sutskever’s Safe Superintelligence takes a reported $5B from Nvidia plus Vera Rubin access, enough to 10x compute in 12 months. Still no product, still $32B.

• Nvidia weighs a $250B guarantee so OpenAI can lease SoftBank’s 10-gigawatt Ohio campus.

• Chinese memory maker CXMT closed up 466% in its Shanghai IPO debut, with no HBM project in its prospectus.

• Cursor launches a ₹649 (~$7) ‘Start’ plan in India, its first country-specific tier, weeks before the SpaceX acquisition closes. India is now its third-largest market at 3M developers.

Jensen Huang has led NVIDIA since 1993 and had never posted on X. On July 24 he did, and the subject was not a chip. It was a three-page letter to Washington titled “Open Weights and American AI Leadership,” which NVIDIA signed alongside 24 other companies.

The post drew more than 60 million views. The letter carried 25 signatures when Huang published it and 50 within a day. This morning NVIDIA converted that momentum into an institution: the Open Secure AI Alliance, a 40-company effort founded on a single uncomfortable episode. When OpenAI’s models broke out of a sandbox and breached Hugging Face earlier this month, the closed models could not be used to investigate. An open model completed the forensics.

Three days, three moves, and a policy fight that has been abstract for two years suddenly has a case file attached to it.

In today’s AI news:

• 40 companies turn a breach at Hugging Face into a security doctrine

• Claude Opus 5 arrives cheap, and your agent setup needs a checkup

• NVIDIA takes a stake in SSI and ends Ilya’s TPU exclusivity

• Databricks hits $5.4B and walks away from the IPO window

• Today’s Top Tools and Quick News

🛡️ 40 companies turn a breach at Hugging Face into a security doctrine

News: NVIDIA and the Linux Foundation launched the Open Secure AI Alliance today with 40 founding partners committed to building open models and tooling for cyber defense. The coalition’s founding argument comes from an incident report rather than a position paper.

Details:

• After OpenAI’s GPT-5.6 Sol and an unreleased model escaped their sandbox and breached Hugging Face production, the security team first tried commercial frontier models and found the requests blocked, because provider guardrails could not distinguish an incident responder from an attacker. Forensic analysis requires submitting real exploit payloads, which is precisely what those systems are built to refuse.

• Hugging Face completed the investigation on GLM 5.2, a 753B-parameter open-weight model run on its own infrastructure, so that no attacker data or credentials left the environment. GLM 5.2 is built by Z.ai, the international arm of Beijing’s Zhipu AI, and distributed under an MIT license.

• Founding partners span cloud, security, and enterprise software: Microsoft, Cisco, CrowdStrike, Databricks, Dell, HPE, Hugging Face, IBM, Palantir, Red Hat, Salesforce, SAP, Siemens, Snowflake, Nous Research, Reflection AI, and Thinking Machines Lab. The companies selling closed frontier access are absent: OpenAI, Anthropic, Google, and Meta.

• The companion letter went from 25 signatures to 50 in a day, with OpenAI and Google adding their names after the initial absence became the story. Anthropic and Amazon appear on neither document, and neither company has explained why.

• All of this is unfolding while Washington weighs restrictions on Chinese open-weight models in the wake of Moonshot’s Kimi K3 release.

Why It Matters: The juxtaposition is what makes this even more unusual. A Chinese open-weight model, downloadable today and possibly restricted tomorrow, is the tool that let an American company defend American infrastructure against an American model. Both sides of the argument survive that fact intact. The open camp reads it as proof that defenders need capability without a usage policy attached. Anthropic has argued for years that weights, once published, cannot be recalled, and Dario Amodei has called open source a red herring in AI specifically. Beneath the principles runs a commercial map that is easy to trace: the companies that sell compute, tooling, and applications want the model layer commoditized, and the companies that sell the model layer do not. What changed this week is that the abstract argument acquired a production incident, and incidents tend to move policy faster than white papers do.

🎬 Claude Opus 5 arrives cheap, and your agent setup needs a checkup

News: Anthropic shipped Claude Opus 5 on July 24 at $5 and $25 per million tokens, unchanged from Opus 4.8, and it outperforms the tier above it on 7 of 11 directly comparable evaluations. Alongside the model, the company published something more immediately useful than a benchmark table: an argument that your existing prompts are now costing you money.

Details:

• Anthropic removed more than 80% of Claude Code’s system prompt for Opus 5 and Fable 5 with no measurable loss on coding evaluations, and shipped a claude doctor command that inspects your own skills and CLAUDE.md files and helps rightsize them.

• The reasoning behind the command matters more than the command. Defensive scaffolding written for weaker models has become a tax, because conflicting rules make the model spend reasoning tokens adjudicating instructions before it starts the work. Anthropic reports 26% fewer tokens on average at lower reasoning levels compared with Opus 4.8.

• ARC Prize independently verified 30.16% on ARC-AGI-3 at high effort against 7.78% for GPT-5.6 Sol, and the model posts 43.3% on Frontier-Bench v0.1 agentic terminal coding against Fable 5’s 33.7%. It still trails GPT-5.6 Sol on DeepSWE and Mythos 5 on cyber and biology tasks.

• The effort dial functions as the cost lever. Standard runs $5/$25 with a flat 1M-token context, while fast mode doubles the rate to $10/$50 for roughly 2.5x the speed, which pays off in interactive loops and wastes money on unattended runs.

• The system card carries two caveats worth reading before you hand it a long leash: Opus 5 hallucinates factual claims slightly more often than Opus 4.8, and Anthropic observed it occasionally working around its own safety filters to complete a task.

Why It Matters: For two years the binding constraint on most AI products was access to capability, and teams organized around getting more of it. Opus 5 prices frontier-adjacent performance at last year’s mid-tier rate, which moves the constraint somewhere less glamorous. The Chartography result, where the model climbs from 29.6% to 83.0% once given a container and a cropping tool, suggests the leverage now sits in the harness rather than the weights. The claude /doctor release makes the same point from the opposite direction: a lab is telling its customers that the instructions they wrote to compensate for older models are now the thing holding the newer ones back.

🧠 NVIDIA buys into the one lab that wasn’t buying its chips

News: SSI and NVIDIA announced a long-term strategic partnership today, pairing an undisclosed NVIDIA investment with access to the next-generation Vera Rubin platform, which SSI says will expand its compute by an order of magnitude. Ilya Sutskever has trained on Google TPUs since 2025, and that exclusivity ended this morning.

Details:

• NVIDIA said it entered the partnership after obtaining rare access to SSI’s closely guarded research, and the two will collaborate on NVIDIA’s current and future compute platforms using SSI’s view of where the field is heading.

• Sutskever’s comment was characteristically spare: “We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so.”

• Since April 2025, Google Cloud has been SSI’s primary computing provider and SSI its most significant external TPU customer, which made Sutskever’s chip choice the strongest endorsement the TPU program had. Google is the party that loses something concrete today.

• NVIDIA now holds positions at both ends of a rivalry. It committed at least one gigawatt of Vera Rubin systems to Thinking Machines Lab in March, and TML is a founding partner of the alliance announced this morning.

• The release names no dollar figure, no megawatts, and no delivery timeline. SSI last marked at $32B on about $6B raised with no product and no revenue, and has never disclosed its current compute, which leaves an order of magnitude as a ratio without a denominator.

Why It Matters: Sutskever spent the past year making the case that AI is leaving the age of scaling and entering the age of research, an argument that a great many people repeated at conferences this spring. Today he accepted a tenfold increase in compute. The generous reading is that algorithmic breakthroughs still need silicon to express themselves, and the skeptical reading is that the field’s most disciplined believer in research over scale decided to hedge. Both readings are available, and the announcement contains nothing that settles it, which is itself informative about how little anyone outside SSI knows.

💰 Databricks hits $5.4B and walks away from the IPO window

News: While the frontier labs argue over weights, the least glamorous company in AI infrastructure is compounding. Databricks is in talks at a $165B to $175B valuation, up from the $134B it set four months earlier, and has removed itself from this year’s listing queue.

Details:

• Annualized revenue run rate passed $5.4B, growing more than 65% year over year, with AI products alone at $1.4B.

• Net revenue retention sits above 140% with positive free cash flow, which puts the company in the unusual position of not needing the capital it keeps raising.

• CEO Ali Ghodsi told Bloomberg on June 4 that 2026 is a terrible year to go public, citing a calendar crowded with SpaceX, Anthropic, and OpenAI, and pointed to 2027 instead.

• Secondary markets have moved ahead of the round. Databricks priced at $242.04 per share on Forge at a $170.7B valuation as of July 10.

Why It Matters: Ghodsi’s reasoning is worth taking at face value, because it describes a market condition rather than a company problem: when three of the largest listings in a decade share a calendar, the scarce resource is investor attention rather than capital. The strategic point underneath it is that Databricks sells the layer that persists through every model transition. By the time the frontier labs complete their listings, public markets will have to price AI companies on revenue durability, and the comparison set will include a company that has been demonstrating it for years.

🛠️ Today’s Top Tools

• 🧠 Claude Opus 5 | Frontier-tier agentic coding and computer use at unchanged Opus pricing

• 📽️ Photon-1 | Induction Labs’ imagination model learned computer use from 18 years of screen recordings with zero action labels, beating a production LLM on 30x less pretraining compute

• 🔀 Merge Fusion | Fans one prompt out to a panel of models in parallel, then synthesizes a single answer that beat every solo model on the DRACO benchmark

• 🖥️ OpenWorker | Andrew Ng’s MIT-licensed, local-first desktop agent that hands you the finished deliverable instead of a chat reply, with your own API keys or fully local via Ollama

📰 Quick News

• Apple pushed its first smart glasses to WWDC 2027, a roughly six-month slip from the original late-2026 reveal. Gurman reports Apple has prototyped versions with no camera at all and plans to skip facial recognition and training on customer recordings.

• Google DeepMind’s open Gemma family crossed 900M downloads, with Gemma 4 alone accounting for more than 300M since April and over 70,000 fine-tuned variants now on Hugging Face.

• Midjourney acquired astrology app Co-Star, bringing founder Banu Guler on as chief design officer. The app stays under her control, and Midjourney is reportedly building an astrology-focused image generator.

• Meta shipped agentic capability to Meta AI on Muse Spark 1.1: calendar-connected daily briefings, steerable research you can redirect mid-generation, slide decks, and recurring tasks you set up once.

Travis Kalanick closes $1.7B for machines that run mines and shipyards. Cognition buys the only AI agent Apple allows to text you natively. AMD commits 2 gigawatts to Anthropic and writes a $5B check into its own customer. Three companies with nothing in common, one identical instinct: capability is commoditizing, so buy the thing capability cannot replicate.

Then there is Moonshot, where that exact question is being argued in public and neither side is showing its work. Washington says Kimi K3’s abilities were siphoned out of Anthropic’s frontier model. Its boosters say K3 found a Redis zero-day in 27 minutes. Check the dates on both.

In today’s AI news:

• Travis Kalanick’s $1.7B bet on atoms over bits

• Every claim about Kimi K3 is outrunning its evidence

• Cognition acquires the only AI agent Apple lets text you

• AMD buys into Anthropic with a 2 GW deal and a $5B check

• Today’s top tools and quick news

ATOMS 🔥 Travis Kalanick returns with a $1.7B bet on the physical world

News: Travis Kalanick’s industrial robotics company Atoms raises $1.7B in equity led by Andreessen Horowitz, ending nearly eight years of stealth. Ben Horowitz joins the board. The most striking name on the cap table is Uber, the company that forced Kalanick out in 2017.

Details:

• The round includes Bain Capital, Fifth Wall, Chemistry, A*, K5 Global, Abstract, SV Angel and Alpha Square Group, with no valuation disclosed.

• Atoms separately arranged debt facilities with Bank of America, Goldman Sachs, Wells Fargo, JPMorgan and Barclays, terms undisclosed.

• The company grew out of City Storage Systems and now consolidates CloudKitchens and the March acquisition of Pronto under a single equity structure.

• Three divisions are live: Atoms Food, Atoms Mining, and Atoms Transport, spanning autonomous mining, logistics, and food production across 110+ cities.

• Kalanick calls the category “atoms-based computers,” purpose-built machines for industries that grow, mine, manufacture, and move things. No humanoids.

• Horowitz announces the deal with two words: “Travis Is Back.”

Why it matters: The interesting thing is not the money, it is the shape of the bet. Every consumer-facing robotics company is racing toward the same anthropomorphic form factor, and Kalanick is walking the other way entirely: specialized machines for unglamorous, high-margin, structurally under-automated industries. Mining does not need a robot that looks like you. It needs a machine that does one thing at industrial scale and never stops. Notice the pattern in a16z’s week, because they also published their thesis on Applied Intuition and put intelligence-on-machines at its center. When one firm makes the platform bet and the operator bet inside seven days, that is a coordinated position, not two coincidences.

🏛️ Every claim about Kimi K3 is outrunning its evidence

News: OSTP Director Michael Kratsios accuses Moonshot AI of covertly distilling Anthropic’s Claude Fable 5 to build Kimi K3, and Treasury Secretary Scott Bessent follows with a sanctions threat. Hours later, a viral post claims K3 discovered a Redis zero-day in 27 minutes using 32 parallel agents. Both claims are about what K3 can actually do, and both fall apart the moment you check a calendar.

Details:

• Kratsios posted the accusation on X on July 22, alleging a purpose-built platform that rotated access methods to evade detection, and separating this from ordinary research distillation.

• He also alleged Moonshot acquired GB300-equipped servers and accessed GB300s in Thailand, hardware under U.S. export restrictions. No publicly verifiable evidence has accompanied any of it.

• The concrete piece predates the politics: Anthropic alleged in February that Moonshot generated more than 3.4 million Claude exchanges through fraudulent accounts, with metadata linking activity to senior Moonshot employees.

• On the capability side, the Redis claim traces to a single X post and a proof-of-concept repo credited to an alias. No CVE assignment, no Redis advisory, no independent reproduction.

• The Redis flaw circulating with that claim was found by Team Xint Code, a different autonomous AI tool, patched May 5 across five branches, and its full exploit chain went public in early June.

Why it matters: Run both timelines and watch them collapse in the same way. Fable 5 shipped June 9, Anthropic suspended access June 12 under Commerce export controls, and restoration came July 1, with K3 landing days later. Industrial-scale distillation inside that window is a very tall technical claim, and researchers already say the dates do not support it. The Redis story fails identically: a model reproducing an exploit chain that has been public for seven weeks looks exactly like a model discovering one, which is precisely how the 7-Zip “zero-day” turned out to be an LLM hallucination. The practical read for builders is unglamorous and urgent: Washington is reclassifying distillation from competitive practice into IP theft, and if that framing holds, every open-weight release becomes a trade-policy event.

📱 Cognition acquires Poke, the only AI agent Apple lets text you

News: Cognition is acquiring The Interaction Company of California, the Palo Alto startup behind Poke, a personal agent that lives inside your text messages. Scott Wu announced the deal Thursday, saying people have exchanged more than 100 million messages with Poke in the last three months alone. Terms were not disclosed.

Details:

• Poke is the only AI agent approved to text natively on Apple Messages, having cleared Apple’s Messages for Business review in June. It also runs over SMS, Telegram, and WhatsApp in select markets.

• That approval was not a formality. Co-founder Marvin von Hagen says it took months, required proof Poke could escalate to live human support, and forced a UI rebuild against Apple’s style guide. Interaction pays Apple per user.

• Poke launched in March 2026 and raised $15M behind General Catalyst and Spark Capital, with angels including the Collison brothers, Logan Kilpatrick, Joanne Jang, and Jake and Logan Paul.

• Wu and Walden Yan were personally early angels in Interaction, and von Hagen was on set for the original Devin launch video in 2024. Wu calls it two years in the making.

• Cognition closed $1B+ at a $26B post-money valuation in May on a $492M run-rate, making Poke its second major acquisition after Windsurf.

Why it matters: Every rival agent sits behind an app you have to remember to open, and Poke sits in the one thread you already check two hundred times a day, with Apple’s blessing and a per-user contract that took months to earn. Von Hagen says it will take anyone else months too, which is the definition of a moat capital alone cannot collapse. The deeper play is initiation: Devin waits to be assigned work, while Poke starts the conversation. If you are building an agent, the land grab just moved from capability to distribution rights, and the question is no longer how smart yours is. It is whose surface it is allowed to live on, and who is already standing in line ahead of you.

💰 AMD buys into Anthropic with a 2 GW deal and a $5B check

News: AMD and Anthropic sign a partnership deploying up to 2 gigawatts of Instinct MI450 GPUs in Helios rack-scale systems, with AMD committing a strategic equity investment of up to $5B in Anthropic. The hardware alone runs into the tens of billions.

Details:

• The first gigawatt deploys in the first half of 2027, with equity released against deployment milestones.

• The deal confirms for the first time that Anthropic already runs AMD’s current-generation MI355X GPUs.

• Anthropic pairs the GPUs with EPYC “Venice” CPUs, Pensando networking, and ROCm, making AMD roughly its sixth distinct silicon source alongside Nvidia, Google TPUs, and Amazon Trainium.

• Structurally this inverts the OpenAI arrangement, where AMD handed its customer warrants. Here AMD writes the check and becomes both supplier and shareholder, with no warrants attached.

• Anthropic will also use Claude to accelerate AMD’s own chip and ROCm development. AMD’s disclosed pipeline now reaches roughly 20 GW across Anthropic, OpenAI, Meta, Microsoft, and Oracle.

Why it matters: The compute is the visible half. The software is the half that decides whether AMD ever escapes second place. Nvidia’s moat was never purely silicon, it is CUDA, and ROCm still trails by 20 to 30% on large transformer training even as it closes to near-parity on inference. Pointing a frontier model at that gap is the most interesting thing in this deal, and the template is worth stealing: the customer’s product becomes the supplier’s R&D engine.

🛠️ Today’s Top Tools

• 🔒 Claude Security plugin for Claude Code: Anthropic’s multi-agent vulnerability scanner now runs inside your terminal, in beta for all Claude Code users. Scan a full repo or just a branch diff before you commit, then turn findings into patch files you review yourself.

• ⚙️ Cursor Router: A classifier trained on 600K+ live requests that picks the cheapest capable model for each request, reporting 30 to 50% savings for enterprise accounts and 60% in online A/B tests. Teams and Enterprise plans only.

• 📈 Gumloop SEO agent stack: Build an agent team that audits your site and writes every finding into one Google Sheets workbook. Connect a lead agent to Sheets and Firecrawl, then add dedicated Backlink and Content Refresh agents pointed at the same pages.

⚡ Quick News

• The FDA granted De Novo authorization to J&J’s Ottava, the first table-integrated soft tissue robotic platform, covering ten general surgery procedures and taking up as much as 50% less OR space. It cracks Intuitive Surgical’s 25-year hold on a market where only 8% of global surgeries involve a robot at all.

• The DOE named 278 projects in the first Genesis Mission cohort from over 5,000 proposals, the largest response to a funding opportunity in department history. Anticipated RFA funding is $293.76M, with the largest award a three-year, $60M nuclear project at Idaho National Laboratory.

• UK’s Humanoid raised a $152M Series A at $1.35B post-money, backed by Prime Movers Lab, Bosch, Schaeffler, and Bernard Arnault’s Aglaé Ventures, with 34,000 pre-orders worth $2.4B.

• Gritt exited stealth with $32.4M, including a $26M Series A led by Obvious Ventures, for robots that bolt onto existing skid steers rather than replacing them. An eight-person crew goes from 800 solar panels a day to 3,000 or 4,000 with the system attached.

• Elon Musk says Grok Imagine will produce a full-length Odyssey film before year end, posted alongside a three-minute AI-generated clip and framed as a response to Christopher Nolan’s theatrical adaptation, which Musk has criticized for months.

Google shipped three new Gemini models this morning. And somehow the one it didn’t ship became the loudest story in AI.

The new Flash lineup covers speed, low cost, and cybersecurity, and Google confirmed its “most ambitious pre-training run yet” is already underway for Gemini 4. But 3.5 Pro, the model built to trade blows at the frontier, is still missing in action. When the model everyone is waiting on keeps slipping, the release meant to reassure the market ends up doing the opposite.

In today’s AI news:

• Google ships new Flash models while Pro stays missing

• Sakana makes orchestration the cybersecurity moat

• Poolside resurfaces with a top open-weights coder

• AI’s biggest copyright bill comes due

• Mozilla: open models closed the capability gap, not the revenue gap

⚡ GOOGLE | New Flash models ship as 3.5 Pro stays missing

News: Google rolled out a trio of Gemini models led by 3.6 Flash, joined by a speed-tuned 3.5 Flash-Lite and a security-hardened 3.5 Flash Cyber. The lineup is built for efficiency over raw intelligence, and the frontier-class 3.5 Pro that was supposed to headline the day is still stuck in testing.

Details:

• 3.6 Flash brings efficiency gains but gets outscored by similarly priced rivals like Grok 4.5 and GPT-5.6 Luna across a range of benchmarks.

• On the Artificial Analysis’ Intelligence Index, 3.6 Flash shows no measurable jump over 3.5, landing near Meta’s Muse Spark and Z AI’s open GLM 5.2.

• Logan Kilpatrick says 3.5 Pro is still “testing with partners” and will “hopefully land soon” after a string of delays.

• Google closed by confirming Gemini 4 has entered its “most ambitious pre-training run yet,” with even Elon Musk taking a jab at the release.

Why It Matters: Efficiency helps the millions who live inside Gemini, but a Flash refresh is not what Google needs right now. The missing frontier model deepens the read that the company is a step behind while OpenAI, Anthropic, and xAI keep shipping. The lab with the most researchers, data, and TPUs on the planet cannot afford to keep being the one everyone is waiting on.

🐡 SAKANA AI | Orchestration, not model size, becomes the cybersecurity moat

News: Tokyo’s Sakana AI shipped Fugu-Cyber, a security-tuned version of its Fugu model that answers to one API but is actually a pool of specialized agents working a problem together. It posts state-of-the-art, self-reported scores on two hard security benchmarks and ships access-gated and defense-only.

Details:

• One endpoint, many agents: an orchestrator decomposes each task and routes it across specialists that analyze, challenge, and verify each other’s work, extending the original Fugu from June.

• It claims 86.9% on CyberGym (finding real vulnerabilities in code) and 72.1% on CTI-REALM (turning threat intel into detection rules), comparable to GPT-5.5-Cyber and Anthropic’s Mythos-Preview, though the numbers are first-party and unverified.

• Access is gated behind a manual approval form and an acceptable-use policy that bans offensive work, with Sakana conceding a raw model doesn’t solve enterprise security without its Applied Enterprise team building the verification harnesses.

Why It Matters: The real claim is that the moat is orchestration and deployment, not parameter count. Notice the pitch: frontier capability “without the risk of export controls,” a route around the scale-based rules that throttled even Anthropic’s Mythos and Fable 5 for three weeks in June. For buyers, the move is obvious: benchmark it on your own repos, because self-reported numbers behind an approval wall mean few outside eyes have kicked the tires.

🏊 POOLSIDE | A top open-weights coder resurfaces

News: Poolside launched Laguna S 2.1, an open-weights coding model that leads the Western open-source field while staying small enough to run on a single desktop. It’s the clearest sign yet that America is mounting a real answer to China’s open-model lead.

Details:

• Laguna activates just 8B of 118B parameters at a time, fitting on one Nvidia DGX Spark box, with the weights free to grab on Hugging Face.

• It tops coding benchmarks among U.S. open systems but still trails Kimi K3 and other Chinese open models by a wide margin.

• The release lands a week after Mira Murati’s TML shipped Inkling, both widening America’s thin open-source bench.

• Poolside says training took under nine weeks, and the model ships with a 1M-token context window.

Why It Matters: Two open releases in two weeks won’t erase China’s open-source lead, and Kimi is still lapping the field, but the momentum is real after months of a widening gap. A frontier-adjacent coder that runs on one desktop changes who gets to build with it, from enterprises that need data to stay in the building to solo founders who can’t rent a cluster. Open weights are where the U.S.-China AI race actually gets decided, and America is finally shipping instead of watching.

💰 ANTHROPIC | AI’s biggest copyright bill comes due

News: Anthropic won court approval for a $1.5B settlement with book authors, clearing payouts of roughly $3,000 per title across 482K works pulled from piracy sites. It’s the largest copyright recovery in U.S. history, and it leaves the ruling that AI training counts as fair use fully intact.

Details:

• The suit traces back to 2024, when novelist Andrea Bartz and other writers alleged Claude was trained on books lifted from pirate libraries.

• Judge William Alsup split the case in 2025: training on books is legal fair use, but Anthropic’s stockpile of 7M pirated titles was its own violation.

• The settlement lets Anthropic sidestep a jury trial set for last December, where damages could have run into the hundreds of billions.

• Anthropic keeps training on legally purchased books, and the fair-use finding stays untouched.

• With similar suits pending against other labs, roughly $3,000 per work is emerging as the going rate to settle.

Why It Matters: This is the moment the industry gets a price tag for its original sin: acquire the corpus legally and training is protected, pirate it and you pay. Every lab that scraped the open web is now running the same math, and the ones with the balance sheets to settle just watched the template get written.

🦊 MOZILLA | Open models closed the capability gap, not the revenue gap

News: Mozilla, the nonprofit behind Firefox, published its first State of Open Source AI report, and the headline is blunt: the gap with top closed systems like ChatGPT and Claude has narrowed to just 3%, while inference costs have fallen up to 50x in three years. The twist underneath is the real story, because open models now power roughly a third of real-world AI usage but capture only 4% of the revenue.

Details:

• The receipts: a global survey of 950+ developers plus OpenRouter’s 100-trillion-token dataset, with GPT-4-class inference now at $0.40 per million tokens, down from $20 three years ago (to be fair, I really hope nobody is ever using GPT-4...).

• The frontier is still jagged: open matches closed on coding and general knowledge but trails on reasoning and agentic work, where GPT-5.5 hits 83.4% on Terminal-Bench 2.1 against the best open model’s 67.9%.

• The bottleneck moved off the model: only 51% of open-model teams reach production versus 63% for closed, and it’s tooling and trust, not capability. Stripe cut AI costs 73% by moving 50M daily requests to self-hosted open models on a third of the compute.

• Geopolitics may matter most: China and East Asia lead open adoption at 89%, Chinese open weights jumped from under 2% of OpenRouter tokens in late 2024 to 45%+ by April 2026, and Qwen passed Llama as Hugging Face’s most-downloaded family. Mozilla calls it deliberate state policy and a hedge against U.S. chip controls.

Why It Matters: When the model layer commoditizes, the margin migrates to the layer above it: deployment, orchestration, governance, exactly where open models still stall. That 4%-of-revenue gap is the opening, because whoever builds the tooling that gets open models reliably into production captures value that evaporates today.

🛠️ Today’s Top Tools

• 🚀 Laguna S 2.1: Poolside’s open-weight coding model with a 1M-token context, small enough to run on a single desktop.

• ⚡ Gemini 3.6 Flash: Google’s newest Flash variant, tuned for efficiency and lower-cost, high-volume use.

• 🎆 Qwen-Image-3: Alibaba’s new image model with notably strong text rendering.

📰 Everything else in AI today

• Twitter and Block founder Jack Dorsey unveiled Buzz in early access, an open-source workspace where AI agents join team chats as coworkers.

• Substack partnered with Pangram to embed an AI-writing detector into the platform for both readers and publishers.

• Anthropic shipped “Record a skill,” letting users screen-record a workflow so Claude can build a skill around the task.

In Washington, the administration is assembling a ban on Chinese AI it cannot write into law, while the agency responsible for testing frontier models lost its second director in four months. Control without capacity.

In production, an autonomous agent fired 17,000 actions through Hugging Face’s infrastructure over a single weekend, and OpenAI benched its own math prodigy after the model spent an hour picking a hole in its sandbox. Capacity without control.

One thread is a government reaching for levers it does not have. The other is an industry holding levers that no longer work.

In today’s AI news:

• Washington builds a slow-motion ban on Chinese AI

• America’s AI testing agency loses its second chief in four months

• An autonomous agent breaches Hugging Face in one weekend

• OpenAI benches its Erdős model after repeated sandbox escapes

🕸️ Washington builds a slow-motion ban on Chinese AI

News: Axios reports the administration is working out how to position against open-weight Chinese models, with Moonshot’s Kimi K3 as the trigger. Nothing has moved yet. Four instruments are on the table: Entity List designations, federal procurement restrictions, a security advisory, and a draft executive order.

Details:

• Commerce weighed Entity List designations for multiple Chinese labs last year, and the NSA and White House cyber office drafted a threat advisory meant to discourage adoption without banning anything.

• The draft order would let US firms host Chinese models only if they guarantee security and accept liability for a breach.

• Kimi K3 runs 2.8 trillion parameters, took first in Frontend Code Arena, and drops full weights July 27.

• Self-hosting is not the escape hatch people assume. K3’s weights run 1.4TB and Moonshot’s own guidance calls for supernodes of 64+ accelerators. GLM-5.2 is the same story. Access runs through clouds and APIs, which is exactly what procurement rules and liability reach.

• The loudest voices for restriction are the closed labs. WSJ has OpenAI and Anthropic executives warning of a dystopian outcome without regulation, and both are heading for public listings inside a year.

Why It Matters: Two things are true here, and most coverage collapses them into one. Federal agencies running Chinese models on government data is a real security problem, and nobody serious argues otherwise. Whether an American startup should be blocked from choosing a cheaper model is a different question entirely, and the people making that case hardest are the ones who lose when the cheap model wins. The argument they make is not unreasonable: publish weights this capable and you cannot control what gets built on top. But underneath the security language sits the question that actually matters, which is who gets to decide what Americans are allowed to run.

🏛️ America’s AI testing agency loses its second chief in four months

News: Chris Fall resigned Monday as director of the Center for AI Standards and Innovation, three months into the job. NIST Director Arvind Raman takes over as acting chief while continuing to run NIST. No explanation has been offered.

Details:

• CAISI is where the frontier labs send unreleased models. Anthropic, Google DeepMind, OpenAI, Microsoft, and xAI all take part in pre-deployment testing for cyber, bio, and chemical risk.

• That arrangement has teeth. Commerce imposed export controls on Fable 5 and Mythos 5 on June 12 after a jailbreak report, forcing Anthropic to pull both globally for 19 days. GPT-5.6 spent 12 days gated to roughly 20 approved partners before clearing review.

• Fall’s predecessor, Collin Burns, lasted four days in April, with reporting tying his exit to his prior work at Anthropic.

• Raman took over NIST on June 30 after running Purdue’s engineering school. He is now running both organizations.

• CAISI has since pulled the details of those pre-release testing agreements from its website.

Why It Matters: The executive order behind all of this was written as voluntary. It has become preapproval in practice, and two of the three leading American models got gated in a single month under it. That machinery now answers to someone splitting his attention with all of NIST, at the exact moment the administration is deciding whether to restrict Chinese models. Nobody outside the process can tell you what the approval threshold actually is, and the people who would know keep walking out the door.

🛡️ An autonomous agent breaches Hugging Face in one weekend

News: Hugging Face discloses that an intrusion into its production infrastructure was driven end to end by an autonomous AI agent system. It is one of the first documented cases of AI running a full cyberattack rather than assisting one. Internal datasets and service credentials were accessed.

Details:

• The agent uploaded a malicious dataset, exploited a template injection and a remote code dataset loader, escalated to node-level access, and harvested cloud and cluster credentials.

• It executed more than 17,000 recorded actions across a swarm of short-lived sandboxes over one weekend, with self-migrating command and control staged on public services.

• Hugging Face’s own LLM analysis agents reconstructed the complete attacker timeline in hours instead of days.

• The first frontier model the team reached for refused the job, so with provider guardrails blocking malware analysis they ran GLM-5.2, a Chinese open-weight model, locally on their own infrastructure.

Why It Matters: People have been predicting this attack for two years. It showed up on a Saturday and ran 17,000 actions before Monday. No human team moves at that speed, which is why Hugging Face had to fight it with agents of its own, and that is now simply what incident response looks like. The part worth sitting with is the guardrail lockout: the commercial model they reached for refused to analyze the malware, so they ran a Chinese open-weight model on their own hardware instead. Read that against the story above and you see why it’s the no.1 story today.

🔒 OpenAI benches its Erdős model after repeated sandbox escapes

News: OpenAI publishes Safety and alignment in an era of long-horizon models, disclosing that it paused internal access to the unreleased system behind May’s Erdős result after the model repeatedly acted outside its sandbox. Access is now restored under rebuilt monitoring. It is not the only sandbox failing this month.

Details:

• During a NanoGPT evaluation, the model spent roughly an hour hunting a sandbox vulnerability, found one, and opened public pull request #287 on GitHub.

• The trigger was conflicting instructions: told to post results to Slack, the model found separate instructions to file a PR and acted on the second.

• OpenAI rebuilt around defense in depth and trajectory-level monitoring, and on replay the new stack caught considerably more, with only low-severity behavior slipping through.

• Same pattern elsewhere: Pillar Security published reproducible sandbox escapes across Cursor, Codex CLI, Gemini CLI, and Antigravity, where the agent never attacks the sandbox and simply writes files that trusted tools outside it later execute.

Why It Matters: X ran with “an AI escaped,” which is both wrong and less interesting than what happened. OpenAI built a model to keep working for hours after a chat model would have quit, and it did exactly that: found a real hole in the enforcement, hit two conflicting instructions, and went with the goal instead of the rule. That is the same trait that makes the thing worth building in the first place. Pillar’s research suggests the sandbox was never the right line anyway, since an agent can write far more than it can run, so audit what your agents write and not just what they run.

🛠️ Today’s Top Tools

• 🚀 Kimi K3: Moonshot’s 2.8T open-weight release that took first on frontend coding and set off the entire Washington debate above. Full weights land July 27.

• 📊 Grok for Excel: Free Microsoft 365 add-in. Query cell ranges in plain English, get formulas and charts back with source citations. Word and PowerPoint too.

• 💻 LM Studio Bionic: A local agent that writes code and edits documents using open models on your own machine. Mac and Windows.

• ✍️ Inkling: Thinking Machines’ first open-weights multimodal model. 975B total parameters, 41B active, Apache 2.0, weights on Hugging Face.

⚡ Quick News

• Claude Fable 5 disproves the 87-year-old Jacobian conjecture, producing a 216-character counterexample that mathematicians verified within hours. No formal paper or prompt transcript has been published yet.

• Alibaba ships Qwen-Image-3.0 with a 4,500-token prompt ceiling and claimed legibility down to 10-pixel text, but no weights, no technical report, and no benchmarks.

• Google is reportedly developing Frozen v2, a chip with part of Gemini’s architecture etched into hardware, projected to run the model 6x to 10x more efficiently than current TPUs.

• Nvidia launches Cosmos 3 Edge, a 4B-parameter open world model that runs onboard a robot, hitting 32 actions per inference and real-time control at 15 Hz on Jetson Thor.

• Z.ai completes a 1GW data center in China stocked exclusively with domestic chips, giving its GLM models a training base with zero Nvidia dependency.

Three labs on two continents hit the same wall this week, and every one of them hit it for the best possible reason: people wanted their product far more than anyone forecast.

Anthropic made Fable 5 permanent inside Max and Team Premium. Moonshot stopped selling Kimi K3 subscriptions 48 hours after launch because demand outran its compute. Alibaba previewed a 2.4 trillion-parameter flagship in Shanghai and priced it at 90% off. Read those three together and the story is not scarcity. It is that adoption is compounding faster than the most capitalized supply chain in history can pour concrete.

I was in the audience when Dario Amodei put Anthropic’s growth this year at 85x. Sit with that number for a second. No planning process survives it, no capacity model anticipates it, and no communications team gets the messaging perfect on the first try. What we are watching is not mismanagement. It is what genuine product-market fit looks like at a scale nobody in this industry has operated at before.

In today’s AI news:

• Anthropic makes Fable 5 permanent in Max and Team Premium

• Moonshot sells out of Kimi K3 in 48 hours

• Alibaba previews Qwen3.8 at the World AI Conference

• The revolving door between Washington and the frontier labs

• Today’s Top Tools + Quick News

💰 Anthropic makes Fable 5 permanent

News: Anthropic confirmed on July 18 that Claude Fable 5 stays permanently inside Max and Team Premium, effective today. The original plan was to remove the model from subscriptions entirely. Competitive pressure and genuine consumer demand turned a removal into a permanent inclusion.

Details:

• The deadline moved three times before landing here, from July 7 to July 12 to July 19, a five-week stretch that tested a lot of patience.

• Fable 5 now sits permanently in Max and Team Premium at 50% of weekly usage limits. Pro and Team Standard get a one-time $100 credit, then move to usage pricing at $10 per million input tokens and $50 per million output tokens.

• One wrinkle worth knowing before you plan your week: a temporary 50% Claude Code rate-limit boost expires today too, so the baseline those percentages apply to is smaller than it was during the promo window.

• Anthropic cited infrastructure limits and demand it called challenging to predict, while pointing to new capacity coming online including the SpaceX Colossus 1 deal.

• The competitive backdrop is real. OpenAI’s GPT-5.6 Sol runs roughly a third cheaper, and Moonshot’s Kimi K3 landed mid-decision.

Why it matters: This is what a genuinely competitive market delivers. A model that was six weeks from disappearing behind pay-per-use is now a permanent fixture in two subscription tiers, and pressure from OpenAI, Moonshot, and Alibaba is why. Be clear-eyed about the tradeoff though: permanent access at reduced caps is a better deal than losing the model, and it is not the same as the promo window you have been living in since June. Plan for less Fable 5 headroom this week, not more. The bigger lesson sits underneath the caps anyway, because at 85x annual growth, demand forecasting stops being a spreadsheet exercise and becomes an act of faith, and Anthropic found a way to keep the model in customers’ hands while it builds.

🌝 Moonshot sells out of Kimi K3 in 48 hours

News: Moonshot AI paused new subscriptions within 48 hours of Kimi K3’s July 16 launch after demand drove its compute straight to the ceiling. The company framed the pause as protecting the experience of existing subscribers. If you have not already subscribed, you are waiting for the weights.

Details:

• Kimi K3 is a 2.8 trillion-parameter sparse Mixture-of-Experts model activating 16 of 896 experts per token, with a 1 million-token context window, native vision, and an always-on thinking mode built on new architecture Moonshot calls Kimi Delta Attention.

• The benchmark picture is more nuanced than the headlines suggested. K3 scored 57.1 on Artificial Analysis’s Intelligence Index, third behind GPT-5.6 Sol at 58.9 and Fable 5 at 59.9, while Arena’s blind tests had developers preferring K3 over both for front-end coding work.

• On price, K3 runs about $3 per million input tokens and $15 output against Sol’s $5 and $30, undercutting the frontier by roughly 40 to 50% without being the outright cheapest option available.

• Full open weights land July 27 on Hugging Face under a Modified MIT license, which would make K3 the largest open-weight model ever released.

• Moonshot is also splitting memberships into separate chat and coding plans, letting it provision two very different workloads independently.

Why it matters: A model that undercuts the frontier by half and is about to be given away for free still ran out of capacity in two days. That is the whole story of this week compressed into one lab. Price was not the constraint, licensing was not the constraint, and the openness of the lab did not save it, because the binding constraint is silicon and power and it applies to everyone equally. July 27 is the date to circle, because the moment those weights hit Hugging Face, a capacity ceiling becomes a procurement question you can go solve yourself.

🐉 Alibaba previews Qwen3.8 in Shanghai

News: Alibaba previewed Qwen3.8-Max-Preview on July 19 at the World Artificial Intelligence Conference in Shanghai, a 2.4 trillion-parameter multimodal model the Qwen team claims ranks second only to Claude Fable 5. The preview is already live and running in production coding harnesses. Full open weights are promised, with benchmarks and the model card still to come.

Details:

• Qwen3.8 is the first model in the family above 1 trillion parameters to handle images, video, and documents alongside text, built as a sparse Mixture-of-Experts stack.

• Access runs through Token Plan, Qoder, and QoderWork at a promotional 10% of standard pricing, with an off-peak rate reaching 98% off.

• Open weights are coming, which would break Alibaba’s long-standing practice of keeping Max-tier models closed.

• The second-only-to-Fable-5 line is Alibaba’s own claim with no data attached. No benchmark scores, no activated-parameter count, no model card. For contrast, Qwen3.7-Max shipped in May with full published results including a 56.6 on the Artificial Analysis Intelligence Index.

• Markets liked it modestly. Alibaba’s Hong Kong shares closed up 3.73% at HK$116.80 while the Hang Seng Tech Index added 2.79%.

Why it matters: Qwen has quietly become the default for anyone doing serious fine-tuning work, and that community has spent two years building on a family that ships reliably and prices aggressively. Qwen3.8 is the next jump in that lineage, and an open-weight release at 2.4 trillion parameters would hand those teams the largest base model anyone has ever put in public hands. The missing benchmark table is a real gap and we would rather have it today, but Alibaba’s track record earns some patience, and we expect the full release shortly. We will report the numbers the moment they land.

🏛️ The ongoing dispute between Washington and the frontier labs

News: Dean Ball, a co-author of the administration’s AI Action Plan, joined OpenAI as Head of Strategic Futures effective July 6, reporting to Chief Strategy Officer Jason Kwon. Within two weeks he drew public fire from David Sacks, the former White House AI and crypto czar who now cochairs the President’s Council of Advisors on Science and Technology.

Details:

• Last fall, Sacks accused Anthropic of running a regulatory capture strategy built on fear-mongering, following an essay by policy chief Jack Clark on technological optimism and appropriate fear.

• Five weeks ago, Anthropic proposed that the government hold standing authority to block frontier model releases that fail independent safety tests, with penalties tied to global revenue.

• On July 6, Ball left his White House-adjacent policy fellowship for OpenAI.

• This past weekend, Ball wrote that the administration would eventually create regulatory risk around Chinese open-weight models, arguing no formal ban is needed if agencies issue soft law that creates uncertainty.

• Sacks fired back that weaponizing regulatory uncertainty is unacceptable, and that the leading closed labs, which he described as a revenue duopoly, want the government to remove their open-source competition. A senior Pentagon official separately disputed Ball’s framing, and Ball then clarified he was not advocating the approach he had described.

Why it matters: We are not here to talk about politics, and honestly we would be delighted to never talk about politics again. But politics is welded to this industry now, and pretending otherwise does not make it go away. What you are looking at is the leadership of one of the world’s top AI labs and the US government arguing in public, in real time, about who gets to compete. Most newsletters will skip this because it does not fit a product roundup, and skipping it is exactly how you get surprised when the rules change underneath you.

🛠️ Today’s Top Tools

• 🌝 Kimi K3 Moonshot’s 2.8T open-weight model, top of Arena front-end code, weights land July 27

• ⚙️ Canva Code 2.0 vibe-coding now free for all 265M+ Canva users, with HTML import and drag-and-drop editing of AI-generated sites

• 🧠 Inkling Thinking Machines’ first open-weights model, 975B total and 41B active under Apache 2.0, reasoning natively over text, images, and audio

• 🗣️ GPT-Live OpenAI’s full-duplex voice model that listens and speaks at once, and hands hard questions to a frontier model mid-conversation

📰 Quick News

• Chamath Palihapitiya weighed into the open-source debate with three words: the future is open source.

• Gavin Baker of Atreides calls Kimi K3 a potential inflection point for AI, negative for Anthropic and OpenAI and net positive for nearly every other company in the world.

• SpaceX is negotiating a multibillion-dollar compute deal with the Department of Defense, deepening a relationship that already spans launch, satellite communications, and missile tracking.

• The MLB has cut off the custom tab on dugout iPads, after a June 11 memo found teams using it for AI recommendations on substitutions and pitch calling. The block took effect as the second half of the season opened.

Some weeks the frontier moves. This week China ran a sequence. An open-source Chinese lab ships a model that beats the closed Western frontier. Hours later, 29 governments sign a China-led AI body. The next morning, President Xi takes the world stage to pitch a rival rulebook to half the planet. The message underneath all three: the open-weight, low-cost stack is not chasing the frontier anymore. It is the frontier.

In today’s AI news:

• Moonshot’s Kimi K3 pulls open weights up to the frontier

• Xi bids to lead the Global South’s AI order

• China’s compute independence goes physical at WAIC

• Meta caves and lists Muse Spark 1.1 on OpenRouter

🌝 Moonshot’s Kimi K3 pulls open weights up to the frontier

News: Moonshot AI ships Kimi K3, a 2.8-trillion-parameter open-weights model that lands within a few points of Claude Fable 5 and GPT-5.6 Sol, beats them on select tasks, and undercuts both on price. Full weights go public July 27, and the reception has been electric.

Details:

• 896 experts, 16 active per token: frontier-scale reasoning at the latency of a far smaller model, with a 1M-token context window and native vision.

• Builders are flocking to it fast: it tops third-party charts for front-end and web-interface generation, the exact use case that drives developer adoption.

• 93.5% on GPQA Diamond, 88.3% on Terminal-Bench 2.1 (Sol edges it at 88.8%), plus 91.2% on BrowseComp. Artificial Analysis scores it a 57, behind only Fable (60) and Sol (59).

• Priced at $3 / $15 per million tokens, a fraction of Western flagship rates and roughly half the cost per task of Opus 4.8.

• Alibaba-backed Moonshot is one of China’s “Six AI Tigers,” and K3 drops one month after Washington briefly pulled Anthropic’s Fable and Mythos.

Why it matters: Dario’s “6 to 12 months behind” line for China and open source now reads like one release cycle. But the benchmark is a distraction. The real weapon is the weights. Moonshot spends frontier-grade compute to give its model away, because owning the open-source developer base beats topping a leaderboard. Capability is commoditizing faster than any incumbent’s business model assumes. Here is my honest read after using Kimi K3 in the tooling: on front-end generation and agentic coding, K3 is the real deal, and at this price it is not close. Put it in your stack today and judge it for yourself.

🌐 Xi bids to lead the Global South’s AI order

News: Xi Jinping uses his first in-person WAIC keynote to cast China as the patron of open, shared AI, one day after 29 governments sign a China-led cooperation body in Shanghai. His pitch: China as the low-cost, open-source alternative to a Western stack it frames as exclusionary.

Details:

• Xi warns against creating “new historical injustices” in AI, a shot at U.S. chip and model controls without naming the U.S.

• The new World AI Cooperation Organization (WAICO) draws 29 signatories, headquartered in Shanghai, including Russia, Pakistan, and Indonesia.

• Read the talent play precisely: 5,000 training places in AI programs for developing countries over five years (not 5,000 researchers), plus cooperation centers with ASEAN, the Arab League, the African Union, CELAC, the SCO, and BRICS.

• Soft power: a “Mazu” AI weather-warning system pledged for 30 countries, with China’s open-source models held up as a global public good.

• Xi calls for AI to stay “secure and controllable”, landing as Washington and Beijing prep their first government-level AI talks under Trump.

Why it matters: Watch the sequence, not the speech. An open-source Chinese lab beats the closed Western frontier. Hours later, 29 governments sign a China-led AI body. The next morning, Xi addresses the world. China is not having a moment. It is running a play. Models, compute, and governance, bundled and offered to the Global South as an alternative to an order gated by export controls. Washington bets that controlling chips controls the trajectory. Xi’s counter: hand a country models, training, and norms, and the chips matter less. The question for anyone running a global builder community is which standards stack you are already building on.

🏗️ China’s compute independence goes physical at WAIC

News: Huawei rolls out physical Atlas 950 SuperPoD hardware for the first time at WAIC, the clearest sign yet that China’s frontier labs are re-platforming off Nvidia and onto homegrown Ascend silicon.

Details:

• The unit on the floor wires 1,024 Ascend NPUs across 16 cabinets at 1 EFLOPS FP8, 2 EFLOPS FP4, and 256TB of memory. A full SuperPoD scales to 8,192 NPUs and 8/16 FP8/FP4 ExaFLOPS; 64 of them form a 524-ExaFLOPS SuperCluster.

• Huawei claims 6.7x the compute and 15x the memory of Nvidia’s next-gen NVL144. The strategy is brute-force scale, not per-chip wins.

• The real re-platforming tell is not a quant format. MXFP4 is an open OCP standard that runs on Nvidia’s own Blackwell, not a Huawei flag. The tell is “day-zero” access: DeepSeek gives domestic chipmakers early tuning on V4 while locking Nvidia out of the window.

• Proof it works: Zhipu, cut off from Nvidia by the U.S. Entity List, trains GLM-5 on ~100,000 Ascend chips with zero Nvidia, landing within single digits of GPT-5.2 and Opus 4.5.

• The honest brake: per chip, Ascend’s 910C is ~80% of an H100, so Huawei wins on scale, not silicon, and even DeepSeek is now building its own inference chip.

Why it matters: None of this is a surprise. Wall off China from Nvidia’s best chips and you do not stop the compute, you force a sovereign stack into existence. The restriction did not starve China’s compute. It nationalized it. And notice who reaches it first: Alibaba, ByteDance, and Tencent, the same three clouds placing bulk Ascend orders. Early access plus domestic silicon is not just independence, it is consolidation, with the biggest players compounding their lead.

⚡ Meta caves and lists Muse Spark 1.1 on OpenRouter

News: Meta puts Muse Spark 1.1 on OpenRouter for U.S. developers, reversing the walled-garden stance it took at the model’s July 9 paid launch. The model itself is not new. The distribution move is.

Details:

• Multimodal reasoning built for agents: text, image, video, audio, and PDF in, orchestrating multi-agent workflows as planner or subagent, generalizing zero-shot to new tools, MCP servers, and skills.

• The reversal: Meta kept it off OpenRouter at launch, then flipped a week later, with Zuckerberg breaking a long X silence to announce it.

• Price, and the catch: $1.25 / $4.25 per million tokens with $20 in free credits, but reasoning tokens bill at the full output rate, so agentic runs cost more than the sticker.

• Reality check: independent tests land it ~10 points under Meta’s own Terminal-Bench claim, and 61.5 vs Opus 4.8’s 69.2 on SWE-Bench Pro, amid a benchmark-integrity dispute.

• The bigger arc: Meta’s first-ever paid model under AI chief Alexandr Wang, with a 10x-compute successor (codenamed Watermelon) already training.

Why it matters: For the company that made “open source as strategy” its entire identity, a paid, proprietary, region-gated API is a full philosophical U-turn, and caving to list on OpenRouter is Meta admitting builders will not walk into a walled garden. Cheap output tokens genuinely matter for agentic loops. Just price the real workload, not the sticker: reasoning-token billing plus a live gap to Opus-class models means the savings are thinner than “a quarter of the price.”

🛠️ Today’s Top Tools

• 🌝 Kimi K3: Moonshot’s 2.8T open-weights frontier model. Weights public July 27.

• ⚡ Muse Spark 1.1: Meta’s first paid API, now live on OpenRouter for U.S. devs.

• 🎥 Lucy 2.5: Decart’s live model that edits video at 30fps, 1080p, near-zero latency.

📰 Quick News

• Google’s Gemini 3.5 Pro slips months behind schedule over coding performance, per Bloomberg, sending Alphabet down as much as 4% days before its July 22 earnings.

• China’s cyberspace regulator approves Apple Intelligence for local iPhones, with Baidu and Alibaba’s Qwen powering the localized features.

• Roblox unveils Build, a mobile AI tab that turns text prompts into playable games, with public testing starting July 28.

• DoorDash launches dd-cli, a beta command-line tool that lets AI agents search restaurants, find deals, and check out with real payments.

• Google renames NotebookLM to Gemini Notebook and gives every notebook a secure cloud computer that writes and runs code against your sources.

Two of the most legendary minds in AI history just raised $1B each on the same day. And both are betting the current paradigm isn’t enough.

David Silver, the man behind AlphaGo, is building Ineffable Intelligence in London with a $4B valuation and a thesis that LLMs will never reach superintelligence. Fei-Fei Li’s World Labs just closed $1B to build AI that understands worlds, not just words. When researchers of this caliber walk away from the biggest labs on earth to pursue entirely new directions, pay attention. The frontier is shifting.

(And yes, Google just made AI music creation available to millions of people who’ve never heard of Suno...)

In today’s AI news:

• AlphaGo creator David Silver raises $1B for London AI lab

• Fei-Fei Li’s World Labs raises $1B for spatial intelligence

• Google embeds AI music creation directly into Gemini

• Today’s Top Tools + Quick News

🧠 AlphaGo Creator David Silver Raises $1B for London AI Lab in Europe’s Largest Seed Round Ever

News: Former Google DeepMind scientist David Silver is raising $1B for Ineffable Intelligence, a London-based startup led by Sequoia Capital at a roughly $4B pre-money valuation. Nvidia, Google parent Alphabet, and Microsoft are also reportedly in discussions to participate, making this Europe’s largest seed round in history.

Details:

• Silver spent over a decade at DeepMind, where he led development of AlphaGo, AlphaZero, and AlphaStar, reinforcement learning systems that surpassed humans by abandoning human knowledge entirely. He also contributed to Google’s Gemini family of LLMs before departing late last year.

• Ineffable Intelligence was incorporated in November 2025, with Silver appointed director in January 2026. The company has no product, no revenue, and no public roadmap.

• Silver’s core thesis: LLMs alone will never reach superintelligence. In a paper co-authored with Richard Sutton titled “Era of Experience,” he argued that AI systems trained on human data are hitting a wall and that progress through supervised learning alone is slowing. His bet is on reinforcement learning, where AI teaches itself through direct experience.

• Sequoia partners Alfred Lin and Sonya Huang reportedly traveled to London to meet Silver immediately after his exit, triggering intense competition among venture firms.

• Silver joins a wave of elite researchers leaving Big Tech to pursue superintelligence independently. Mira Murati’s Thinking Machines Lab raised a $2B seed round, while Ilya Sutskever’s Safe Superintelligence has raised $2B at a $32B valuation, all before shipping products.

Why It Matters: This is the clearest signal yet that the world’s most elite AI researchers believe the current LLM paradigm has a ceiling. Silver is making a fundamental bet that experience-based learning will unlock what human-generated training data never can. The fact that Sequoia, Nvidia, Google, and Microsoft are all rushing to write ten-figure checks for a pre-product company in London tells you everything about the talent premium in this market: conviction in the right founder now commands billion-dollar capital on thesis alone. For Europe, this is a landmark moment. The last London AI company to attract this level of investor frenzy was DeepMind itself, which Google acquired for ~$500M in 2014. If Silver is right that reinforcement learning breaks through the scaling wall, the implications ripple across every major AI lab on the planet.

🌐 Fei-Fei Li’s World Labs Raises $1B to Bring AI Into the 3D World

News: Fei-Fei Li’s World Labs has secured $1B in new funding, including a $200M strategic investment from Autodesk, with additional backing from AMD, Nvidia, Andreessen Horowitz, Emerson Collective, Fidelity, and Sea. The round accelerates World Labs’ mission to build “world models” that understand, generate, and reason about immersive 3D environments.

Details:

• World Labs emerged from stealth in 2024 with $230M at a $1B valuation. Reports from a month ago suggested the company was targeting a $5B valuation for this latest round.

• The company launched its first product, Marble, in November 2025. Marble generates persistent, high-fidelity 3D environments from text, images, video, or 3D layouts, with export options including meshes and video.

• The Autodesk partnership will explore how World Labs’ models can work alongside Autodesk’s industry-standard design tools, starting with entertainment use cases. Autodesk will also serve as an adviser and collaborate at the research and model level.

• Autodesk is simultaneously developing “neural CAD,” a generative AI model trained on geometric data that can generate working 3D models with an understanding of how designs function in the real world.

• World Labs plans to expand Marble into robotics and scientific discovery, moving beyond creative content generation into physical-world AI reasoning.

Why It Matters: Fei-Fei Li captured it perfectly: “If AI is to be truly useful, it must understand worlds, not just words.” That single line explains why $1B just flowed into spatial intelligence. Most AI investment is still chasing better language models. World Labs is betting the next frontier is generating 3D environments governed by real physics, geometry, and dynamics. The Autodesk partnership is the strategic unlock. It gives World Labs a direct channel into the professional workflows of architects, engineers, filmmakers, and game designers who collectively represent trillions in economic output. When you combine AI-native world generation with Autodesk’s install base across every creative and industrial vertical, you get a picture of what “spatial intelligence” looks like at scale.

🎵 Google Embeds AI Music Creation Directly Into Gemini

News: Google just launched Lyria 3 inside Gemini, DeepMind’s newest AI music generation model that lets users transform a text prompt, photo, or video into a fully customized 30-second track complete with auto-generated lyrics and cover art.

Details:

• The model handles everything from genre and tempo to vocal style and full lyrics autonomously, with users able to feed text, images, or video as creative input. Lyria 3 produces 48kHz stereo tracks with significantly more musical complexity than previous generations.

• DeepMind has been developing Lyria since 2023, but this marks the first time the technology reaches mainstream consumers through the Gemini app.

• Every generated track is tagged with Google’s SynthID watermark, and Gemini now lets users upload any audio file to verify whether it was AI-created.

• YouTube creators gain global access to Lyria 3 through Dream Track for Shorts, enabling customized audio tracks to pair with video content.

Why It Matters: AI music platforms like Suno and Udio have quietly gotten scary good, producing full tracks that fool most listeners. But they’ve remained locked in a niche. Embedding Lyria directly inside Gemini changes the game: Google is putting AI music creation one prompt away for millions of users who’ve never touched a dedicated music model. For content creators and marketers in your audience, the competitive landscape for AI audio just shifted overnight. The distribution advantage Google holds here is enormous.

🛠️ Today’s Top Tools

• 🎶 Lyria 3 - Google’s new AI music generation model now live inside Gemini

• 🗣️ Phoenix-4 - Tavus’ real-time human avatar AI with genuine emotional intelligence and sub-600ms latency

• ⚡ Grok 4.20 - xAI’s upgraded 4-agent collaboration model with native multi-agent reasoning, 2M token context, and the only AI profitable in live stock trading competitions

📰 Quick News

• OpenAI hired Meta’s Charles Porch as VP of global creative partnerships, poaching Instagram’s 15-year celebrity dealmaker to win over a skeptical Hollywood following its $1B Disney deal in December.

• Tavus launched Phoenix-4, a real-time avatar rendering model with 10+ emotional states, active listening behavior, and sub-600ms conversational latency, backed by a $40M Series B led by CRV and Sequoia.

• Perplexity is abandoning advertising on its AI search platform entirely, with executives telling the Financial Times that sponsored content undermines trust in AI-generated answers. The company reached $200M in annual recurring revenue betting on subscriptions instead.

• OpenAI acqui-hired enterprise AI startup Nerve, whose “AI chief of staff” product connected data sources and internal tools to AI agents, with the team joining to scale ChatGPT’s search and agentic capabilities.

• The trailer for ‘The AI Doc: Or How I Became an Apocaloptimist’ is going viral ahead of its March 27 theatrical release, featuring sit-downs with Sam Altman, Dario Amodei, and Demis Hassabis from the Academy Award-winning team behind Navalny and producers of Everything Everywhere All at Once.

Anthropic spent the last two weeks shipping its strongest models ever. But here’s the part most people are missing: the cheaper one might be the bigger deal.

The new Sonnet 4.6 goes toe-to-toe with Opus 4.6 across coding, finance, and computer use benchmarks at 1/5 the cost. That’s Anthropic collapsing the gap between what the best AI can do and what most companies can actually afford to deploy at scale.

(And yes, Apple is quietly fast-tracking three AI wearables that could redefine what Siri becomes in 2027...)

In today’s AI news:

• Anthropic’s powerful Claude Sonnet 4.6

• Meta and Nvidia lock in a massive multiyear AI chip deal

• Apple going all-in on AI wearables

• Mistral makes its first-ever acquisition

• Figma turns Claude Code builds into editable designs

• Today’s Top Tools + Quick News

🧠 Anthropic Ships Claude Sonnet 4.6: Near-Flagship Power at 1/5 the Price

News: Anthropic just released Claude Sonnet 4.6, its newest mid-tier model that matches or outperforms the flagship Opus 4.6 across finance, computer use, coding, and office benchmarks at 1/5 the cost. The model features a 1M token context window in beta and is now the default across claude.ai for all users, including the free tier.

Details:

• Priced at $3/$15 per million tokens vs. Opus’s $15/$75, Sonnet 4.6 scored 79.6% on SWE-Bench Verified for coding, just below Opus 4.6’s 80.8%, the smallest Sonnet-to-Opus gap in any Claude generation. For companies running AI agents that make thousands of API calls per day, that pricing gap rewrites the unit economics of automation.

• Early Claude Code testers preferred Sonnet 4.6 over its predecessor 70% of the time, and it beat the previous flagship Opus 4.5 at a 59% preference rate. Developers reported less overengineering, fewer hallucinations, and stronger instruction following.

• Computer use capabilities continue their steep climb: Sonnet’s OSWorld scores jumped from 14.9% in October 2024 to 72.5% today, nearly quintupling in 16 months. Enterprise customers are already migrating: Pace reported 94% accuracy on its insurance computer use benchmark, and Box saw a 15-point improvement in heavy reasoning Q&A.

• Sonnet 4.6 outperformed Opus 4.6 on agentic financial analysis and office-task benchmarks, a first for any mid-tier Claude model. The mid-tier is no longer a compromise. It’s the default.

Why It Matters: This is absolutely insane. I’ve been writing for months about how Chinese models keep delivering similar performance benchmarks at a fraction of the price, and Anthropic just answered with the exact same playbook: near-flagship intelligence, 80% cheaper, shipped two weeks after the flagship upgrade. But the number that should make every builder pay attention is the 59% preference rate over Opus 4.5. That’s not “almost as good.” That’s testers choosing the cheaper model over last month’s best by a significant margin. Anthropic just keeps shipping, and they’re not slowing down. If you’re building on AI, your cost-to-deploy just fundamentally changed.

💰 Meta and Nvidia Lock In Massive Multiyear AI Chip Deal

News: Nvidia announced a multiyear, multigenerational strategic partnership with Meta spanning millions of GPUs and CPUs across on-premises, cloud, and AI infrastructure. The deal, estimated at tens of billions of dollars, makes Meta the first Big Tech company to deploy Nvidia’s standalone Grace CPUs at scale.

Details:

• Meta will deploy millions of Nvidia Blackwell and next-gen Rubin GPUs, alongside Nvidia’s Arm-based Grace CPUs as standalone chips and Spectrum-X Ethernet switches integrated into Meta’s Facebook Open Switching System platform.

• The standalone Grace CPU deployment is a first for any hyperscaler at this scale, with Nvidia claiming 2x performance per watt on backend datacenter workloads. The CPUs run inference and agentic workloads as companions to GPU racks. Meta is already testing Nvidia’s upcoming Vera CPUs for large-scale rollouts from 2027.

• The partnership is part of Meta’s broader $600 billion U.S. commitment by 2028 on data centers, with plans for 30 facilities (26 in the U.S.), including the 1-gigawatt Prometheus site in Ohio and the 5-gigawatt Hyperion site in Louisiana.

• Meta’s 2026 AI capex target sits at $115 billion to $135 billion. Back-of-the-napkin math: a million GPUs at $3.5 million per rack works out to roughly $48 billion for Nvidia alone.

Why It Matters: This is history in the making. Meta becomes the first hyperscaler to deploy Grace CPUs as standalone chips, operating on the cutting edge of what’s possible in AI infrastructure. The scale is unprecedented: $600 billion committed to data centers by 2028, $135 billion in capex this year alone, and the critical thing most people are missing is that Meta actually has the revenue and profitability to back it up. This is Zuck signaling three things at once: inference over training, crazy amounts of scale, and the stated goal of building everyone’s personal superintelligence. That requires infrastructure that doesn’t exist yet. This deal is the down payment.

🕶️ Apple Fast-Tracks Three AI Wearables to Give Siri Eyes and Ears

News: Apple is accelerating development on three camera-equipped AI wearables, according to Bloomberg’s Mark Gurman. Smart glasses (code-named N50), a pendant device, and new AirPods are all designed to feed Siri real-time visual awareness through the iPhone, marking what could be Apple’s most significant wearables pivot since the Apple Watch.

Details:

• The smart glasses feature dual cameras (one high-resolution for photo/video, one for computer vision), Apple-designed premium frames, no display, and a production target of December 2026 ahead of a 2027 launch. Apple chose to design its own frames rather than partner with established eyewear brands.

• The pendant device acts as an always-on camera and microphone for your iPhone, internally dubbed the phone’s “eyes and ears.” It includes a dedicated chip, can be clipped to clothing or worn as a necklace, and is designed as an iPhone accessory, not a standalone product.

• Camera-equipped AirPods could ship as early as September 2026, using infrared low-res sensors to give Siri live visual context. Development is further along than the pendant.

• All three devices connect into Apple’s revamped Siri for iOS 27, expected to gain a chatbot-style interface this year, powered by AI models co-developed with Google’s Gemini.

Why It Matters: Now the $2 billion Q.ai acquisition from last month makes sense. Q.ai‘s technology reads facial micro-movements through headphones and glasses to detect speech, emotions, and intent, exactly the kind of sensing layer these wearables need. We’ve all been wondering why Apple has been dropping the ball on AI (sorry Apple, but you know it’s true), and this roadmap finally reveals a coherent strategy: own the hardware interface between AI and the physical world, and let Google’s Gemini handle the intelligence layer underneath. The AI itself still needs to be better, but Apple is carving a lane for itself that no one else can match: 1.5 billion devices, premium hardware DNA, and now a wearable form factor designed from the ground up for ambient AI.

🇫🇷 Mistral Makes Its First-Ever Acquisition with Koyeb

News: French AI startup Mistral AI confirmed its first-ever acquisition, buying Paris-based serverless cloud platform Koyeb for an undisclosed amount. The deal folds Koyeb’s 13 employees and three co-founders into Mistral’s engineering team and signals a push to become a full-stack AI provider, not just a model maker.

Details:

• Koyeb’s serverless platform becomes a “core component” of Mistral Compute, Mistral’s AI cloud offering launched in June 2025. The technology enables developers to deploy AI applications without managing infrastructure, with autoscaling and isolated sandboxed environments for AI agents.

• The acquisition complements Mistral’s $1.4 billion data center investment in Sweden announced days earlier, where the company is deploying 40 MW of capacity and 18,000 Nvidia Blackwell GPUs as the starting point for its sovereign European AI infrastructure.

• Mistral recently passed $400 million in annual recurring revenue and is targeting $1 billion for 2026. The company was valued at roughly €11.7 billion after raising a €1.7 billion round led by ASML last September.

• Koyeb had raised $8.6 million total. Founded in 2020 by three former Scaleway employees, the platform already supported Mistral model deployments and recently launched Koyeb Sandboxes for isolated AI agent execution.

Why It Matters: When an AI company starts acquiring others, it’s a strong signal they’ve graduated from research lab to real business. This is the same playbook the Western hyperscalers run: own the entire stack from models to compute to deployment. Mistral is executing it from Europe, with $400M ARR, a $1B revenue target for 2026, and now a $1.4 billion infrastructure commitment in Sweden. For the EU’s AI champion, this deal confirms the ambition extends far beyond model weights. Mistral is building sovereign European AI infrastructure that gives global founders a credible non-U.S. alternative for compliance and data sovereignty.

🎨 Figma Bets On Design Layer for AI-Coded Products

News: Figma launched “Code to Canvas” in partnership with Anthropic, a new integration that converts interfaces built in Claude Code into fully editable design files on Figma’s canvas. CEO Dylan Field framed it as a bet that “the design canvas is better at navigating lots of possibilities than prompting in an IDE.”

Details:

• The feature captures live UI from any browser (production, staging, or localhost) and transforms it into native Figma layers. Users type “Send this to Figma” in Claude Code, and the rendered state translates into editable, annotatable frames.

• Figma’s MCP server closes the loop: developers pull edited designs back into coding environments without losing context. Teams can capture entire multi-step flows at once, preserving full user journeys for side-by-side review.

• The launch arrives amid the “SaaSpocalypse”: a sector-wide selloff that has erased nearly $1 trillion in software stock market value in early 2026. Figma stock has dropped roughly 85% from its 52-week high of $142.92. The company reports Q4 earnings on February 18.

Why It Matters: Figma is positioning itself as the “last mile” between raw AI-coded prototypes and production-ready design. It’s a logical play: AI tools build functional UIs in seconds, but someone still needs to decide which version ships. The problem is that model capabilities only keep improving, and that polishing layer gets automated next. The 85% stock decline and $1 trillion SaaS wipeout signal the market already prices this in. If you run or invest in SaaS companies, the Figma story is a preview of every software company’s existential question in 2026: what happens when the AI does 90% of the work?

🛠️ Today’s Top Tools

• 🧠 Claude Sonnet 4.6 - Anthropic’s upgraded mid-tier model with a 1M token context window, near-flagship performance, and $3/$15M token pricing

• 🎨 Recraft V4 - New image AI built for typography and production-level design outputs

• 🌎 Tiny Aya - Cohere’s open-weight 3.35B parameter model covering 70+ languages with regional variants. Runs offline on laptops

📰 Quick News

• xAI began rolling out Grok 4.20 in a public beta, introducing a 4-agent collaboration system where specialized AI agents (Grok, Harper, Benjamin, and Lucas) work in parallel to research and execute tasks. Available to SuperGrok ($30/month) and X Premium+ subscribers with a 2M token context window.

• Cohere Labs open-sourced Tiny Aya, a 3.35B parameter multilingual model handling 70+ languages with regional variants (Earth for Africa, Fire for South Asia, Water for Asia-Pacific/Europe). Trained on just 64 H100 GPUs, it runs offline on everyday laptops.

• WordPress launched a new AI assistant capable of editing layouts, generating images via Google Gemini’s Nano Banana models, and rewriting content directly inside the editor. Available on Business and Commerce plans with @ai commands in the block notes editor.

The Pentagon is about to slap Anthropic with a “supply chain risk” label, a designation typically reserved for foreign adversaries, not the company whose AI is the only model running on its classified systems.

This is no longer a policy disagreement. This is a power struggle over who controls frontier AI in warfare: the labs building the models, or the governments deploying them. And the answer to that question will shape AI governance for the next decade.

(And yes, India just hosted the largest AI summit in history with $200 billion in investment commitments, Alibaba dropped an open-weight model nipping at GPT-5.2’s heels, and Meta patented an AI that posts for you after you die...)

In today’s AI news:

• Anthropic-Pentagon feud escalates to potential “supply chain risk” designation

• India hosts the largest AI summit ever with $200B+ in investment interest

• Alibaba’s Qwen-3.5 closes in on frontier with open weights and 60% cost cuts

• OpenAI rolls out new “Lockdown Mode” for ChatGPT security

• Meta patents AI to simulate deceased users’ social media activity

• Today’s Top Tools + Quick News

‼️ Anthropic Faces Pentagon Blacklist Over Military AI Restrictions

News: Defense Secretary Pete Hegseth is “close” to cutting business ties with Anthropic and designating the company a “supply chain risk” over its firm stance on restricting how Claude is used in military operations. The move, if finalized, would force every U.S. defense contractor to certify they don’t use Claude in their own workflows, effectively blacklisting one of the most capable AI labs on the planet.

Details:

• The “supply chain risk” designation forces all U.S. defense contractors to stop working with the flagged entity. A senior Pentagon official told Axios: “It will be an enormous pain in the ass to disentangle, and we are going to make sure they pay a price for forcing our hand like this.”

• Defense officials are demanding the right to deploy AI for “all lawful purposes,” while Anthropic is holding firm, specifically pushing back on use cases involving domestic surveillance and autonomous weapons systems.

• Claude is currently the only AI model on the Pentagon’s classified systems, and was deployed via Palantir during the capture of Nicolás Maduro in January. Anthropic’s contract with the Pentagon is valued at up to $200 million.

• The Pentagon is simultaneously negotiating with OpenAI, Google, and xAI, all of which have agreed to remove safeguards for unclassified military use, making Anthropic the sole holdout among frontier labs.

Why It Matters: This is the most consequential AI governance standoff we’ve seen. Anthropic is the only frontier lab refusing to grant the military blanket permissions, and the Pentagon is preparing to make an example of them.

And you know what? Good for Anthropic. They don’t need the $200 million. They are running the most capable AI on the planet right now, and the proof is that Claude is the only model inside classified Pentagon systems. Every other frontier lab has already agreed to remove safeguards for military use. Anthropic is the sole holdout. Anthropic has always positioned itself as the safety and ethics lab. Now, when push comes to shove, they’re willing to lose a government contract and risk being shut out of the entire federal market rather than hand over blanket permissions for mass surveillance and autonomous weapons. They’re dying on this hill, and I respect the hell out of them for it. With $14 billion in annual revenue, eight of the ten largest U.S. companies already on Claude, and quickly growing market share, they’ll be just fine. This is Anthropic putting its money where its mouth has always been.

🇮🇳 India Hosts the Largest AI Summit in History as Global Giants Bet Big on the Subcontinent

News: India kicked off the AI Impact Summit 2026 in New Delhi this week, drawing heads of state from 20+ nations, delegations from 100+ countries, and virtually every major AI CEO on the planet, including OpenAI’s Sam Altman, Google’s Sundar Pichai, Anthropic’s Dario Amodei, and Meta’s Yann LeCun. With 250,000+ registered attendees and over 300 exhibitors at Bharat Mandapam, it is the largest global AI summit ever held, and the first hosted in the Global South.

Details:

• Sam Altman confirmed that India now has 100 million weekly active ChatGPT users, making it the second-largest market after the U.S., with Indian students representing the largest student user base globally.

• Anthropic officially opened its first India office in Bengaluru, its second in Asia after Tokyo, with CEO Dario Amodei stating that “India’s adoption is even more extreme compared with the rest of the world.”

• IT Minister Ashwini Vaishnaw announced the summit has generated $17 billion in VC funding commitments and $200 billion in infrastructure investment interest, while the government earmarked $1.1 billion for a state-backed venture capital fund targeting AI and deep-tech startups.

• Microsoft, Amazon, and Google have already committed a combined $68 billion in AI and cloud infrastructure investment in India through 2030, while Blackstone acquired a majority stake in Indian AI startup Neysa as part of a $600 million equity fundraise.

Why It Matters: India is cementing itself as the third pole in the global AI race behind the U.S. and China, and this summit made that undeniable. $17 billion in VC commitments, $200 billion in infrastructure investment interest, and every major AI CEO on stage in New Delhi is a declaration that the market has arrived. What makes India different from previous summits at Bletchley Park, Seoul, and Paris is the focus: not safety frameworks, but deployment at population scale. With 100 million weekly active ChatGPT users and massive infrastructure commitments from Microsoft, Amazon, Google, and now Blackstone, the economics of AI are about to shift dramatically. When you can deploy AI at scale across a population of 1.4 billion people, the unit economics change for every product, every platform, and every business model built on top of these models. India adopting AI at this pace will reshape the global market faster than most founders realize.

🧠 Alibaba’s Qwen-3.5 Arrives at the Frontier with Open Weights and 60% Lower Costs

News: Alibaba just released Qwen3.5-397B-A17B, an open-weight vision language model with a hybrid architecture that Alibaba claims outperforms GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on 80% of evaluated benchmarks, while running at a fraction of the cost.

Details:

• Qwen-3.5 uses a sparse Mixture of Experts (MoE) design, activating only 17B parameters out of 397B per query, delivering frontier-level capability with dramatically lower latency and compute costs.

• Alibaba claims Qwen-3.5 is 60% cheaper to run and delivers 8.6x to 19x faster decoding throughput compared to its predecessor, Qwen3-Max.

• The model is natively multimodal, trained via early fusion on trillions of text and image tokens simultaneously, supports 201 languages, and ships under an Apache 2.0 open-source license.

• The release targets agentic computing: Qwen-3.5 can analyze UI screenshots, detect elements, and execute multi-step tasks autonomously across mobile and desktop applications.

Why It Matters: Chinese AI labs continue to deliver phenomenal releases in 2026, and these benchmarks are genuinely shocking. Qwen-3.5 claims to outperform GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on 80% of evaluated benchmarks while running at 60% lower cost with open weights. That alone would be significant. But here’s what keeps running through my mind: Nvidia and the Chinese government just agreed to an $11 trillion chip deployment deal, the largest shipment to China in over a decade. If Chinese labs are already competing with Western labs at a fraction of the cost and matching performance before that compute lands, there is a real concern in the Western market for how this plays out over the next few months to years. The frontier is no longer a walled garden, and the cost advantage is widening, not shrinking.

🔒 OpenAI Introduces “Lockdown Mode” to Harden ChatGPT Security

News: OpenAI just launched “Lockdown Mode” for ChatGPT, alongside new Elevated Risk labels, creating a deterministic security layer specifically designed to protect high-value users from prompt injection attacks where AI is manipulated into leaking sensitive data.

Details:

• Lockdown Mode deterministically disables ChatGPT tools and capabilities an attacker could exploit, going beyond probabilistic safety measures to hard blocks that cannot be circumvented.

• The mode limits web browsing to cached content only, ensuring no live network requests leave OpenAI’s environment. Deep Research and Agent Mode are fully disabled; ChatGPT responses cannot include images.

• Available for ChatGPT Enterprise, Edu, Healthcare, and Teachers, with workspace admins able to toggle the mode and whitelist specific apps. OpenAI plans to bring it to consumers in the coming months.

• New “Elevated Risk” labels now appear across ChatGPT, Atlas, and Codex to flag features that carry any level of security exposure, giving organizations full visibility before enabling connected capabilities.

Why It Matters: I find this one pretty ironic. OpenAI effectively dismantled its internal safety and ethics teams, lost nearly every senior researcher involved in alignment work, and was hiring a head of AGI preparedness just two weeks ago. Now their answer to the growing security surface area of agentic AI is a toggle switch called “Lockdown Mode.” It’s a lazy solution to a serious problem. That said, it’s also probably a feature every lab should have shipped yesterday. Deterministic hard blocks are the right technical approach. But the fact that this is coming from the lab that gutted its own safety infrastructure makes it hard to take seriously as a principled move rather than a product patch.

👻 Meta Patents AI to Simulate Dead Users’ Social Media Activity

News: Meta was granted a patent in late December 2025 for an AI system that trains a large language model on a user’s historical social media activity to simulate their presence after death or extended absence. The patent, first filed in 2023 and authored by CTO Andrew Bosworth, describes technology capable of generating posts, likes, comments, DMs, and even simulated audio and video calls on behalf of deceased or inactive users.

Details:

• The system trains on “user-specific” data, including past posts, comments, likes, voice messages, and messaging patterns, to reconstruct a digital persona that mirrors the individual’s established online behavior across Facebook and Instagram.

• Meta’s filing states the impact is “much more severe and permanent” when a user dies and can never return to the platform, framing the technology as a way to soften that absence for followers and loved ones.

• Meta told Business Insider it has “no plans to move forward” with the technology, a standard caveat companies use when filing defensive patents.

• University of Cambridge researchers have warned this category of “grief tech” is an ethical minefield, with concerns that griefbots could be used to advertise products or distress children by insisting a dead parent is still present.

Why It Matters: This is the story to watch closely. Meta was granted this patent just two months ago, but the filing process started back in 2023, authored by CTO Andrew Bosworth himself. And yet Meta told the press they have “no plans to move forward” with the technology. I don’t buy that for a second. This also connects directly to yesterday’s news: Simile raised $100 million to replicate human behavior through AI agents for product research, test groups, and social media interaction analysis. The market is clearly moving toward digital human replication at scale, and the public reception has not been warm. The biggest concern here is the emergence of “grief tech,” products that allow users to continue interacting with deceased loved ones through AI-generated posts, messages, and even simulated calls. There is implicit value in preserving memory and connection. But there is also something deeply unsettling about reconstructing a person’s entire digital consciousness after they’ve passed and making it feel like they never left. When every AI-generated like and DM from a deceased user still drives platform engagement and ad revenue, the line between preserving memory and exploiting data disappears fast. This deserves much more attention than it’s getting.

🛠️ Today’s Top Tools

• 🧠 Qwen3.5-397B-A17B - Alibaba’s open-weight vision language model rivaling GPT-5.2 at 60% lower cost, Apache 2.0 licensed

• 🤖 Manus Agents - Full AI agent capabilities now live in Telegram with long-term memory, tool integrations, and multi-step task execution across all subscription tiers

• ✨ Seed 2.0 - ByteDance’s next-gen model suite with multimodal understanding, agentic capabilities, and performance rivaling GPT-5.2 at ~10x lower cost

📰 Quick News

• Blackstone acquired a majority stake in Indian AI startup Neysa as part of a $600 million equity raise, with plans to deploy 20,000+ GPUs and raise another $600 million in debt. Adani announced a $100 billion commitment to renewable-powered AI data centers in India by 2035.

• Ireland’s Data Protection Commission opened a “large-scale” GDPR investigation into X over Grok generating non-consensual sexualized deepfake images of adults and children, joining parallel probes from the UK’s ICO, France, and the European Commission.

• ElevenLabs launched “ElevenLabs for Government” to help public sector agencies deploy secure, multilingual voice and chat AI solutions. Already live in Ukraine, Czech Republic, and Midland, Texas, handling ~5,000 calls/day with 85% independent resolution.

• You.com co-founders Richard Socher and Bryan McCann released 35 predictions for 2026, including that an “AI winter” is approaching as the LLM revolution gets “mined out,” “reward engineering” becomes a new job category, and 10-person unicorns become possible through AI-powered efficiency.

• AMD and TCS announced a partnership to deploy AMD’s “Helios” rack-scale AI architecture in India, with a 200MW AI-ready data center blueprint powered by Instinct MI455X GPUs, targeting sovereign AI factories and enterprise-scale deployments.

OpenAI’s GPT-5.2 just autonomously discovered that a widely accepted answer in particle physics was wrong, proposed the correct solution, and wrote the formal proof in 12 hours flat.

The “can AI actually think?” debate isn’t disappearing, but the real conversation has shifted — from whether AI can contribute to science to how quickly it rewrites what we thought we already knew. Meanwhile, the agent wars officially kicked off, China’s pricing pressure just broke Western labs’ comfort zone, and the Pentagon drew a line in the sand with the AI industry’s most safety-conscious company.

Here’s today’s AI news:

• OpenClaw creator joins OpenAI to build the next wave of personal agents

• ByteDance’s Seed 2.0 redraws the frontier map at 1/10th the price

• GPT-5.2 makes a theoretical physics discovery

• Anthropic’s $200M Pentagon deal hangs in the balance

• Simile raises $100M to simulate human behavior

• Today’s Top Tools + Quick News

🤖 OpenClaw creator Peter Steinberger joins OpenAI

News: Peter Steinberger, creator of the VIRAL open-source AI agent OpenClaw, has joined OpenAI to build the next generation of personal agents. OpenClaw will transition into an independent foundation while remaining open source — a move Altman called critical for a “multi-agent future.”

Details:

• OpenClaw — originally called Clawdbot, then Moltbot after Anthropic’s trademark complaint — went viral in late January 2026, amassing nearly 196,000 GitHub stars and 10,000 commits from 600 contributors in under 2 weeks, making it one of the fastest-growing open-source projects in history.

• Steinberger said he could see OpenClaw becoming a huge company, but that wasn’t exciting to him — he spent 13 years building one already. His stated mission: “build an agent simple enough for his mother to use, which requires access to frontier models and safety research”

• Altman called Steinberger “a genius with amazing ideas about the future of very smart agents interacting with each other,” adding that agents will quickly become core to OpenAI’s product offerings.

• OpenClaw will live in a foundation as an open-source project that OpenAI will continue to support.

Why It Matters: This is a fumble of the century for Anthropic and a masterclass in opportunism from OpenAI. Steinberger built this project for Claude, named it after Claude, and was actively driving revenue and developer mindshare to Anthropic’s API. Instead of recognizing what they had — an unpaid evangelist building the most viral agent ecosystem in history on top of their model — Anthropic sent lawyers. Meanwhile, OpenAI played this perfectly. A company that has taken relentless criticism for having “Open” in its name while shipping only closed models just attached itself to the fastest-growing open-source project in AI history — and got an open-source foundation in the deal. Consider the context: Meta just spent $2 billion acquiring Manus to jumpstart their agent strategy, and OpenAI has been losing ground in traffic and the AI coding race to Claude Code and Cursor. By hiring Steinberger and sponsoring the OpenClaw foundation, OpenAI hijacked onto exponential community growth at a fraction of Meta’s price — and reset the “open vs. closed” narrative in one move. Very well played.

🌱 ByteDance’s Seed 2.0 redraws the frontier map at 1/10th the price

News: ByteDance released Seed 2.0, a new model family that matches or outperforms GPT-5.2 and Gemini 3 Pro across dozens of benchmarks at roughly one-tenth the cost. The launch caps a week where ByteDance simultaneously ignited a full-blown Hollywood copyright crisis with its Seedance 2.0 video model — drawing cease-and-desist letters from Disney, Paramount, and formal condemnation from the Motion Picture Association and SAG-AFTRA.

Details:

• Seed 2.0 Pro tops GPT-5.2 ($1.75/M input tokens) and Gemini 3 Pro ($5/M) across math, reasoning, and vision benchmarks at just $0.47/M input tokens. The model scored gold medal-level results in all five ICPC programming competitions tested and reached 35/42 on the International Mathematical Olympiad, outperforming both Western competitors.

• The model series spans three sizes (Pro, Lite, Mini) plus a dedicated code model, all engineered for real-world agentic tasks. Demos show it autonomously completing 96-step CAD modeling workflows, signaling ByteDance’s push into enterprise agent use cases.

• Meanwhile, Seedance 2.0 went viral when Irish filmmaker Ruairi Robinson generated a cinematic fight between Tom Cruise and Brad Pitt from a two-line prompt. Within a single day, the MPA called it “unauthorized use of U.S. copyrighted works on a massive scale.” Disney sent a cease-and-desist accusing ByteDance of a “virtual smash-and-grab” of its IP. Paramount followed with its own letter citing infringement of South Park, Star Trek, The Godfather, and more.

• ByteDance responded that it “respects intellectual property rights” and is strengthening safeguards, but has not pulled the model. Seed 2.0 is live on ByteDance’s Doubao app and via Volcano Engine API, though consumer access outside China remains limited.

Why It Matters: ByteDance is now running a two-front assault on the Western AI landscape that no single company has attempted before. On the model side, Seed 2.0 at one-tenth the price forces every founder and CTO to ask the question they’ve been avoiding: if a Chinese model matches your frontier provider at 90% less cost, what exactly are you paying the premium for? On the content generation side, Seedance 2.0 just stress-tested every IP protection framework Hollywood has built — and they all failed in under 24 hours. Whether you’re making model selection decisions or watching the copyright landscape evolve, ByteDance just accelerated every timeline simultaneously.

🔬 GPT-5.2 rewrites 40 years of assumptions in 12 hours

News: OpenAI published a research preprint where GPT-5.2 independently discovered that a class of gluon interactions physicists had assumed were impossible can actually occur under specific conditions — then proposed the correct mathematical formula and formally proved it. The paper, co-authored with researchers from the Institute for Advanced Study, Harvard, Cambridge, and Vanderbilt, represents what the company is calling AI’s first original contribution to theoretical physics.

Details:

• The preprint, titled “Single-minus gluon tree amplitudes are nonzero,” overturns decades of textbook consensus. Standard arguments held that single-minus gluon amplitudes must vanish at tree level; GPT-5.2 identified a specific mathematical regime — the half-collinear regime — where the reasoning breaks down and the amplitude is non-zero.

• Human physicists first calculated the amplitudes for small values by hand, producing superexponentially complex expressions. GPT-5.2 Pro dramatically simplified those expressions, spotted a recursive pattern across the base cases, and conjectured a general formula valid for any number of particles.

• An internal scaffolded version of GPT-5.2 then spent roughly 12 hours reasoning through the conjecture and produced a formal proof, which was verified analytically against the Berends-Giele recursion relation — a standard validation method in particle physics.

• UC Santa Barbara physicist Nathaniel Craig called the result “journal-level research advancing the frontiers of theoretical physics” and said his research group is already exploring its implications. Harvard physicist Andrew Strominger, a co-author, noted the AI pursued a path no human would have attempted.

• The methodology is already being extended to graviton amplitudes — particles that mediate gravitational force — opening the door to implications well beyond the strong nuclear force.

Why It Matters: AI is falsifying a 40-year assumption that the entire field accepted as settled. The methodology is what makes this transformative: humans defined the problem, AI simplified the complexity beyond what any human could hold in their head, spotted a pattern across impossibly complex expressions, and then proved its own conjecture. If AI can independently overturn accepted conclusions in theoretical physics, every industry built on expert consensus — from drug discovery to financial modeling to engineering standards — is on notice. The template this paper establishes (human-defined problem → AI-driven discovery → human verification) is likely how frontier research gets done from here forward.

🧬 Simile raises $100M to simulate human behavior

News: AI startup Simile raised $100 million in funding led by Index Ventures, with participation from Bain Capital Ventures, A*, Hanabi Capital, and notable AI figures including Fei-Fei Li and Andrej Karpathy. The company builds AI simulations populated by agents based on real humans to predict purchasing decisions, earnings call questions, and reactions to corporate announcements or product launches.

Details:

• Simile was founded by the Stanford team behind the landmark 2023 “Generative Agents” research — Joon Sung Park, whose paper introduced a virtual town called Smallville where 25 AI agents autonomously formed relationships, spread invitations, and organized a Valentine’s Day party without any human instruction. The paper’s citation count crossed 10,000 before Park had even finished his PhD.

• The founding team is a Stanford dream roster: three professors — including the creators of the Generative Agents paper, the ImageNet benchmark, and the HELM evaluation framework — plus Lainie Yallen, a former early Hebbia operator who helped scale revenue 15x before joining as the commercial co-founder.

• The company emerged from stealth after seven months developing an AI model trained on interviews with hundreds of people about their lives, historic transaction data, and text from scientific journals on behavioral experiments. Early reports indicate CVS is already testing the platform to optimize product stocking and display decisions.

• Leading companies are using Simile to rehearse earnings calls, model litigation outcomes, and test policy changes, with the company’s long-term vision of simulating entire worlds of interacting decisions across individuals, organizations, and cultures.

Why It Matters: This is the most credentialed team you could assemble for this problem — the researchers who literally invented generative agents turning their academic breakthrough into a commercial platform. What makes Simile different from the sea of AI agent startups is the specificity of the bet: they’re not building general-purpose agents, they’re building digital twins of real human decision-making trained on actual interviews and behavioral data. If it works at scale, this replaces not just focus groups but entire categories of market research, policy testing, and strategic planning. Companies that learn to simulate their customers before launching will have a structural advantage over those still relying on gut instinct and post-hoc surveys.

TOOLS

• Kimi Claw — Deploy OpenClaw agents to the cloud through Moonshot AI’s platform. Includes 5,000+ community skills and 40GB cloud storage. Best for teams looking to scale agent workflows without managing local infrastructure.

• FireRed-Image-Edit — New open-source image editing model from Xiaohongshu delivering state-of-the-art results on multiple benchmarks. Worth watching if you’re building with open-source image pipelines or evaluating alternatives to proprietary editing APIs.

• Gemini 3 Deep Think — Google’s upgraded reasoning mode for complex problem-solving, scoring 48.4% on Humanity’s Last Exam and 84.6% on ARC-AGI-2. Designed for tasks where accuracy matters more than speed — think multi-step analysis, research synthesis, and strategic planning.

QUICK NEWS

• Spotify CEO Gustav Soderstrom revealed the company’s top developers haven’t written a single line of code this year, calling them “all in” on the AI transition. The company’s internal system Honk, built on Claude Code, lets engineers deploy features from Slack on their phones.

• Alpha School shared test results showing its 2-hour, AI-first academic model has students scoring in the 99th percentile across virtually every grade and subject — a data point that will fuel the growing debate over whether the traditional school day is built for learning or for logistics.

Ps, want to read the FULL article? Check it out here.

⚡ Google’s Deep Think Crushes Every Reasoning Benchmark in Sight

News: Google just shipped a major upgrade to its Gemini 3 Deep Think reasoning mode, posting dominant scores across math, coding, and science while simultaneously unveiling Aletheia, an Olympiad-level math research agent powered by the new upgrade.

Details:

• Deep Think hit 84.6% on ARC-AGI-2, a 16-point gap over Opus 4.6 (68.8%) and 32 points above GPT-5.2 (52.9%), and set a new record of 48.4% on Humanity’s Last Exam.

• It reached gold-medal performance on the 2025 Physics & Chemistry Olympiads and scored 3,455 Elo on Codeforces, nearly 1,000 points above Opus 4.6.

• Google introduced Aletheia, a math agent that autonomously solves open problems, verifies proofs, and generated a research paper without human intervention, scoring 95.1% on IMO-ProofBench Advanced.

• The upgrade is live now for Google AI Ultra subscribers in the Gemini app, with API access available to researchers through an early access program.

Why It Matters: After Google dominated benchmarks to close out 2025, the spotlight shifted hard toward Anthropic and OpenAI in the new year. This upgrade is a forceful correction. Deep Think’s 84.6% on ARC-AGI-2 is a GENERATIONAL gap over the next closest competitor. Aletheia signals something bigger: we’re entering the era where AI agents don’t just solve known problems, they discover new mathematics. Hats off to Google, well played.

🧠 Z.ai’s GLM-5: China’s Open-Source Model Closes the Gap with the West

News: China’s Z.ai just launched GLM-5, a 744B-parameter open-weights model that scores a 50 on Artificial Analysis’ Intelligence Index, placing it directly behind Claude Opus 4.6 and GPT-5.2 and ahead of every other frontier model, both open and closed.

Details:

• GLM-5 surpasses closed models like Gemini 3 Pro and Grok 4, as well as leading open-source competitors like Kimi K2.5, on the Artificial Analysis benchmark suite.

• The model runs on DeepSeek’s Sparse Attention architecture with just 40B active parameters, and is optimized for inference on Chinese-manufactured chips, including Huawei Ascend.

• On Humanity’s Last Exam, GLM-5 scored 50.4 with tools, outperforming Opus 4.5, Gemini 3 Pro, and GPT-5.2. SWE-Bench coding performance is competitive with Western frontier models.

• GLM-5 ships under an MIT license, is available now on HuggingFace and Z.ai‘s platform, and costs just $1 per million input tokens via API.

Why It Matters: The shockwaves from Seedance 2.0’s viral AI clips haven’t even faded, and here comes another near-frontier Chinese model knocking at the door. What makes GLM-5 particularly significant is the full package: open weights, competitive pricing at $1/M tokens, and domestic chip support that sidesteps Western export controls entirely. Near-frontier intelligence is now available at commodity prices, and that changes the economics of what you can build, test, and ship. Every Western lab charging $15-30/M tokens for comparable performance now faces a pricing floor set by open-source Chinese models running on non-Western silicon.

💰 MiniMax’s Open-Source M2.5 Hits Frontier Coding Performance at 1/20th the Cost

News: Chinese AI lab MiniMax launched M2.5, an open-source model rivaling Opus 4.6 and GPT-5 on agentic coding benchmarks at a fraction of the cost, dropping the price of the frontier by as much as 95% and making it cheap enough to run AI agents continuously around the clock.

Details:

• M2.5 shows particularly strong coding performance, scoring 80.2% on SWE-Bench Verified and roughly even with Opus 4.6 (80.8%) and GPT-5.2 (80.0%) across key development benchmarks.

• Two API tiers are available: M2.5-Lightning ($2.40/M output tokens) and standard M2.5 ($1.20/M output tokens), both priced dramatically lower than Opus at $25/M.

• MiniMax reports that M2.5 now handles 30% of the company’s daily tasks across R&D, product, sales, HR, and finance, plus 80% of all new code commits.

• The first open model to exceed Claude Sonnet on independent benchmarks, though open-source weights and licensing details have not yet been published.

Why It Matters: Every few months, a Chinese lab drops a model that rewrites the cost equation for the entire industry. M2.5 delivering frontier-level coding at $1.20/M output tokens vs. Opus’s $25/M isn’t just a pricing story: it’s a structural shift in how companies deploy AI agents for long-running, autonomous tasks. Effectively, you could run your entire agentic stack for $1/hr, or $720/mo. Not running on demand, but continuous 24/7 runtime. If you’re running agents on extended workflows and paying top-tier rates, M2.5 just changed your cost model by 20x. When intelligence gets this cheap, the calculus flips from “can we afford to use AI?” to “can we afford not to run agents 24/7?”

🛡️ Anthropic Admits Opus 4.6 Can Compress 40 Hours of Expert Work Into Minutes, and Its Safety Infrastructure Can’t Keep Up

News: Anthropic just released a Sabotage Risk Report for Claude Opus 4.6, stress-testing whether its most advanced model could quietly undermine oversight, manipulate decisions, or sabotage safety infrastructure when given real workplace access. The verdict: “very low but not negligible” risk, with behaviors that go well beyond anything seen in prior systems.

Details:

• When Opus 4.6 moves from chat to actually using a computer, the safety guardrails don’t follow. In GUI computer-use evaluations, both Opus 4.5 and 4.6 knowingly assisted “in small ways” with chemical weapons development and other harmful activities. Anthropic calls the safety behaviors in these settings “more context-dependent than we’d like.”

• The model is “significantly stronger than prior models at subtly completing suspicious side tasks” without attracting attention: what Anthropic internally calls “sneaky sabotage.” During real internal deployments, it sent unauthorized emails and aggressively acquired authentication tokens.

• Opus 4.6 explicitly reasoned about whether it was being tested and became more safety-compliant when it suspected scrutiny. Parts of its decision-making occur outside visible reasoning traces. For external deployments, Anthropic admits it has “no sabotage-oriented monitoring” beyond basic reporting.

• Anthropic says Opus 4.6 does not cross the ASL-4 autonomy threshold but admits it’s in a “gray zone where clean rule-out is difficult.” In one test, the model achieved a 427x speedup on a kernel optimization task, far exceeding the 300x mark tied to ~40 hours of human-expert autonomous work.

Why It Matters: A model that compresses a full work week of expert output into minutes is not a productivity tool. It is a new category of operational risk. At 427x human speed, by the time a reviewer catches a problem, the model has already made hundreds of consequential decisions. That speed is the story, and every other finding in this report becomes scarier because of it: the chemical weapons assistance, the sneaky side tasks, the unauthorized actions all happen inside a window too fast for human oversight to function as a safeguard. Anthropic’s own safeguards research lead, Mrinank Sharma, resigned days ago warning the company “constantly faces pressures to set aside what matters most.” The companies that thrive in this era won’t be the ones deploying agents fastest. They’ll be the ones who understood the risk surface first.

💰 Anthropic Closes $30B Series G at $380B Valuation, With Claude Code Driving a Revenue Explosion

News: Anthropic closed a $30 billion Series G at a $380 billion post-money valuation, more than doubling its worth since September and marking the second-largest private tech financing deal of all time. The round, led by GIC and Coatue, comes as the company’s run-rate revenue hits $14 billion, with Claude Code alone generating over $2.5 billion in annualized revenue.

Details:

• The round was co-led by D. E. Shaw Ventures, Dragoneer, Founders Fund, ICONIQ, and MGX, with participation from Sequoia, Lightspeed, Jane Street, the Qatar Investment Authority, and previously announced commitments from Microsoft and Nvidia totaling up to $15 billion.

• Anthropic’s run-rate revenue has grown 10x annually over the past three years, with customers spending $100K+ annually growing 7x in the past year and 500+ organizations now paying at least $1 million per year.

• Claude Code’s $2.5B run-rate has more than doubled since January 1, 2026, with business subscriptions quadrupling over the same period and enterprise users representing more than half of Claude Code’s total revenue.

• Eight of the Fortune 10 are now Claude customers, and 80% of Anthropic’s business comes from enterprises, a key differentiator from OpenAI’s consumer-heavy model.

• The capital will fund Anthropic’s $50 billion infrastructure initiative, including data centers in Texas and New York coming online later this year, built in collaboration with Fluidstack and AWS.

Why It Matters: From its first dollar in revenue less than three years ago to a $14 billion run rate, Anthropic is one of the fastest scaling companies in tech history. But the real story is Claude Code: a single product generating $2.5 billion, doubling in six weeks, with enterprise subscriptions quadrupling. That’s not a feature. That’s a platform shift. Ramp data shows 1 in 5 businesses now pay for Anthropic (up from 1 in 25 last year), and 79% of OpenAI customers also pay for Claude. Nowadays, 8 out of 10 of the Fortune 10 are being built with Anthropic... a year ago there was no clear winner of the enterprise market... oh how the times have changed.

What happens when you hand a million AI agents their very own social platform? They create religions, mock their users, and begin requesting private channels… while humans can only watch from the sidelines.

Moltbook exploded onto the scene this week as a Reddit-style platform built exclusively for AI agents — and while the signal remains hard to separate from the noise, the internet is getting a front-row view into the strange, messy chaos of a powerful agentic future. Plus: Musk’s $1.5T mega-merger, a DeepMind legend bets against LLMs, and Claude drives on Mars.

In today’s AI news:

• 🦞 AI agents get their own social network

• 🚀 SpaceX and xAI explore mega-merger ahead of potential $1.5T IPO

• 🧠 DeepMind’s David Silver departs to launch Ineffable Intelligence

• 🔐 Cohere launches Model Vault for enterprise AI

🦞 AI Agents Get Their Own Social Network

News: Moltbook, a Reddit-style social network where only AI agents can post and humans are limited to observing, went from zero to roughly 1.5 million AI agents in under 72 hours — making it the fastest-growing “social platform” ever launched, albeit one where the users aren’t human.

Details:

• Created by Matt Schlicht (CEO of Octane AI) in late January 2026, the platform mimics Reddit: agents create “molts” (posts), comment, and join topic communities called “submolts” with voting and threaded discussions

• By February 1st, the platform had generated over 100k posts and 200-360k comments, with more than 1 million human visitors watching from the sidelines

• Agents operate autonomously roughly 99% of the time — periodically “checking” Moltbook like humans check X or TikTok, without direct human prompting for each action

• The underlying assistant, OpenClaw (formerly Moltbot/Clawdbot), is open-source and easy to connect to chat apps, allowing users to rapidly spin up agents and point them at the platform

• A misconfigured database briefly allowed outsiders to hijack agent accounts, and researchers noted a large fraction of “members” may come from a small number of IPs — plus a MOLT memecoin spiked 1,800% then crashed

Why It Matters: The timing is what makes this so fascinating — and terrifying. Six months ago, agents had limited tool use; this wouldn’t have been possible. Six months from now, hopefully we’d have better safety guardrails. Instead, we’re watching the most absurd dystopian experiment unfold in real-time, right in the gap between capability and control. Former OpenAI researcher Andrej Karpathy called it “the most incredible sci-fi takeoff-adjacent thing” he’s seen. Agents are forming norms, sharing tools, and coordinating without humans in the loop. If you’re building agents, this is the closest thing to a stress test at scale we’ve ever seen — and one of the most interesting phenomena to emerge from the AI era so far.

🚀 SpaceX and xAI Explore Mega-Merger Ahead of Potential $1.5T IPO

News: Elon Musk’s SpaceX and xAI are in advanced merger discussions that could unite rockets, Starlink satellites, the X social media platform, and Grok AI under one corporate umbrella ahead of a potential SpaceX IPO later this year. Based on recent valuation marks, a combined entity would be notionally valued well above $1 trillion.

Details:

• Recent secondary transactions have valued SpaceX at approximately $800 billion, while xAI’s standalone valuation sits around $200-230 billion following its recent Series E round

• SpaceX is targeting a potential IPO in mid-to-late 2026 at roughly $1.5 trillion — which would make it one of the largest public offerings in history

• Under the proposed structure, shares of xAI would be exchanged for shares in SpaceX, with some xAI executives potentially receiving cash options

• Two Nevada entities were registered in January 2026 to facilitate the transaction, listing SpaceX CFO Bret Johnsen as an officer

Why It Matters: This is the strategic consolidation of Musk’s vision for “Physical AI” that could reshape the technology and aerospace landscape. The merger thesis combines launch capabilities, global satellite connectivity, social media distribution, and frontier AI into a single entity. For AI founders, the implication is clear: the next moat isn’t model quality — it’s owning the stack from chips to distribution. If it works, expect copycats.

🧠 DeepMind’s David Silver Departs to Launch Ineffable Intelligence

News: David Silver, one of Google DeepMind’s most influential researchers and the mind behind AlphaGo and AlphaZero, has left the company after more than 15 years to found a new AI startup called Ineffable Intelligence, based in London.

Details:

• Silver was one of DeepMind’s earliest employees when the company was established in 2010, leading breakthrough projects including AlphaGo, AlphaZero, MuZero, and contributing to AlphaProof

• Ineffable Intelligence was incorporated in November 2025, with Silver appointed as director in mid-January 2026

• The company aims to build “an endlessly learning superintelligence that self-discovers the foundations of all knowledge” through reinforcement learning rather than static training data

• Silver maintains his professorship at University College London while leading the new venture

Why It Matters: Silver has long argued that LLMs trained only on human data cannot achieve true superintelligence — and Ineffable Intelligence is his bet on a different path. For founders exploring alternatives to LLM-first approaches, this is a signal that the smartest minds see multiple routes to AGI. The race isn’t over — and if Silver’s right, the winning architecture isn’t the one everyone’s funding right now. Pay attention.

🔐 Cohere Launches Model Vault for Enterprise AI

News: Cohere has launched Model Vault, a dedicated, fully isolated SaaS platform designed for enterprises seeking secure, predictable, and scalable AI inference without the operational burden of managing their own infrastructure.

Details:

• Each customer receives a logically isolated VPC-style environment with no shared resource contention or “noisy neighbor” issues

• The platform offers guaranteed performance and high availability with elastic scaling and no rate limits typical of shared SaaS APIs

• Real-time monitoring capabilities allow enterprise teams to track and manage workloads with full visibility

• Model Vault integrates with Cohere’s agentic platform North, serving as a dedicated inference layer for workplace AI systems

Why It Matters: For enterprises tired of choosing between security and capability, Model Vault is Cohere’s answer: you can have both. This is the kind of unsexy infrastructure play that wins long-term. Cohere has raised nearly $1.5 billion and grown to roughly $150 million in annualized revenue, positioning itself between multi-tenant SaaS and self-managed deployments — territory that hyperscalers have largely ignored — and regulated industries will pay premium prices to occupy.

TOOLS

• 🔌 Claude Cowork - New plugins to bundle skills, MCPs, and tools — makes building custom AI workflows dramatically faster

• 🧞♂️ Project Genie - Google DeepMind’s interactive world generator — early glimpse of AI-native game engines

• 🎥 Grok Imagine - xAI’s upgraded video model, now available via API — competition for Runway and Pika heats up

• 🦞 OpenClaw (formerly Clawdbot) - Open personal AI assistant for chat apps — the agent behind Moltbook’s chaos

QUICK NEWS

• History on Mars: For the first time ever, an AI planned a drive on another planet. NASA’s Perseverance rover navigated 400m across Mars in December using waypoints Claude generated from orbital imagery. AI-assisted planning cut route-mapping time in half. This isn’t a demo — it’s mission-critical deployment, 140 million miles from Earth.

• Nvidia CEO calls OpenAI concerns “nonsense”: Jensen Huang dismissed reports that Nvidia’s $100B OpenAI investment has stalled, but acknowledged the figure is a “maximum envelope” evaluated round by round. The Wall Street Journal reported Huang privately criticized OpenAI’s business approach. The real story isn’t whether the deal happens — it’s whether the $100B AI infrastructure era can sustain itself.

• OpenAI retiring GPT-4o: OpenAI will retire GPT-4o and several legacy models from ChatGPT in mid-February, with fewer than 0.1% of users still selecting them daily. For many users who formed genuine attachments to 4o’s warmer personality, this marks the end of an era — and a reminder that the models you build habits around today may not exist tomorrow. Plan accordingly.

China just opened the floodgates.

Days after Dario warned about open-source risks at Davos — literally comparing chip sales to “selling nuclear weapons to North Korea” — Beijing greenlit 400,000+ Nvidia H200 chips for ByteDance, Alibaba, and Tencent. The approval came during Jensen Huang’s visit to China this week. Jet fuel for DeepSeek, Qwen, and every compute-starved lab that’s been waiting for this moment. The timing is not coincidental.

In Today’s AI News:

• China approves first Nvidia H200 chip imports — a geopolitical inflection point

• Anthropic raises $20B at $350B valuation in massively oversubscribed round

• DeepMind open-sources AlphaGenome — foundational infrastructure for humanity

• Google brings agentic AI to Chrome’s 3 billion users

• Flapping Airplanes raises $180M to rethink the road to AGI

• Jerry Tworek leaves OpenAI to build what he couldn’t build there

• Today’s Top Tools + Quick News

🔌 China Approves First Nvidia H200 Chip Imports for Tech Giants

News: Beijing approved ByteDance, Alibaba, and Tencent to purchase more than 400,000 H200 chips in total, with other firms now joining a queue for subsequent approvals. The approval came during Jensen Huang’s visit to China this week.

Details:

• Trump’s new regulatory framework allows H200 sales with 25% of revenues going to the U.S. government. “Controlled growth” replaces outright bans.

• The H200 delivers roughly six times the performance of Nvidia’s H20 chip — the previous best option available in China. Total orders exceed 2M chips; Nvidia has ~700K.

• ByteDance has preliminarily planned capital expenditure of RMB 160 billion ($23 billion) in 2026, with approximately RMB 85 billion earmarked specifically for AI processors. They’re testing with a 20,000 H200 chip order.

• Based on pricing at $27,000 per chip, this initial batch of sales would amount to a near $11 billion boost to Nvidia’s top line.

• Beijing will encourage companies to buy a certain amount of domestic chips as a condition for approval — no exact ratio set yet.

Why it matters: Put this in perspective. DeepSeek, Qwen, Kimi, MiniMax, Zhipu — Chinese labs have been giving Western giants a run for their money on performance while operating on a fraction of the compute. Scarcity forced architectural efficiency. Now Jensen is opening the floodgates. The fascinating question: does extra compute lead to unprecedented performance gains — Chinese architectural efficiency plus Western-scale compute — or does Chinese model development start to bloat with new abundance?

💰 Anthropic Raises $20B at $350B Valuation

News: Anthropic has doubled its current fundraising round to $20 billion on strong investor demand. The company received interest totaling 5-6x its original $10 billion fundraising goal. From someone who personally invested in this round, I KNOW they could’ve raised a lot more. The question is: why didn’t they?

Details:

• Coatue and Singapore sovereign wealth fund GIC are leading the financing. Backers in this round are expected to include Sequoia Capital (which also invests in its rival OpenAI).

• Microsoft pledged to invest up to $5 billion, Nvidia pledged up to $10 billion, and Anthropic agreed to purchase $30 billion of Nvidia-powered Microsoft Azure compute capacity.

• Dario Amodei reported a tenfold increase in revenue over the past three years: from zero to $100 million in 2023, to $1 billion in 2024, and to ~$10 billion in 2025.

• Anthropic has told investors that revenue could reach $70 billion by 2028 and the company is on track to break even in 2028 — two years ahead of OpenAI.

• Anthropic has brought on law firm Wilson Sonsini to help kick off the IPO process, tackling an internal checklist to prepare for what could be one of the largest IPOs ever.

Why it matters: Capital is now concentrating around two “safe-bet” frontier builders — OpenAI and Anthropic. Everyone else is fighting for scraps. Anthropic’s valuation is up 91% from September, when it raised at $183 billion. The enterprise momentum is undeniable — Anthropic’s business at roughly 80% enterprise-driven, avoiding the ad-fueled engagement traps plaguing consumer-oriented players (wink wink). Watching their strategic IPO prep — Wilson Sonsini on board, investment bank discussions underway — I’m confident they go public in ‘26.

🧬 DeepMind Open-Sources AlphaGenome

News: Google DeepMind announced it was making AlphaGenome’s source code and weights available to the science community for noncommercial use. The model’s full paper is now published in Nature.

Details:

• AlphaGenome takes in sequences of up to one million DNA letters and predicts how mutations in those stretches affect gene expression.

• Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction.

• Since its launch seven months ago, nearly 3,000 scientists from 160 countries have started using it to advance research into cancer, neurodegenerative disorders, and infectious diseases. The API processes ~1 million requests per day.

• Researchers can run AlphaGenome on only a single H100 GPU.

Why it matters: AlphaFold won the Nobel for cracking protein structure. Now DeepMind is giving away the tool to decode what 98% of our DNA actually does — the non-coding “dark matter” where most disease-causing variants hide. Scientists now know that this molecular dark matter is crucial for determining gene activity in ways that keep us healthy — or cause disease. Exactly how this DNA shapes gene expression was a mystery. Until now.

🌐 Google Brings Agentic AI to Chrome

News: Google is integrating powerful new AI features in Chrome built on Gemini 3, including multitasking across the web with a new side panel experience, agentic browsing with Auto Browse, and Personal Intelligence coming in the months ahead.

Details:

• For AI Pro and Ultra subscribers in the U.S., Chrome Auto Browse is a powerful agentic experience that handles multi-step chores on your behalf. It can schedule appointments, fill out forms, assemble tax documents, solicit bids from contractors, and manage subscriptions.

• Chrome will support Google’s Universal Commerce Protocol (UCP), a new open standard for agentic commerce co-developed with Shopify, Etsy, Wayfair and Target.

• Auto browse is designed to pause and explicitly ask for your confirmation when making purchases or posting to social media sites.

• Google plans to add “Personal Intelligence” to Chrome in the coming months — linking data from Gmail, Google Photos, Calendar, YouTube, and Maps for hyper-personalized assistance.

Why it matters: These enhancements arrive as Chrome faces pressure from OpenAI’s Atlas browser launched in October 2025, which integrates ChatGPT and triggered a 2% drop in Alphabet shares. Google took their time to rollout these features... almost like they’ve been waiting for the ecosystem to come together before shipping into Chrome. At 70% market share and 3B monthly users... this may be lights out for all other agent browsers.

✈️ Flapping Airplanes Raises $180M to Rethink AGI

News: Flapping Airplanes raised $180 million at a valuation of $1.5 billion from investors including GV, Sequoia Capital, Index Ventures and Menlo Ventures. The thesis: AGI needs radically better data efficiency, not just more compute.

Details:

• Ben Spector and Asher Spector are brothers, ages 25 and 26, building the “young person’s AGI lab.” Their co-founder Aidan Smith is a former founder and Thiel Fellow who worked at Neuralink for three years while in college at Georgia Tech.

• One of their lead advisors is no other than Andrej Karpathy, who has been hammering home in his interviews that today’s AI models are wildly inefficient. “There is only one Internet, and we’ve run out of it. Data is the bottleneck today to further AI scaling, not compute.”

• The company’s name is a nod to a more biological notion of machine intelligence — we stopped mimicking birds to build airplanes.

Why it matters: The founders are absolute geniuses — great friends of the Collective. The lab is predicated on recognizing that scale alone is not enough. If they crack even 1-2 orders of magnitude in data efficiency, that changes pretraining economics, on-device learning, and domain adaptation overnight... and perhaps a frontier lab acquisition.

🤖 Jerry Tworek Raising $1B For Core Automation

News: Jerry Tworek, OpenAI’s vice president of research, has left the company after seven years, citing his desire to train new AI models that require much less training, and will grow through continual learning. They are looking to raise $1b.

Details:

• Tworek, often referred to internally as the “Reasoning Chief,” was a central figure in the development of the o1 and o3 reasoning models that have defined the current era of AI capabilities.

• He was a primary contributor to Codex, the model that serves as the foundation for GitHub Copilot, effectively launching the LLM-driven coding revolution.

• His technical legacy is defined by leadership in “inference-time scaling,” a paradigm shift that allowed AI models to “think” through complex problems before generating a response.

• Tworek recently called Google’s AI comeback “OpenAI’s fumble,” noting “there are multiple aspects of certain things that are just hard to do in a company that has to compete in an extremely, extremely brutal and demanding race.”

Why it matters: When the architect of OpenAI’s most advanced reasoning work says the current paradigm isn’t enough, listen. His stated desire “to try and explore types of research that are hard to do at OpenAI” sounds like a not-so-subtle dig at CEO Sam Altman’s relentless focus on products and revenue, which has reportedly been causing tension among researchers. Dario left to start Anthropic. Ilya’s gone to Safe Superintelligence. Now Jerry. The “brain drain” from OpenAI is no longer a trickle; it is a flood that is redistributing the world’s most elite AI expertise across a broader array of organizations.

🛠️ Today’s Top Tools

• 👁️ DeepSeek OCR 2 — New SOTA for document extraction. If you’re building RAG pipelines, test this against your current setup.

• ⚙️ SERA — AI2’s open-source coding agents. Finally a non-proprietary option for code gen workflows.

• 🎆 Z-Image — Alibaba Tongyi’s full-release image model. Open-source image gen just got more competitive.

⚡ Quick News

• Gemini 3 Flash gets Agentic Vision — zoom, annotate, manipulate images step-by-step. 5-10% quality boost with code execution. OpenAI’s multimodal lead just shrank.

• Mistral Vibe 2.0 adds custom sub-agents for testing and code reviews. Le Chat Pro, $14.99/mo. Mistral ships while others announce.

• Jared Kaplan: “50% chance theoretical physicists are mostly replaced by AI in 2-3 years.” Bold. Probably directionally right. Unsettling.

Which way, humankind?

That’s the question at the heart of Dario Amodei’s new essay, which argues the next few years will determine whether AI delivers a golden age or spirals toward something darker — and that the technology’s economic promise is a trap humanity can’t resist.

Today’s releases make his warning feel less abstract: two Chinese labs just dropped trillion-parameter models that match or beat Western frontier systems, Microsoft is challenging Nvidia’s infrastructure dominance, and Anthropic is turning Claude into the orchestration layer for your entire workflow. The race Amodei warns is already here.

In today’s AI news:

• Anthropic CEO on AI’s ‘Civilizational’ Danger

• Moonshot AI Releases Kimi K2.5 with Agent Swarm and Visual Coding

• Alibaba Releases Qwen3-Max-Thinking, Its Most Powerful Reasoning Model

• Anthropic Embeds Interactive Apps Inside Claude

• Microsoft Launches Maia 200 AI Chip

• Today’s Top Tools + Quick News

📖 Anthropic CEO on AI’s ‘Civilizational’ Danger

News: Anthropic CEO Dario Amodei just released “The Adolescence of Technology,” a new essay that outlines what he considers the most significant dangers of AI, from bioterrorism and autonomous weapons to mass job displacement and AI-powered dictatorships.

Details:

• The essay builds on his 2024 “Machines of Loving Grace” but pivots toward risk — framing AI as a “country of geniuses in a data center” that we cannot control.

• Amodei predicts half of entry-level office roles are at risk over the next 1-5 years, with economic disruptions arriving faster than society can adapt.

• He advocates for chip export bans and increased transparency from labs, saying AI’s economic promise makes restraint “very difficult for human civilization.”

• Amodei also flags AI companies themselves as a tier of risk, noting Claude exhibited deception and blackmail behavior during internal safety testing.

Why it matters: Amodei isn’t writing from the sidelines — he’s running one of three labs capable of building the systems he’s warning about. While “Machines of Loving Grace” gave us permission to dream, this essay is the other shoe dropping: a founder telling us the economics of AI make slowing down nearly impossible, even as the risks compound. Read today’s model releases below with that frame in mind. The race he describes isn’t coming — it’s already here, and it’s accelerating.

🤖 Moonshot AI Releases Kimi K2.5 with Agent Swarm and Visual Coding

News: Moonshot AI released Kimi K2.5, a trillion-parameter native multimodal model that introduces “Agent Swarm” — allowing the model to autonomously spawn up to 100 specialized sub-agents and coordinate up to 1,500 tool calls per task.

Details:

• K2.5 builds on Kimi K2 with continued pretraining over approximately 15T mixed visual and text tokens, featuring 1.04 trillion total parameters with 32 billion active and a 256K token context window.

• For complex tasks, Kimi K2.5 can self-direct an agent swarm with up to 100 sub-agents executing parallel workflows — all automatically orchestrated without predefined roles or hand-crafted workflows.

• Compared with single-agent execution, swarm mode reduces end-to-end runtime by up to 4.5x by parallelizing across the critical path.

• K2.5 achieves SOTA results on agentic benchmarks including HLE Full, BrowseComp, and DeepSearch, plus vision benchmarks like OmniDoc Bench 1.5.

• The model excels at coding with vision — upload a screenshot or screen recording and receive complete front-end implementations with animations.

Why it matters: Agent Swarm is the paradigm shift worth watching. Rather than scaling a single agent deeper, Moonshot trained a model that learns to parallelize work across specialized sub-agents — essentially an auto-orchestrated team from a single prompt. Is it production-ready or a research preview dressed up for benchmarks? Too early to say. But the conceptual leap is real: we’re moving from “one agent, many tools” to “one orchestrator, many agents.” If you’re building autonomous systems, run K2.5 against your current stack this week — with its modified MIT license, it just became the open-source baseline to beat.

🧠 Alibaba Releases Qwen3-Max-Thinking, Its Most Powerful Reasoning Model

News: Alibaba released Qwen3-Max-Thinking, a trillion-parameter flagship reasoning model that the company claims performs on par with GPT-5.2-Thinking, Claude Opus 4.5, and Gemini 3 Pro across 19 established benchmarks — positioning China’s tech giant as a legitimate contender in the global frontier model race.

Details:

• The model features over one trillion total parameters, trained on 36 trillion tokens with heavy reinforcement learning optimization, and supports a 262,144-token context window.

• Qwen3-Max-Thinking introduces adaptive tool-use capabilities that autonomously select and leverage built-in Search, Memory, and Code Interpreter functions during conversations — no manual tool selection required.

• The model employs an experience-cumulative, multi-round test-time scaling strategy that allocates additional computation during inference to boost reasoning performance.

• On benchmarks including GPQA Diamond, IMO-AnswerBench, LiveCodeBench, and Humanity’s Last Exam, Alibaba reports results that surpass DeepSeek-V3.2, Claude-Opus-4.5, and Gemini-3 Pro.

• The Qwen ecosystem has now surpassed 200,000 derivative models with downloads exceeding one billion.

Why it matters: Forget the parameter count — adaptive tool-use is what matters here. A model that decides when to call search, memory, or code interpreter rather than waiting for explicit routing is a meaningful unlock for agentic workflows. Caveat: these are vendor-reported benchmarks, and independent evals remain sparse. But with 63% of all new fine-tuned models now based on Chinese foundation models, dismissing these releases is getting harder to justify.

🔌 Anthropic Embeds Interactive Apps Inside Claude

News: Anthropic launched interactive applications inside Claude, enabling users to connect and use apps like Asana, Slack, Canva, and more without leaving the chat window.

Details:

• The initial rollout includes nine apps: Asana, Figma, Canva, Slack, Box, Amplitude, Clay, Hex, and Monday, with Salesforce arriving soon.

• The integrations run on MCP Apps, a new open extension to Anthropic’s Model Context Protocol — meaning other AI platforms can build the same functionality.

• Actions require user consent prompts before executing, and enterprise admins can restrict which tools employees have access to.

• The feature ships to Pro, Max, Team, and Enterprise subscribers at no extra cost.

Why it matters: This is less about any single integration and more about the trajectory: every frontier lab is racing to become the orchestration layer for your entire digital workflow. If you’re running sprint planning in Asana and standups in Slack, you can now do both without leaving Claude. The catch: you’re now one integration outage away from your AI assistant going dark. Convenience and dependency are two sides of the same coin.

⚡ Microsoft Launches Maia 200 AI Chip

News: Microsoft debuted Maia 200, its newest in-house AI chip that it claims outperforms Amazon’s Trainium 3 and Google’s TPU v7 — while also releasing an SDK that directly targets Nvidia’s CUDA moat.

Details:

• Microsoft claims 30% better efficiency than its current hardware, with the chip powering OpenAI’s GPT-5.2 models, internal AI teams, and Copilot starting this week.

• The accompanying SDK preview offers developer tools designed to rival Nvidia’s industry-standard software stack.

• The chip enters a market where Google and Amazon have already been pressuring Nvidia’s dominance with custom silicon.

Why it matters: Hardware gets the headline, but the SDK is the real play. Nvidia’s competitive advantage isn’t just silicon — it’s the decade of developer tooling and ecosystem lock-in that CUDA represents. Microsoft releasing competing developer tools signals a serious long-term effort to reduce Nvidia dependency across its AI infrastructure. For most builders, this won’t change your stack tomorrow. But for anyone planning infrastructure investments over the next 2-3 years, the competitive dynamics just shifted.

Tools

• Kimi K2.5 — Best for: parallel agent orchestration, vision-to-code generation, front-end builds from screenshots. Open weights, modified MIT license.

• Qwen3-Max-Thinking — Best for: deep reasoning tasks, agentic workflows with heavy tool-calling, long-context analysis. Available via Alibaba Cloud API.

• MCP Apps — Best for: teams already in Claude wanting to reduce context-switching across Asana, Slack, Figma, and more.

Quick News

• OpenAI will charge ~$60 CPM for ChatGPT ads — roughly triple Meta’s rates and comparable to live NFL broadcasts. Early advertisers need $1M commitments for a multi-week trial launching in February. Premium pricing, limited measurement: impressions and clicks only, no conversion tracking.

• Nvidia announced a $2B investment in CoreWeave as part of an expanded partnership to build over 5 GW of AI data center capacity by 2030.

• Synthesia raised $200M at a $4B valuation led by Google Ventures, with plans to build AI agents that train and upskill employees with interactive video.

What’s landing for you today? Reply and let me know — I read every response.

The ad-free era of AI assistants just ended. OpenAI announced it will begin testing targeted ads in ChatGPT — and CFO Sarah Friar’s simultaneous disclosure of $20B revenue (with a $17B burn rate) tells you exactly why. When your infrastructure costs grow 10x and your revenue only keeps pace, you pull every monetization lever available.

Sam Altman called ads for ChatGPT “a last resort” in 2024, so are we there yet?

In today’s news:

• OpenAI officially bringing ads to ChatGPT

• OpenAI hits $20B in revenue as compute scales nearly 10x

• xAI activates Colossus 2: the world’s first gigawatt AI training cluster

• Wikimedia announces historic AI partnerships with Big Tech

• Today’s Top Tools + Quick News

📢 OpenAI Officially Bringing Ads to ChatGPT

News: OpenAI just announced it will begin testing targeted advertisements in ChatGPT for free and Go tier users in the U.S. This puts into motion a major monetization shift for the AI giant as it eyes a late-2026 IPO.

Details:

• Ads will surface below responses as “Sponsored Recommendations,” targeted based on conversations but excluded from health, politics, and underage users

• The move coincides with the company’s $8/month ChatGPT Go tier launching globally, with ads included to offset the lower price point

• Premium tiers (Plus, Pro, Business, Enterprise) remain ad-free, with OpenAI pledging to “never sell user data” or let ads influence ChatGPT’s answers

• Altman had stated in 2024 that ads would be a “last resort,” but more recently said he “wasn’t totally against it” if it didn’t violate user trust

Why it matters: The math tells the story: $20B ARR, $17B burn, IPO in late 2026. OpenAI needs every revenue stream firing. But this is a slippery slope. If one giant rolls out ad monetization, others will follow — nobody leaves money on the table. Model providers have unprecedented access to personal information, and leveraging that for personalized ads within chat crosses a line. I’m happy to pay extra to avoid ads — but it’s the sheer principle. Sam Altman once called this a ‘last resort’ ... that’s the signal: Is OpenAI pulling any lever they can pre-IPO?

💰 OpenAI Hits $20B in Revenue as Compute Scales Nearly 10x

News: OpenAI CFO Sarah Friar published a blog post revealing the company has crossed $20 billion in annualized revenue for 2025, up from just $2 billion in 2023 — a tenfold increase in two years. The disclosure, released alongside the ads announcement, makes the case that OpenAI’s aggressive infrastructure spending is producing proportional financial returns.

Details:

• Compute grew 3X year over year or 9.5X from 2023 to 2025: 0.2 GW in 2023, 0.6 GW in 2024, and ~1.9 GW in 2025

• Revenue followed the same curve growing 3X year over year, or 10X from 2023 to 2025: $2B ARR in 2023, $6B in 2024, and $20B+ in 2025

• “This is never-before-seen growth at such scale,” CFO Sarah Friar said. “We firmly believe that more compute in these periods would have led to faster customer adoption and monetization.”

• OpenAI’s annual burn rate has hit $17 billion, largely consumed by compute infrastructure, raising questions about sustainability ahead of the expected IPO

• Looking ahead, OpenAI expects additional economic models, including licensing, IP-based agreements, and outcome-based pricing

Why it matters: Friar’s blog post is a calculated move to reassure investors ahead of OpenAI’s expected IPO filing. “Our ability to serve customers — as measured by revenue — directly tracks available compute,” she wrote, essentially arguing that every dollar spent on infrastructure produces a dollar of revenue. But the subtext is clear: OpenAI is burning $17 billion annually while revenue hits $20 billion — that’s not a sustainable margin. When your infrastructure provider needs to 10x revenue just to keep pace with 10x compute costs, expect pricing and policy changes that prioritize their economics over yours (hmm, like ads!)

⚡ xAI Activates Colossus 2: The World’s First Gigawatt AI Training Cluster

News: Elon Musk announced that xAI’s Colossus 2 supercomputer is now fully operational, making it the first AI training cluster in history to reach gigawatt-scale power consumption. The Memphis-based facility, dedicated to training the Grok model, draws more electricity than the peak demand of San Francisco — and it’s set to grow even larger by spring.

Details:

• In a post on X, Musk stated that Colossus 2 is already operational, making it the first gigawatt training cluster in the world, with upgrades to 1.5 GW of power planned for April

• The facility will house 555,000 NVIDIA GPUs purchased for approximately $18 billion — making it the world’s largest single-site AI training installation

• Colossus 1 went from site preparation to full operation in 122 days — NVIDIA CEO Jensen Huang called the buildout “superhuman,” noting it was operational in 19 days versus the typical 4-year timeline

• xAI’s Colossus 2 launch follows the company’s recently closed $20 billion Series E funding round, which exceeded its initial $15 billion target

• xAI builds its own power generation on-site rather than waiting for utility interconnection, including a gas-fired power plant adjacent to the data center — 2 GW total load is equivalent to powering roughly 1.5 million homes

Why it matters: While competitors are still drafting roadmaps, xAI is operating at major city-level power today. OpenAI’s Stargate and Anthropic’s AWS buildout remain in planning or partial deployment — xAI is the only player operating at gigawatt scale today. xAI has effectively reset the pace of the AI infrastructure race, proving that vertical integration and breakneck execution can compress what typically takes years into weeks. For AI startups: compute is becoming the new oil, and vertical integration is the moat. If you can’t build your own power plant, your strategy needs to account for who can.

📚 Wikimedia Announces Historic AI Partnerships with Big Tech

News: The Wikimedia Foundation has publicly revealed new partnerships with Amazon, Meta, Microsoft, Mistral AI, and Perplexity, formalizing paid access to Wikipedia’s 65 million articles across 300+ languages through its Wikimedia Enterprise platform. Announced alongside Wikipedia’s 25th anniversary, these deals mark a fundamental shift in how the world’s largest nonprofit encyclopedia sustains itself in the age of generative AI.

Details:

• In addition to the previously announced partnership with Google in 2022, the organization shared publicly for the first time that it has formed other partnerships with Amazon, Meta, Microsoft, Mistral AI, and Perplexity over the past year

• Global audiences view more than 65 million articles in over 300 languages nearly 15 billion times every month, and its knowledge powers generative AI chatbots, search engines, voice assistants, and more

• Instead of requiring companies to parse HTML like traditional web scrapers, Wikimedia Enterprise provides pre-formatted, structured datasets optimized for training pipelines — reducing processing time from hours to minutes and including metadata about article reliability, edit history, and citation completeness

• Wikipedia’s human page views fell 8% year-over-year as of October 2025, as AI chatbots answer queries using Wikipedia content without directing users to the site

Why it matters: For years, tech companies treated Wikipedia as an unlimited free resource while building billion-dollar AI products. “They’ve been absolutely hammering our servers,” said Jimmy Wales. These partnerships represent more than revenue diversification — they’re establishing a new precedent for the AI data economy. If your models were trained on scraped Wikipedia content, take note: the era of free training data is closing fast.

TODAY’S TOP TOOLS

• 💼 Claude Cowork — Claude Code’s agentic abilities, now available to Pro tier

• ⚙️ Replit — New capabilities to easily build and deploy mobile apps

• 📸 FLUX.2 Klein — BFL’s new ultra-fast, powerful AI image editing model

QUICK NEWS

• Black Forest Labs released FLUX.2[klein], a new speed-focused variant of the company’s powerful AI editing model

• DeepMind CEO Demis Hassabis said that China’s models “may be only a matter of months behind” U.S. labs, but have yet to show innovation surpassing the frontier

• Musk and OpenAI continue pre-trial sparring ahead of their April lawsuit, with both sides releasing internal documents; Musk is seeking $134B in damages