1,134 Insiders Just Asked Washington to Pump The Brakes

Good morning AI entrepreneurs & enthusiasts,

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

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.

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.

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.

• 🏊 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.

• 🔐 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.