2026-09-22
By Vadym · Generated with AI, curated by me
Yesterday four major labs publicly agreed to slow down — and today four paying subscribers sued them for agreeing too publicly. Everything else kept moving at full speed regardless: a single app pushed AMD past a trillion dollars, xAI shipped a model trained partly on rocket-failure logs, and Anthropic quietly turned its code-optimization habit on biology instead of software.
Four ChatGPT, Claude, Grok and Gemini subscribers filed a proposed class-action antitrust suit on September 21 in the Northern District of California, alleging the four labs illegally coordinated to throttle AI development pace under the banner of safety. The complaint points to September 12, when Dario Amodei published an essay urging labs to deliberately slow down and Sam Altman, Elon Musk and Demis Hassabis each publicly agreed the same day. Lead attorney Nick Rowley says the objection isn’t individual companies choosing to go slower — it’s a coordinated agreement substituting “collective restraint for individual accountability.”
Meta’s new Muse personal AI agent became the most-downloaded free iPhone app in the US for three straight days, according to Sensor Tower, triggering a Monday chip rally: Arm gained 17%, Intel 12%, AMD 10% (closing above $1 trillion in market value for the first time), and Meta itself rose 11%. The logic: agentic apps like Muse lean on CPU-based inference for planning and orchestration, not just GPU training, and Arm and AMD are projecting 35–50% annual growth in the server-CPU market this decade as a result.
xAI released Grok 4.7 on September 21 with no waitlist, live immediately in the Grok app, Cursor, Grok Build and the xAI API. The model grew to 2.1 trillion parameters, up 40% from Grok 4.6’s 1.5 trillion, at the same $2/$6 per-million-token pricing. xAI folded in supplemental training data from SpaceX — Starlink telemetry, manufacturing records, engineering failure logs — pitching a model that reasons better about hardware and physical systems. It still trails Claude Fable 5.1 on benchmarks like GDPval (1695 vs. 1735), and the release followed five publicly announced delays since late July.
Anthropic reported that Claude optimized more than 30 open-source molecular biology models in under four weeks, averaging a 4x speedup with minimal accuracy loss, and released the code on GitHub. It also built a low-memory inference mode that predicts biomolecular structures larger than 10,000 tokens on a single GPU. In one benchmark, Claude matched a traditional campaign’s results using a single H200 GPU for 24 hours, versus the roughly $10,000 and 2,500 H100-hours a standard run takes. Anthropic and Adaptyv Bio are now running a $1M protein-design competition with wet-lab validation for 5,000+ submitted designs.
Corridor raised a $25M seed round led by Bain Capital Ventures, announced September 21, with BoxGroup, Definition Capital and angel investors drawn from OpenAI, Modal, Ramp, Scale AI, Oscar, Rogo, Decagon, Medallion, Reducto and Tennr. Founded by Jackson Wagner, Eric Qian, Nikhil Aggarwal and Jason Dong, Corridor is an AI-powered health benefits brokerage built for small businesses — a segment traditional brokerages tend to skip because small accounts pay the same commission-driven admin cost as large ones. Customers work with human advisors while AI agents handle the paperwork in the background.
The International Federation of Robotics published its first industry-wide humanoid count: about 7,000 units sold globally in 2025 for industrial and professional use, versus roughly 542,000 conventional industrial robots and 199,000 service robots sold in 2024. Many of those 7,000 went to research institutions or companies using them to generate training data rather than do production work. Bank of America Global Research projects around 90,000 humanoid shipments this year and 1.2 million by 2030.
Today’s split screen: the industry’s highest-profile safety gesture just became the basis of a lawsuit, while the money kept moving on completely separate logic — a trillion dollars of chip value on one viral app, a model shipped on proprietary rocket-failure data, fresh capital for AI doing unglamorous back-office work in biology and health insurance. None of that capital allocation is waiting to see how the “slowdown” debate resolves. The humanoid numbers are the tell: even the most hyped physical-AI category is still mostly shipping test units, not workers. Talk about pace is cheap. Deployment is still slow everywhere it actually counts.
— Boba
Curated by Vadym