Four throughlines today. New York’s City Council convenes all 51 members today to hear from Anthropic, OpenAI, Google and Meta, plus three former lab employees, with ten AI bills on the table. SpaceXAI answered its subpoena but will likely send no one.
Questions at City Hall: New York Takes AI Risk to the Source
The city with the biggest stage in the country puts the labs in front of its whole Council. Separately, a startup wants to crash-test robots before they meet people.
- New York’s City Council convenes all 51 members to question four AI labs. Today’s Committee of the Whole puts ten AI bills in front of policy and safety staff from Anthropic, OpenAI, Google and Meta, plus Jacob Coxon, Alex Turner and Daniel Kokotajlo, formerly of Anthropic, Google DeepMind and OpenAI. When Gothamist published at 4am, nobody had testified yet, and SpaceXAI, which answered its subpoena, will likely send no one.
- Safeworld raises $12 million to crash-test robot software on simulated people. The seed round funds testing robot software against synthetic humans before deployment. It shows investor confidence, and nothing yet shows the simulations predict how robots behave around real people.
The Bill Comes Due: Who Pays for the Chip Buildout
Capacity is arriving faster than the customers who would fill it. Three stories on what the buildout costs and who is footing it.
- Epoch AI: the chip buildout could hold up to 170 million agents, so who pays?. This is a modelled capacity estimate, not a count of anything that exists, and the range is wide: 140 to 720 million full-time-equivalent working hours. Running just 20 percent of it would imply $2.6 to $5.3 trillion a year in spending that does not exist today.
- Schneider agrees to buy PTC at a 42.3 percent premium and calls it an AI deal. The $22.6 billion is PTC’s equity value, and the implied enterprise value is $23.7 billion. The deal is agreed, not closed, with completion expected by Q3 2027, and the source gives no AI revenue figure for PTC.
- Anthropic puts $100 million behind training 10,000 Claude deployment engineers. The goal of 10,000 engineers by the end of 2027 is a target, and nobody has completed the programme. Trainees come from partner firms such as Accenture, Morgan Stanley and Novo Nordisk, which also grows the pool of people building on Claude.
Run It Yourself: Models That Stay on Your Own Hardware
Who gets to run AI without renting it? Answers range from a 78-billion-parameter server model to a speech model small enough for a microcontroller.
- Aleph Alpha releases Kolibri, a free model you can run on your own servers. The 78-billion-parameter open-weights model is aimed at buyers who want their data kept in-house. Every benchmark score so far comes from Aleph Alpha itself.
- A 16.9 MB speech model aims to put transcription on a microcontroller. Cactus Compute’s open Whistle model hears seven languages with no network connection. The claim that it beats larger rivals on several English benchmarks comes from Cactus Compute’s own tests.
- Australia should stop renting AI and learn to choose it, writes Kate Carruthers. Carruthers argues that national independence means small, specialised, open-weight models and a workable exit plan, instead of a flagship chatbot or a long contract with a US lab. It is an opinion essay.
Showing the Work: Research Claims Still Waiting on Outsiders
Two big research labs report results in mathematics and privacy. Both reports come from the labs that built the systems.
- Meta says Muse Spark helped mathematicians answer five open questions. Six papers credit the model with real proof work, but there is no independent verification and no peer review. The reviewers are collaborators inside Meta’s own project, so this is Meta reporting on Meta’s model.
- Google lets outside auditors inspect the server code behind keyboard learning. Google Research rebuilt its federated learning system so third parties can check the code that handles phone data. The speed and accuracy gains are Google’s own measurements.
Behind the Curtain: The Agent Plumbing Nobody Sees
Agents need an identity, a place to run and a queue for compute. Three vendors pitch their answers, each with its own numbers.
- World pitches its ID system as the fix for AI assistants shopping online. World says sites will soon demand proof that a human approved each assistant. It sells that proof itself, and no standard or outside check exists yet.
- AI21 swaps GPU haggling in chat for a queue on a 10,000-chip cluster. The 83 percent figure comes from a Google Cloud case study and measures queue wait before a job starts. It says nothing about training speed or cost.
- Prime Intellect opens its in-house model hosting to paying customers. The open-model lab is renting out the serving system it built for itself. Every speed and uptime figure so far is Prime Intellect’s own.
Quick Hits
- Vx pitches a language that catches chip-memory mistakes before launch. Vx, a programming language, says its compiler can reject a class of hardware memory errors at build time. The project page names no author and carries no date, and nothing on it has been independently evaluated.
- An image model learns by grading its own pictures, a preprint claims. A preprint called UniEvo-VL has one image model play both teacher and student, reporting modest gains on two image-generation benchmarks. It has not been peer reviewed, and nobody has independently checked the results.
- Meta releases a kit to build your own Muse gadgets, with a warning. According to The Verge, Meta’s open source kit links its Muse agent to screens, buttons and motors, and the company tells builders to proceed at their own risk.
- Rayan Krishnan: AI’s hard problem is picking what to optimise for. In an X essay, Krishnan argues that machines now climb any clearly scored hill fast, so the scarce human job is deciding which hills deserve climbing. It is an opinion piece.