The White House and six major AI companies have signed a voluntary safety accord that asks the companies to oversee themselves through four layers of internal and external review. The two-page document covers Google, Anthropic, Meta, OpenAI, xAI and Nvidia. It calls for technical controls, internal monitoring teams, outside auditors and board-level committees, but it sets no legal penalties or public disclosure requirement. Why it matters A policy official deciding whether the accord fills a regulatory gap has to separate the controls it names from the authority it lacks. Companies are promising a review structure; they are not accepting a government inspection regime or an enforceable standard. ...

Hewlett Packard Enterprise has won a $1.2 billion order from Vultr for AMD Helios AI racks, the first customer order HPE has announced for the system. Vultr plans to install the racks in US cloud data centres for model training and inference. Each rack combines 72 AMD accelerators with HPE’s Juniper networking, CPUs, network cards, cooling and AMD’s ROCm software. Why it matters A cloud operator adding AI capacity is buying an integrated rack rather than assembling chips, switches and cooling separately. The order gives AMD and HPE a large deployment in a market where Nvidia’s rack-scale systems set the commercial reference point. ...

ElevenLabs has completed a $300 million employee tender offer at a $22 billion valuation, twice the price attached to its February financing. The transaction lets employees and existing investors sell shares to new buyers. It is a secondary sale, so the headline amount does not represent $300 million of new operating cash for the voice-AI company. Why it matters An engineer weighing an offer from a private AI company cares about whether stock can become cash before an IPO. ElevenLabs is using a large tender to provide that liquidity while competing with much larger laboratories for specialised staff. ...

The Dutch agency collecting mandatory data-centre reports lists 104 facilities, while the national industry association counts 186 commercial sites large enough to report. The gap comes from a year-long investigation by Lighthouse Reports and European media partners into energy and water disclosure. It shows that a legal reporting system can exist without giving the public a complete facility list or usable environmental data. Why it matters A city planner deciding whether the grid can support housing, industry and a new AI facility needs actual consumption records, not an estimate built from an incomplete registry. Missing sites make regional planning and public comparison less reliable. ...

Open-source developers pushed AI toward smaller machines and stranger architectures, while researchers and forum users argued about who should control machine-generated mathematics and government chat systems. The strongest community items are demonstrations and early releases, not independently reproduced results. Their value lies in the design choices they expose and the questions practitioners are asking around them. In brief PSSA tests a self-modifying language model in Rust Developer Sparticle62ops released PSSA, a small recurrent state-space language model written without a machine-learning framework. The repository says it updates part of its weights while running and outpaces a matched transformer on one CPU test; those performance claims remain author-reported. It deserves attention because the project makes an unusual architecture inspectable at hobbyist scale. Direct source: https://github.com/Sparticle62ops/pssa ...

Anthropic reported on September 29 that Z.ai’s open-weight GLM-5.3 model could build working browser exploits and that researchers could substantially weaken its refusal safeguards. The finding concerns a model that anyone can download, modify and run. Anthropic’s Frontier Red Team tested whether GLM-5.3 could turn known software defects into working attacks and whether it would follow explicitly harmful instructions after common safeguard-bypass techniques. Anthropic researchers Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao and Tripp Gallagher authored the report. The team ran models in isolated environments and combined automated benchmarks with sessions in which security researchers directed the model while examining unfamiliar software targets. ...

OpenAI launched Dots on September 29, introducing persistent agents that run on cloud computers, use connected applications and continue assigned work when the user is absent. The product changes the unit of interaction from a conversation to an ongoing working relationship. A Dot can keep several projects active, retain context across ChatGPT, Slack and Microsoft Teams, and contact its user with progress reports or decisions that require approval. Why it matters: A worker delegating a continuing responsibility gives the system more time, context and opportunity to act than a single chat permits. That makes permissions, audit records and interruption controls part of the product rather than optional deployment work. ...

OpenAI released GPT-6.1 Sol on September 29, pricing the model at $2 per million input tokens and $10 per million output tokens through its API. The company positions the model between its existing GPT-6 Sol and flagship GPT-6 Astra. OpenAI says the new version approaches Astra on coding, computer-use and professional-work evaluations while charging one-fifth of Astra’s standard input and output prices. Why it matters: Developers running agents repeatedly pay for long prompts, tool results and generated output. A lower price at near-flagship capability can change which model they leave active for routine work and which tasks still justify Astra. ...

A British Transport Police facial-recognition trial scanned more than half a million faces but produced one incorrect watchlist alert and no arrests caused by an alert, the Guardian reported on September 29. The six-month pilot covered 18 deployments at busy London railway stations between February and July. Records obtained through a freedom-of-information request put equipment and staffing costs at £320,786 and police time at almost 100 hours. Why it matters: Rail passengers had their biometric data processed at scale, while the system generated no correct watchlist match during the reported period. That result gives lawmakers and oversight bodies a concrete deployment record for judging whether the intrusion was proportionate. ...