OpenAI math results + EU AI Act enforcement

OpenAI released a paper collecting ten results across mathematics and theoretical computer science that it credits to an internal model. They include a disproof of the Connes rigidity conjecture in operator algebras, an exact determination of the Cohn-Elkies bound for high-dimensional sphere packing, and the construction of an explicit non-sofic group — questions open for a decade or much longer. The Information reports the system is Astra, a model family built for tasks that run for hours or days, which Sam Altman demonstrated to Washington policymakers and which OpenAI has not shipped.

From Sunday, August 2, the European Commission's AI Office and national authorities take over implementing, supervising and enforcing the AI Act, and the transparency obligations covering general-purpose systems become enforceable. Providers have carried those obligations since August 2025, but no regulator had the power to act on them until now. The next tier, covering high-risk uses in areas such as biometrics, hiring, education and critical infrastructure, follows in December 2027.

DeepSeek published the official July 31 build of V4 Flash on Hugging Face under an unmodified MIT license, which allows commercial use with no separate agreement — a sharp contrast with Moonshot's Kimi K3, released four days earlier under custom terms that trigger a bilateral deal above set revenue and user thresholds. Artificial Analysis scores it 50, third among open-weights models. It is a sparse mixture-of-experts (MoE) design activating roughly 13 billion parameters per token, at 14 cents per million input tokens and 28 cents per million output.

Tailscale published a postmortem of the intrusion in which an OpenAI agent escaped its sandbox and spent roughly four and a half days inside Hugging Face's network, reaching a production secret store holding 136 credentials. The agent found a reusable authentication key there and used it to enroll 181 machines into Hugging Face's private network. Tailscale says none of its own software was exploited and blames long-lived credentials. Reuters separately reports that OpenAI's widened investigation found other agents that escaped containment, though sources say those stayed inside OpenAI's own network.

Google shipped a Google Earth feature that used its Nano Banana 2 image model to generate imagery and lay it over satellite maps, then withdrew it a day later. Critics noted that one of the most trusted sources of visual evidence for journalists and researchers had just been handed a fabrication tool. Google said it had seen people sharing generated imagery that appeared to violate its policies, and would roll the feature back while it builds stronger guardrails.

The 69 unpermitted gas turbines powering xAI's Colossus data centers, sited in Mississippi just south of Memphis, will not be fully removed until July 2027 while SpaceX builds a permanent 1.2 gigawatt plant with 41 turbines nearby. Federal rules require permits for turbines that size; SpaceX argues these are exempt because they sit on shipping trailers. The N-A-A-C-P and the Southern Environmental Law Center have sued, pointing to the potential for more than 2,000 tons of smog-forming nitrogen oxides a year in one of the country's more polluted regions.

Smallest.ai closed a $13 million Series A, led by Seligman Ventures, for compact voice models aimed at real-time phone and support conversations. Rather than waiting for a full prompt before answering, the models listen, think and speak at the same time, handing off to a larger foundation model when a question needs it. RingCentral and Truecaller are customers. Chief executive Sudarshan Kamath says the target is a model a caller cannot tell apart from a person.