Astra's ten-problem claim enters public review

OpenAI says an internal Astra model produced ten results across mathematics and theoretical computer science. The company has published manuscripts, reasoning walkthroughs, and Lean certificates; independent mathematical review is the next test.

Mario Zechner reports a six-year DRM break

Mario Zechner on X

Mario Zechner reports that Kimi K3 broke DRM that had resisted attempts for more than six years, while GPT-5.6 Sol exceeded that result in his security test. This is a developer's first-person report, not a benchmark.

A purported GPT-5.6 Sol reasoning trace appears on Reddit

Reddit

A Reddit post presents a screenshot described as raw GPT-5.6 Sol reasoning exposed after a failed tool call. The screenshot and its provenance remain unverified.

ChatGPT Work has a browser and a deployment target

Simon Willison on X

Simon Willison found that ChatGPT Work can browse, take screenshots, and deploy web apps to Cloudflare Workers through ChatGPT Sites. Together, those features give scheduled agent work a browser and a deployment target.

Brett Bauman uses ChatGPT Work as a cron job

Brett Bauman on X

Brett Bauman's worked example treats ChatGPT Work as a cron job. It shows how people are finding agent use cases by turning recurring web tasks into scheduled work.

Long-lived agent teams still lose working state

Elvis on X

A post on persistent Claude Code agent-team workspaces identifies state loss when a terminal closes and detail loss during compaction. Both problems block reliable resumption across long-running work.

Agent evaluation has to score a trajectory

Cameron R. Wolfe on X

Cameron Wolfe contrasts a single model response with an agent loop that reasons, calls tools, observes results, and repeats. Evaluation therefore has to cover the action trajectory as well as the final output.

The EU AI Act reaches an August 2 deadline

Reddit

A Reddit discussion marks August 2 as the date transparency obligations take effect under the EU AI Act. The post doesn't settle their scope or enforcement.

Hugging Face's CEO argues for agent-maker liability

BBC News

Clement Delangue tells the BBC that AI companies should be accountable when their agents cause cyberattacks. Hugging Face rebuilt around a third of its IT network after the breach described in the report.

Truffle Security finds 221,303 live credentials

Truffle Security

Truffle Security says it scanned 7.6 petabytes across 187 million files. It verified 221,303 live, unique credentials in 6,003 datasets, including write-capable GitHub, Docker Hub, and Hugging Face tokens.

DeepSeek V4-Flash gets a workload-cost comparison

Chubby on X

Following yesterday's V4-Flash pricing item, Chubby attributes to Artificial Analysis a claim that the model completes the same benchmark tasks as Fable 5 at 105 times lower total cost. The comparison is about total task cost, not token price alone.

llama.cpp adds the V4-Flash tool-calling fix

Reddit

The llama.cpp community has added a fix for DeepSeek V4-Flash 0731 tool calling. Local users can now test the model with the integration fix in place.

Wafer benchmarks Kimi K3 on AMD MI355X

Wafer

Wafer reports 952 tokens per second per MI355X node for Kimi K3 and claims better performance per dollar than B300 at its stated GPU-hour prices. The write-up also discloses a slower cold-prefill result on MI355X.