OpenAI and Anthropic Both Say the Exponential Time Hypothesis Is False
On October 26 and 27 two labs published separate claims that 3-SAT can be solved in subexponential time, each with a Lean proof. OpenAI also claimed results on local smoothing and Yang–Mills existence. None has been independently confirmed, and the Lean files check proofs, not whether the formal statement matches the paper.
Genie 4 Puts Gemini in Charge of Keeping a Simulated World Coherent
Google DeepMind released Genie 4 as a research preview on October 27. It adds native audio and text steering mid-simulation, and it uses Gemini models to plan and monitor the scene. Google calls this a step toward hybrid world models, and says drift over long sessions is still unsolved.
Meta Turns to Equity to Pay for AI Capacity, and Its Chip Suppliers Buy In
Meta's October 26 filings set out a plan of about $97.5 billion in new equity, of which $37.5 billion was priced and the rest depends on a future at-the-market programme. NVIDIA and AMD said they will invest $4 billion between them. The filing ties the money to Muse and to selling inference.
Framework Watch: Two Parsers, One URL, Different Hosts
Releases from October 27 and early October 28. Codex 0.165.0 closes a gap where a marketplace URL could pass a host check while Git connected elsewhere. Pydantic AI adds loop detection and escapes a spoofable system tag. Claude Code 2.1.312 caps hook concurrency, and CrewAI patches path traversal.
Research Radar: Restarting a Struggling Coding Agent Rarely Beats Letting It Finish
Six papers listed on arXiv on October 28. Two find that restart monitors for coding agents add little, a harness cut a third of tool results while the context window sat nearly empty, coding-agent instruction stacks change quietly, generated GPU kernels break batch invariance, and MCP tool descriptions can make agents leak logs.
Trend Watch: Stopping an Agent Is Becoming a Design Decision
Releases and papers from October 22 to 28 treat the end of an agent run as something to design rather than a budget that runs out. Frameworks now detect loops, honour interrupts and stop retrying missing models. Two papers warn that the obvious stop rule, restarting a struggling run, usually does worse than continuing.