Loose posts

A new study argues debate plus voting only picks the most presentable wrong answer, and that having agents critique and patch each other's work can solve problems every single model gets wrong.

Given a loosely described request, the model produced a ready-to-merge pull request with a video demo of the changes, and the author says the AI-made PRs beat his hand-written ones.

XYEval injects one confident, misleading hint into tasks from tau2-bench, SWE-bench and others, and Gemini, Claude Opus 4.8 and GPT 5.5 scores drop by up to 46.7%, with agents often arguing against the hint in their reasoning and then following it anyway.

Google DeepMind chief Koray Kavukcuoglu said the company trusts AI agents to autonomously conduct experiments, analyze results and propose new hypotheses under human supervision, something he said wasn't happening six months ago.

After finding himself playing project manager to three agents in three tabs, the author looks at Hyperagent Rooms, a shared space where several specialist agents and several people work the same job in parallel.

An argument that agents should call editing primitives like clips, keyframes, layers, timing, motion and sound directly instead of mimicking a human with a mouse, which makes taste rather than software knowledge the moat.

Odyssey trained its two-model world simulator on real gameplay footage and player inputs, with humans and AI bots that chase them sharing one map.

Citing the Codex team's two PMs, one designer and roughly 40 engineers covering 10-plus surfaces, the author argues code generation is no longer the bottleneck and the scarce skill is defining what counts as correct before building.

A new Microsoft and City University of Hong Kong paper calls an agent's ability to pick the right next step on long tasks "taste" and built a 502-fork decision test set, on which the strongest model scored only 59.7%.

The author argues the real takeaway is the process of large-scale parallel exploration, anomaly detection and human verification, not the discovery itself.

A new method from Google and others uses an optimized harness only during training, raising macro task success from 23.3% to 44.3% without a specialized harness at deployment.

Korea University and Microsoft Research Asia measured time and CPU use across three agent types, finding that adding cores made a question-answering agent 13x faster but a web search agent only 4% faster.