Tools & Demos 1
The release bundle adds ChatGPT Space for shared human-agent work, cloud Codex environments, continuous repository scanning, a low-latency Decisions API, computer use in the Agents API, plugin extensions, marketplace distribution and an Ultrafast tier reaching a claimed 300 tokens per second.
Lab Updates 1
Its draft framework calls for evidence on alignment, containment and monitoring; immutable transcripts; independent dissents; multiple executives with veto power; fail-closed controls; and public postmortems after severe incidents. OpenAI says these practices are still being implemented, not yet a completed external standard.
Podcasts 1
In a 92-minute Latent Space interview, he explains Claude Mods, cheap forked supervisors, multiplayer permissions and why better initial context can save more tokens than repeated correction. He also separates model safeguards, user-intent review and sandboxing as distinct security layers.
TLDR: Projects and Mods turn coding agents into customisable execution systems rather than fixed chat loops. Shihipar's counterintuitive claim is that smarter models can be cheaper on simple work because they need fewer retries and verification steps. Practical advice: state intent and constraints upfront, preserve implementation decisions, and enforce permissions outside the model. Key sections: 00:20:53, 00:23:49, 00:32:49, 01:23:26 and 01:28:59.
HN Threads 2
EFF says the company trains a model on first-party betting records, identifies losing gamblers and sends promotions intended to bring them back. The 529-point HN thread treated this as a clearer near-term AI-regulation case than speculative superintelligence risk; DraftKings' underlying methodology was reported by The New York Times.
The estimate assumes annual AI infrastructure investment reaches $1.5 trillion and represents roughly a quarter of industry revenue. HN's 284-comment debate challenged whether this means direct AI sales or broader enabled economic activity, and whether the buildout is an overshoot or early infrastructure.
Industry 1
CEO Arthur Mensch says capable agents need containment and monitoring, but argues recent U.S. safety rhetoric obscures operational failures rather than justifying slower model development. The stance also serves Mistral's competitive interest as it raises capital to narrow the capability gap.
HF Trending Papers 5
GRAFT replaces all-fail rollout groups with compatible trajectories from a peer model, using weighting and clipping to limit off-policy mismatch. Across three heterogeneous pairs, both models improve by 2.1 points on average at unchanged per-model rollout budgets (12 HF upvotes).
Simple-WAM finds that discarding future representations at inference removes the generalisation benefit of world-action training, while one noisy forward pass preserves it. The authors report better generalisation than explicit video-denoising models at efficiency comparable to latent variants (10 HF upvotes).
A coupled rollout keeps accepted student tokens under reverse-KL training and applies direct teacher correction at disagreement points. The implementation preserves the target distribution while reporting 4.22× higher matched-workload rollout throughput (7 HF upvotes).
The system preserves executable programs, edit history and rendered evidence so an agent can revise images and video rather than stop when code merely runs. GPT-6 Astra met every generation-success check, but only 76.9% of video tasks met all quality thresholds (4 HF upvotes).
AnisoWM replaces a fixed spherical latent target with learned directional variance while leaving the predictor and Euclidean planner unchanged. It improves planning success over LeWorldModel in four visual-control environments, according to the preprint (5 HF upvotes).