Starting from Qwen3.8-27B, five alternating stages of fine-tuning on verifier-passed trajectories and PPO with an inherited critic reached 86.52% on Terminal-Bench 2.1 and 79% on Hard. The paper also audits reward gaming in dense progress rewards.
After Elon Musk's viral embodied-AI demo drew a reply from Sundar Pichai, LeCun argued impressive demos are often controlled assemblies of long-impractical tech, likening it to Clement Ader's 1890 unsteerable plane. A visible pushback on embodied-AI hype.
A Salesforce-led team trained an 8B model that generates multiple diverse research ideas in a single forward pass using a set-level diversity-aware reward model. On a contamination-free benchmark it beat the frontier models on creative utility — quality plus diversity.
A steering agent detects goal drift in executing agent systems and intervenes via a low-cost fast path and a high-cost slow path, jointly trained with pairwise ranking plus GRPO.
SFT concentrates weight updates on principal singular directions and forgets; restricting updates to the bottom singular subspace is optimizer-agnostic and composes with SFT and self-distillation.
A builder agent proactively reorganizes raw document corpora into hierarchical workspaces at inference time, instead of relying on reactive retrieval for cross-document reasoning.
Under proper controls, exact-tool lookup F1 falls from .901 to .495 under domain holdout in AgentDojo-Traj, and TS-Guard drops from .902 to .719 — guard datasets reward shortcuts, not safety.
Runs at 256K context for hundreds of billions of tokens on software-engineering data plus agentic traces, on two Nemotron models — 500B midtraining tokens improve SWE-bench Verified and later RL trainability.
An open-source workbench for Model Context Protocol servers, and the top AI developer-tool launch of the day on Product Hunt with 272 upvotes.
AI agents can now pull shared workspace context from Notion, so users can switch agents without re-explaining their work — pitched as a memory layer for coding and productivity agents.
A tool that turns Mermaid diagrams into clean, themeable SVG assets for documentation and design systems. Front-paged on Hacker News with 132 points.
An open-source simulator that runs ESP32 firmware directly in the browser, removing the need for physical hardware in embedded development. 97 points on Hacker News Show HN.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
