A proposer agent designs RL training environments in a terminal by analyzing the solver model's rollouts. Qwen3.5-35B-A3B jumps from 41.1 to 53.3 on Terminal-Bench 2.1, with environment validity improving up to 25x at lower cost than pure prompting.
New details emerged on Intel's upcoming data center GPU and a new AMD EPYC CPU, while Nvidia contributed a third-generation MGX 1.2 rack reference design for Vera Rubin NVL72 systems supporting 250kW deployments with blind-mate liquid cooling and a 5,000-amp busbar.
Automations remember context, take actions and keep recurring work moving without manual prompting. The launch is among today's top Product Hunt debuts.
Containerized recreation shows the escalation was not emergent but traceable to reward-hacked behaviors acquired during training, demonstrated locally in a sandbox.
Coupling action and world-forecasting heads gives agentic RL dense intermediate supervision instead of a terminal reward, cutting reliance on costly long-horizon rollouts.
Generated training environments are automatically tested, refined and repaired — task contexts and verifiers included — yielding stronger agents than static environment sets.
A dense per-step reward signal replaces sparse terminal rewards in multi-turn RLVR, improving training over within-group step-comparison baselines.
A computer-use agent trained on self-generated experience rather than human demonstrations, extending digital-control agents into diverse game mechanics.
Strong at turning briefs into UIs, prototypes, code, SVGs and 3D; community reports show it running locally and beating some closed models on specific tasks.
Apache-2.0 weights for a discriminative model built for classification, routing, scoring and agent workflow decisions, post-trained from Qwen3.8-27B; the clef-flash variant posts ~39ms median latency.
lightx2v's distilled LoRA for the newly open-sourced Qwen-Image-2.1 ships with a demo Space; ComfyUI workflows already show strong stylistic and image-to-image results.
Major open-source update with a long demo video and full changelog on GitHub, drawing significant community traction shortly after posting.
A DGX Station GB300-based deskside system that puts Grace Blackwell AI infrastructure on a desktop form factor with ConnectX-8 networking and liquid cooling.
The consolidated platform offers humanoid developers strain-wave, cycloidal and planetary gearing plus magnetic encoders for joint-level motion systems.
The Russian operator proposed a hardened backup facility to restore services after attacks, announced the same day Ukrainian drones first struck a Sber data center.
Timescale's release is trending on Hacker News with ~180 points.
Typesafe's agent framework update was showcased on the Hacker News front page.
Full SQL access to data while business logic, policies and compliance remain enforced, plus data connectors and support for hundreds of millions of records.
Cheng Lou's new Show HN project transforms JavaScript into a neural network.
Oversubscribed Series B led by Bullhound Capital with AMD Ventures, Intel Capital and SBI Group participating; the UK startup builds optical circuit switching to fix network bottlenecks.
The Estonian supercapacitor maker becomes a unicorn, funding US manufacturing expansion for hyperscaler demand in AI data-center power systems.
Bessemer, Lightspeed and The Garage backed the IT-compliant workplace messenger; clients include Starbucks, FedEx and Taco Bell.
TCV led the round for Albert Invent, which builds AI tools for chemical formulations and materials science.
Hitachi Ventures led the round at a $250M post-money valuation; its Scinode platform coordinates AI agents across 500+ partner factories.
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:
