Pareto matches Astra's top benchmark score at a third of the cost

A hobby project runs an AI agent called Jev through Pokémon Red, streaming every decision and its odds in a live feed. The creator says Jev decides fast but not fast enough to play Doom, and credits a guide for knowing where to go next.

Documents obtained by the EFF via FOIA show the WISeR pilot, launched in January, was rushed and error-ridden, with one vendor warning CMS a working product by launch was unrealistic. It runs in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington through 2031.

A C++ CUDA test harness hands kernels to LangGraph-based AI agents that run them, collect benchmarks, and profile via Nsight. Built by bertaye as a side project to learn LangGraph.

CLM-8B is the first open model in a new class called Contrastive Language Models: it does not generate text, but scores candidate actions against the current state and returns probabilities. Tests show it runs up to 9x faster than Jev, the proprietary System One model it targets as its main baseline.

A LocalLLaMA poster worked through the rent-vs-buy decision for an 8-GPU HGX H200 server, saying the break-even point landed somewhere they did not expect. The post shares the full calculation and invites corrections.

Google's Product Security team disclosed PageBreak, an internal Gemini-based agent that has uncovered more than 500 XSS vulnerabilities in Google's first-party web apps. It only reports a bug after confirming it with a working exploit against a live environment, giving it a near-zero false-positive rate.