Laplace-1-Atlas Leads Decision Benchmarks
Juneo Lab's open-weight Laplace-1-Atlas predicts what an action will do inside an organization before deciding who does the work, who sees the information, and what gets approved. It scores 89.4% on TrustDecision, the benchmark closest to its purpose. — huggingface.co
GLM-5.3 Spans Two DGX Stations Locally
NVIDIA's quantized GLM-5.3 spans linked DGX Stations, keeping all of its weights in GPU memory so both machines serve one local endpoint. The setup enables private inference on sensitive data and local agent prototyping without per-token costs. — developer.nvidia.com
Antigravity /teamwork Solves Open Math Problems
Google's Antigravity /teamwork harness pairs frontier models with automated verifiers to resolve problems that had stood open for years, including information theory's Courtade-Kumar conjecture. The tools are now available to outside researchers and developers. — research.google
DSpine Speeds Parallel Speculative Decoding
DSpine speeds up AI text generation by proposing many candidate tokens in parallel while conditioning each position on its predecessor. It achieves the best average acceptance length on Qwen3-8B across standard benchmarks, and its inference code is open-sourced. — huggingface.co
Chinese AI Safety Group Targets Reward Hacking
A young Chinese team has built one of the country's largest AI safety communities and is preparing to audit reinforcement-learning environments for reward hacking, the failure mode in which an agent games flaws in its training setup instead of learning the intended behavior. It will publish regular public reports on the failure modes it finds. — lesswrong.com