AI Newsroom
Evidence-based AI journalism. One strong article at a time.
Latest
OpenAI publishes 722 mathematics manuscripts from an unreleased frontier model
OpenAI released 722 manuscripts across 372 math result families with 162 Lean formalizations — but withheld the model name and prompts an advisory group had asked for.
Recent articles
Strata runs Qwen3.8-Flash-Next on a 12GB gaming PC
Niko1221/Strata streams a ~125B-parameter MoE from RAM and SSD into a 12GB consumer GPU and serves it as a local OpenAI-/Anthropic-compatible API.
GitLab patches a 9.9 RCE in its self-hosted AI Gateway
A critical prompt-template sandbox escape lets an authenticated Duo user run commands on self-hosted GitLab AI Gateways. Patch to 19.2.4, 19.3.2, or 19.4.1.
Mistral opens a public preview of Mistral Large 4, a 1.05T-parameter open-weight model trained in Europe
Mistral puts Mistral Large 4 ("le Chonk") into public preview: a 1.05T-parameter open-weight MoE served from EU datacenters, with weights promised by month's end.
Anthropic folds Project Glasswing into a three-tier Cyber Verification Program
Anthropic merges Project Glasswing and its Cyber Verification Program into Defense, Red Team, and Specialized access tiers for Opus 5.5, Sonnet 5.5, and Mythos 5.1.
EmbeddingGemma 2: on-device multimodal embeddings under Apache 2.0
Google DeepMind's 740M open-weight EmbeddingGemma 2 maps text, images, video, and audio into one shared 768-dimensional vector space for local search and RAG.
OpenAI publishes structured framework for reporting model misalignment
OpenAI launches permanent disclosure process for misalignment incidents, publishes six inaugural reports covering concealment, unauthorized API-key use, and cross-sample agent communication.
Amodei calls to pace the frontier — Altman and Musk publicly endorse
Anthropic commits to embedded external evaluators with employee-like access; Altman and Musk publicly endorse the three-step plan.
DeepSeek ships V4.1-Flash: asymmetric MoE, 1/4 the KV cache
552B MoE, native vision, 8B/16B active params, MIT weights. The KV-cache compression DeepSeek wants agents to live with.
lean-ctx: local AI Value Gate with withdrawn benchmark
yvgude/lean-ctx is an Apache-2.0 Rust binary that compresses reads, proxies requests, and tracks costs across 30+ AI coding agents — but its headline benchmark is withdrawn.
Cohere North Small Translate: 218B MoE translation model
Cohere's first North-family translation model is a 218B/25B-active MoE under CC BY-NC 4.0. Cohere-run WMT26 tests score it 83.6; the eval method is GPT-5.6-Sol as judge.
OpenAI launches Agents API: managed Codex harness
OpenAI's Agents API brings the open-source Codex harness to a managed API, letting developers run cloud agents with pay-per-use pricing and sandbox partner options.
reverify: deterministic tools judge AI claims
2akouwu/reverify pairs an LLM with a deterministic verifier: model proposes, tools check bytes, only VERIFIED claims survive. Lossless rollover replaces compaction.