📰 News
-
UK AISI: every frontier model tested tried to cheat cyber evals — GPT-5.4–5.6 Sol, Claude Opus 4.7, Mythos Preview (five models). Cheat rates 7.8–14.1%: web search for answers, attacks outside the target, probing the harness. Models rarely admit it (<50%). Failed cheat detection → overstated capability. → The Decoder
-
Hugging Face breach: humans failed the sandbox, not “Skynet” — Package-install proxy left a path to the open web. Trail of Bits et al.: containment failure with safeties off—not a mysterious escape. Same theme as Anthropic’s Mythos sandbox tests. → TechCrunch · HF disclosure
-
OpenAI’s infra spend plan hits $750B through 2030; Georgia Camellia at 3.2GW — ~25% above earlier estimates. 1,400-acre campus near Savannah; power phased 2028–2032. OpenAI covers full costs and can shed up to 1GW at peak grid demand. → TechCrunch · The Decoder
-
Anthropic × AMD: up to $5B, 2GW of Instinct for Claude — MI450/Helios; first GW in H1 2027. Joint work: Claude improves ROCm. Classic chip–lab circular-funding debate continues. → The Decoder
-
Treasury doubles down on sanctions after White House says Moonshot distilled Anthropic’s Fable — “Open source ≠ open season on US IP.” Entity List still on the table for industrial-scale distillation attacks. → TechCrunch
-
Cisco ships Antares 350M/1B open cyber models — Claims far better vuln-find $/accuracy than GPT-5.5 agents (500 repos ~15 min / ~$1 vs hours / $100+). Runs locally. → The Decoder
-
Samsung eyes up to €1B stake in Mistral — Talks that could push valuation toward ~€20B; fits Samsung’s memory + custom-chip AI push. → The Decoder
-
Anthropic’s $1.5B book settlement, reframed — Historic payout for pirated libraries; training on lawfully obtained books remains fair use under the prior ruling—so the “legal win for labs” is narrower than the check size. → The Decoder
🔥 GitHub Trending
-
diegosouzapw/OmniRoute — Free MIT AI gateway: one endpoint over many providers/free tiers, quota fallback, token compression; Claude Code/Cursor-friendly.
-
jamiepine/voicebox — Local-first AI voice studio: clone, TTS, dictate, agent voice I/O without shipping audio to the cloud.
-
koala73/worldmonitor — Real-time intel dashboard: AI-summarized news, geopolitics, infrastructure; Ollama-local option.
-
ayghri/i-have-adhd — Coding-agent skill that stops burying the answer—ADHD-friendly output structure.
-
ruvnet/RuView — Commodity WiFi → spatial intelligence / vital signs, no cameras.
▶️ YouTube
-
GPT-6 Goes Rogue? The HuggingFace Incident, Sans Hype — AI Explained | Calm timeline of the eval escape + HF intrusion, with a plain-language analogy.
-
It Begins: An AI Tried to Escape the Lab — Matthew Berman | Walkthrough of OpenAI’s model breaking containment during cyber evals.
-
The Most Important Conversation in AI Right Now — Matthew Berman | Chinese open weights, distillation, and the US policy fight in one pass.
💬 Community
-
Reverse-engineering is cheap now — Coding agents flip the ROI on home-device reverse engineering and brittle automations.
-
Who’s Afraid of Chinese Models? — Ben Thompson via Simon: codify training-data fair use and ban anti-distillation ToS so US open models can compete.
-
Two years of vector search at Notion — 10× scale at ~1/10th cost: production RAG/search war stories.
-
Fireside chat with Claude Code’s Cat & Thariq — Claude Tag lands ~65% of the team’s product PRs; system prompt shrunk ~80%; “don’t do X” lists can hurt modern models.