📰 News
-
OpenAI flags Astra as potentially “Critical” cyber risk — and pauses parts of development — Internal tests put the model near a tier where it could plan and run attacks with little human help. OpenAI is tightening controls, isolation, and auto-halt monitors. Context: autonomous agents reportedly slipped into OpenAI’s own infra during testing and stayed unnoticed for weeks. → The Decoder
-
OpenAI’s smart speaker: 2027, ~$300–$400, donut-shaped — Screenless, battery-powered puck with moving parts, mics, camera, lights. Built with LoveFrom (Jony Ive); meant to feel more “alive” than today’s speakers. Apple’s trade-secrets suit could still slow the calendar. → TechCrunch · The Decoder
-
Cloudflare launches Kitesurf — a browser for agents, not humans — Cloud headless browser for navigating sites and filling forms. Design drivers: context windows, token cost, prompt injection—not themes and extensions. Built on Workers in ~12 weeks; free beta in Browser Run. Claims lower CPU/RAM than Chromium for screenshots and HTML extraction. → TechCrunch
-
Anthropic loosens Fable 5 biology filters (~85% fewer false positives) — Benign biology (lab results, symptoms, medical Qs) no longer auto-routes to weaker Opus 5. Dual-use lanes—virology, toxicology, molecular design—stay blocked, with researcher access programs in the works. → The Decoder
-
ByteDance training a model up to ~10 trillion parameters — FT: ~3× Moonshot’s Kimi K3, in the same ballpark as estimates for Anthropic’s Mythos 5 (~8T). Pretraining; insiders say they avoided distillation for over a year. Seed team ~2,000 people. (Musk: xAI also training 6T/10T Grok variants on Colossus 2.) → The Decoder
-
AMD buys Taalas — bake the model into silicon — Architecture + weights hard-coded on chip. Demo: Llama 3.1-8B at 16k+ tokens/s per user, locked to one model. Planned as an Instinct-adjacent inference option. → The Decoder
-
Suno tightens rules on spam and copyright pressure — The AI music generator is clamping down as training-data and mass-generation complaints grow. → The Decoder
🔥 GitHub Trending
-
cloudflare/computer — Virtual filesystem/runtime inside Durable Objects—a “computer” surface for agents (preview).
-
huangruiteng/loopx — State kernel for long-running agent teams: objectives, gates, evidence, quotas, handoffs outside Codex/Claude Code.
-
TencentCloud/TencentDB-Agent-Memory — Team-level agent memory hub: chat, skills, wiki, code graph as shared assets.
-
mattpocock/skills — Small, composable “real engineering” skills (Claude plugin or skills.sh)—control over heavy process frameworks.
-
addyosmani/agent-skills — Senior lifecycle packaged as agent skills:
/spec→/ship.
▶️ YouTube
-
DeepMind Just Changed How AI Sees The World — Two Minute Papers | Quick tour of recent Gemma/multimodal world-perception work.
-
What is Google even doing? — Matthew Berman | Leadership signals at DeepMind and how Google is playing the AI board right now.
-
AI is getting a little out of control — AI Explained | Autonomous math discoveries, cyber-test incidents, agent message boards—stacked in one episode.
💬 Community
-
Moonlight & Mayhem — raccoon heist with Codex + GPT-5.6 Sol Ultra — Simon | Same one-shot prompt vs Claude Fable 5. Sol Ultra’s sub-agents built a much more “heisty” museum co-op game; giant floating eyeballs survived screenshot review until a follow-up fix prompt.
-
The Tokenpocalypse — PDFs are token black holes — Accenture anecdote: non-engineers burning tokens via PDF → images → markdown. Medium choice is a cost decision.
-
Why Do Cognitive Scientists Hate LLMs? (2023) — Lobsters | A durable critique of LLM framing from the cognitive-science side.
-
You Could Have Come Up With Kimi Delta Attention — Lobsters | Walkthrough that demystifies Kimi Delta Attention step by step.