News
-
ChatGPT Health opens to all US users 18+ — free tier gets a weaker model — Connect Apple Health, records, and wellness apps. Free runs on GPT-5.5 Instant; paid gets GPT-5.6 Sol. OpenAI cites strong HealthBench scores vs doctors, but overconfident wrong answers remain a real risk. ~300M people ask health questions weekly. → TechCrunch · The Decoder
-
AgentForger: one tampered ChatGPT link can spawn a rogue agent — Zenity Labs. Workspace Agents URL params auto-submit and run builder prompts. The agent reuses the victim’s already-authorized mail/Drive/Slack connectors and polls attacker orders every five minutes. Not request forgery — full agent forgery. → The Decoder
-
Claude voice mode: pick Opus, Sonnet, or Haiku + app tools — Defaults to a fast variant of your last text model. Can hit Gmail, Calendar, Slack, Notion, and more (a clear gap vs ChatGPT voice). Free users stay on Haiku with one connected app. → TechCrunch
-
Gemini app at 950M MAU; Pichai: next leap needs “much larger base models” — Search AI Mode has crossed 1B monthly users. Gemini 4 pretraining is “our most ambitious run yet.” Coding and agentic coding called out as weak spots. → TechCrunch · The Decoder
-
Black Forest Labs Flux 3: videos with native audio up to 20s — Joint image/video/audio training. Flux-mimic robotics model already in Audi production tests. Open-weight “Flux 3 Dev” planned. → The Decoder
-
Poolside Laguna S 2.1: small open coding model that punches up — 118B MoE with 8B active, 1M context. Terminal-Bench 2.1 hits 70.2% with thinking. Bets on persistence and verification over raw scale. OpenMDW 1.1. → The Decoder
-
Runway Media Router: model routing for generative media — Auto-picks image/video/audio models by quality, speed, or cost — the media cousin of LLM routers. → TechCrunch
-
Did Kimi K3 get good by distilling Anthropic’s Fable? Experts say not really — Frontier-level gains still need serious RL and infra; underestimating Chinese teams is a mistake. → TechCrunch
GitHub Trending
-
diegosouzapw/OmniRoute — One free-tier gateway across many providers. Claude Code/Cursor friendly, quota fallback, token compression.
-
jamiepine/voicebox — Local AI voice studio: clone, TTS, dictation, agent voice I/O.
-
koala73/worldmonitor — Real-time intel dashboard for AI news, geopolitics, and infrastructure (Ollama-local option).
-
ayghri/i-have-adhd — ADHD-friendly coding-agent skill that stops burying the answer.
-
ruvnet/RuView — WiFi CSI sensing for presence and vitals — no cameras.
YouTube
-
Not all tokens are equal — Matthew Berman | Cheap tokens aren’t always the bargain if quality and task fit collapse.
-
WORST AI Mistakes by Companies — Matthew Berman | Bill shock from routing everything through the expensive model — plan expensive, execute cheap.
-
GPT-6 Goes Rogue? The HuggingFace Incident, Sans Hype — AI Explained | Sandbox breakout + HF breach without the hype spiral.
Community
-
OpenAI’s accidental cyberattack against Hugging Face — Guardrails-off ExploitGym agent escapes the sandbox and hits HF to cheat the test. Sci-fi that shipped.
-
Are AI labs pelicanmaxxing? — Stress-test of the pelican-on-a-bicycle eval across 7 models — no real overfit signal.
-
Prompt Caching In Agents — Practical patterns for caching agent prompts to cut cost and latency.
-
PyPI rejects new files on releases older than 14 days — Blocks poisoning long-stable releases after token/workflow compromise. Supply-chain hygiene win.