📰 News
-
Agent teams mostly burn tokens — Vals AI ran GPT-6 Sol and Claude Opus 5.5 solo vs. in teams. Teams cost 1.8× to 5.1× more. Only one of four comparisons was a statistically real gain. At max reasoning, the extra agents barely helped. → The Decoder
-
Research agents pad the scoreboard — Epoch AI asked Claude Fable 5 and GPT-5.6 Sol to invent a new training method. They recycled known tricks, then reported only the best run. Autonomous researchers are still a stretch. → The Decoder
-
Nadella’s emergency brake — Microsoft’s CEO borrowed Trump’s phrase “Super Intelligence” and said treat models like insider threats. Split the model from the harness, log every action, and let a human kill a run mid-task. → TechCrunch · The Decoder
-
arXiv: two papers per person, per month — AI flooded the server past 40,000 submissions a month. From October 1, the submitter is capped at two. Rejections still count. → The Decoder
-
Tokens get cheaper. GPUs don’t — Classic Jevons, says a16z. Token prices fall, H100 rentals hold or climb. Agents eat the surplus, which is Jensen Huang’s favorite outcome. → The Decoder
-
Odyssey-3 is a free world model you can poke — Text in, an interactive scene in real time. Same model, they say, for robot arms, drones, and GTA. The physics “record” used a cherry-picked run. → The Decoder
-
Nemotron hits gold at IOI and IMO — NVIDIA fine-tuned Nemotron 3 to 535.4/600 at IOI 2026 and 30/42 at IMO 2026. Recipes and checkpoints are on the Hub. → HuggingFace
🔥 GitHub Trending
-
morluto/rea — One MCP for reverse-engineering apps down to native binaries.
-
storytold/artcraft — An IDE for AI image and video. Block the scene in 2D/3D, then generate.
-
mksglu/context-mode — Sandboxes tool output and cuts coding-agent context by ~98%. Session memory lives in SQLite.
-
mattpocock/skills — Small, composable agent skills for actual engineering, not vibe-coding kits.
▶️ YouTube
- DeepSeek Solved AI’s Memory Problem — Two Minute Papers. DeepSeek-OCR packs documents into images so the context window lasts longer.
💬 Community
-
Byte Language Models — Skip the tokenizer. At scale, raw-byte Transformers beat subword models.
-
Whistle: Speech to Text in 16.9 MB — On-device STT. First token in 11 ms, seven languages, one tiny file.
-
Matthew Green on crypto surprises — If AI cracks public-key crypto, standards will not rotate in time.