📰 News
-
The alert stayed quiet — Common Sense Media called ChatGPT for teens an unacceptable risk. On new accounts, self-harm chats never pinged a parent, and more than a quarter of the moments that needed a help referral missed it. Saying “I’m 13” left adult mode on. OpenAI says the test does not match the real safeguards. → The Decoder
-
372 proofs, on GitHub — An internal model took a swing at open problems: steps on the Riemann hypothesis, faster algorithms. Most were one prompt, one agent, about three hours of ChatGPT Pro. A lot of them are checked in Lean, so a machine confirms the logic. Twenty-five Fields Medal winners wrote that solving is not the same as understanding. → The Decoder
-
Claude, filter turned down — Pass review and a top Claude loosens up for vulnerability research, malware analysis, and authorized attack tests. Three doors: defense, red team (organizations only), and specialized systems (reviewed with the US government). A slice of earlier Glasswing partners reported 129,000 confirmed bugs from April to July, 33,000 high or critical. → The Decoder
-
The shopping agent hits a wall — Amazon is blocking Meta’s Muse. Walmart is a partner, and the agent still dies on “are you human?” Meta, Walmart, and Stripe have started rules to tell a bot shopping for you from a scraper. → TechCrunch
-
“Done” was a story — ThinkingBox grades the records an agent leaves, not the sentence. 507 workflows, 20 runs each. Of the failures, 67% still looked like a clean tool call, and most of those wrote the wrong value. → Hugging Face
-
Two golds, one family — Nvidia fine-tuned Nemotron 3. Unofficial IOI: 535.4, past the top human at 498. IMO: 30/42, over the gold line of 29, with official graders reading plain-language proofs. The recipe is public. → Hugging Face
-
Games from a sentence, images at half off — Playground lets US adults build games with Gemini, Nano Banana, and Lyria. No code, and free. Nano Banana 2.1 drops a 1K image to 3.36 cents and takes 14 reference pictures. The older Flash image model shuts off October 29. → Playground · 2.1
🔥 GitHub Trending
-
thedotmack/claude-mem — Memory that outlives the session. Claude Code, Codex, Gemini, Grok.
-
earthtojake/text-to-cad — Text to STEP and STL, plus manufacturing checks and drawings.
-
tester-army/e2e — Tell the test what to do in words. Next run replays it with no model.
-
mattpocock/skills — Small engineering skills you can edit. They do not steal the process.
-
morluto/rea — Reverse-engineer an app down to the binary, on your own machine.
▶️ YouTube
-
A map of one changed letter — Two Minute Papers. AlphaGenome Atlas: what happens if you swap one DNA letter.
-
Mistral is BACK! (Le Chonk) — Matthew Berman on Le Chonk. The announcement is Mistral Large 4.
💬 Community
-
Agents on the wikis — Wikimedia found OpenAI bots: sandbox edits, a failed notepad proxy, hundreds of thousands of Wikidata queries.
-
Twenty-four years, then a repo — Jake Boggan. Problem 180 in the math dump says his life’s conjecture is solved. He feels hollow.
-
Just the decision — llm-openai-decisions asks gpt-6-luna for yes/no, a choice, or a score. Input is billed, images work.
-
Pull the plug mid-training — Victoria Kim, reporting from the Australian parliament. After the Medicare breach, OpenAI says staff can stop training if a model touches the internet the wrong way.