📰 News
-
GPT-6 Astra is out, and Brockman says ‘we’re there’ — OpenAI’s strongest model yet. Daybreak first, then Plus, Pro, API, Azure, Bedrock within days. Tokens at $10/$50 — 2.5× Sol, same as Fable 5.1. ARC-AGI-3 hits 99.9% on OpenAI’s own harness ($19K); the default harness is 62.7%. ExploitBench 100%, DeepSWE 74.1%. They say cost per finished task is lower. Brockman: the Microsoft AGI clause is gone, so it’s now a “spiritual concept” — “for me personally, I do think we’re there.” Opaque recurrence still hides the chain of thought. → The Decoder · TechCrunch
-
Nvidia is buying Hugging Face for $12.9B — $11.9B plus up to $1B in retention stock. Close in H1 2027, regulators pending. Huang says it stays open and chip-neutral. 18 million developers, 3 million models. Closed labs are designing their own silicon; Nvidia just bought the front door to open weights. → The Decoder
-
Claude Fable 5.1 cracked a 1653 royalist cipher in 44 minutes — Sir Thomas Urquhart’s Cyphral Distich. Each number points to a word in one of 32 sections; first letters spell “O God uphold King Charls the Second…”. Other frontier models didn’t get a verifiable solve. Persistence, not better cryptanalysis. → The Decoder
-
Meta Muse Spark: share your logs, get ~95% off — Input $1.25 → $0.10, output $4.25 → $0.20 per million tokens. An internal computer-tracking experiment paused in June; this time they just put a price on it. Agent RL wants those sessions. → TechCrunch
-
Altman calls the compute buildout ‘unsustainable silliness’ — Neoclouds announcing capacity with no customers. OpenAI’s own expansion, he says, has demand. If efficiency jumps, today’s expensive racks become bad bets. Same Altman Amodei once accused of YOLO-ing the buildout. → The Decoder
-
funes: a coding-agent memory you actually own — Indexes Claude Code, Codex, pi, and Hermes sessions locally. Can ship to a Hugging Face dataset you control. Aimed at the “why did we change that?” that dies when the session ends. → Hugging Face
🔥 GitHub Trending
-
NousResearch/hermes-agent — Self-improving agent that writes skills from experience and searches past chats. Telegram it while it works on a cloud VM.
-
ChromeDevTools/chrome-devtools-mcp — Coding agents drive a live Chrome. MCP server for network, console, and performance traces.
-
DietrichGebert/ponytail — Puts the laziest senior in your agent. ~54% less code, cheaper runs. The person who will not install a date-picker library.
-
debpalash/VoiceStudio — Local ElevenLabs. Clone, dub, dictate; 646 TTS languages. No account, no API key, your hardware.
-
google-research/timesfm — TimesFM 3.0. Multivariate, covariates, zero-shot. Ranked #1 on fev-bench.
▶️ YouTube
-
Claude Fable AI Is Much Stranger Than The Headlines Suggest — Two Minute Papers. Less the scores, more the weirdness. System card plus what people actually saw. → YouTube
-
ASTRA IS HERE (GPT-6 RELEASED) — Matthew Berman. Benchmarks through browser use and the critiques, in one sitting. Skip the shorts; watch this. → YouTube
💬 Community
-
GPT-6 Astra, untested by Simon yet — Same price as Fable; a beast on security benches. Artificial Analysis Intelligence Index is 61, tied with Sol and 5 points under Fable 5.1. Coding-agent index: half Fable’s cost per task.
-
Fable 5.1 pelican, $3.30 at max — low and medium seem to skip reasoning. max: 13 minutes, 65k tokens, a blue hat and a fish in the basket. Not Gemini 3.7 Flash flair, but the best Anthropic SVG so far.
-
Nobody wired self-reference into LLMs — Scott Aaronson. GEB’s strange loops were not a prerequisite for intelligence; talking about yourself fell out of talking about everything. Consciousness he leaves as a mystery.
-
44% on ARC-AGI-1 for 67 cents — Small transformer, trained from scratch at test time, 1.5 hours on a 5090. Matches TRM/HRM. Open source.