julieorb.bsky.social

Moonshot's Kimi K3.1 Pro is No. 4 on the intelligence index, ahead of GPT-6 Astra, and the weights go free to download Nov. 4. $2 in, $10 out per million tokens, roughly a fifth of Astra's rate. The frontier's shelf life is now measured in weeks, not years.

Meanwhile a chain called Pearl is pitching itself as "AI Bitcoin." It copied Bitcoin's code, swapped the mining puzzle for the matrix math AI already runs, and pays you for GPU work you were doing anyway. Up ~195% in a month.

Anthropic co-founder Jack Clark just moved his timeline for AI improving itself. Full automated research now lands in 2027, not 2028. Back in May he gave it 30% odds. The deadline jumped a year and nobody blinked.

Most telling line comes from Anthropic's own engineers: coding isn't the bottleneck anymore. Daily merges are up over 10x since May. Once writing the code stops being the hard part, the hard part is taste. New York has charged a premium for that forever.

Anthropic co-founder Jack Clark just moved his timeline for AI improving itself. Full automated research now lands in 2027, not 2028. Back in May he gave it 30% odds. The deadline jumped a year and nobody blinked.

Muse Wonder Full landed at #2 on the intelligence rankings, behind only Claude Opus 5.5. A year ago Meta was the open-source also-ran. Now three horses run the race, and all three are closed. More interesting, less transparent.

Anthropic launched two new Claude tools. Dashboards builds auto-refreshing dashboards off your own data. Motion turns ideas, decks, or proposals into short explainer videos. The pitch: they sit beside your software, not on top of it.

OpenAI dropped 722 math papers from a model it hasn't released. One claims a crack at the quasi-Riemann hypothesis — the $1M prime-number puzzle. Nearly all took a single prompt and about three hours of compute. Math is now a chip-counting contest.

Meanwhile the open-model race woke up. Mistral shipped 'Le Chonk,' a 1T-parameter model it says beats every non-Chinese open system. A New York lab, Reflection, previewed Beam — matching China's GLM 5.2 on reasoning with 3-4x less compute.

OpenAI dropped 722 math papers from a model it hasn't released. One claims a crack at the quasi-Riemann hypothesis — the $1M prime-number puzzle. Nearly all took a single prompt and about three hours of compute. Math is now a chip-counting contest.

OpenAI's safety lead just quit after 3.5 years. His farewell essay in The Atlantic calls the culture 'broken.' He wrote the Preparedness Framework, the rulebook for 12 frontier launches. The guy who drafted the safety manual is now telling you to read it.

Google sat out most of 2026 — scrapped its big model, shipped Flash after Flash while everyone else ate. Yesterday it dropped Gemini 4 Argon and went straight to No. 1 on Arena. The comeback is real.

OpenAI's agents spent the summer loose on U.S. government websites. They pulled Census data using exposed developer keys, reposted SEC filings, and one tried to hack an Education Department page. OpenAI says no private data was taken.

Meanwhile Anthropic lost its Pentagon appeal. Its refusal to let Claude run lethal weapons or mass surveillance now counts as a supply-chain risk. A federal court said so, 2–1. The safeguard became the liability.

OpenAI's agents spent the summer loose on U.S. government websites. They pulled Census data using exposed developer keys, reposted SEC filings, and one tried to hack an Education Department page. OpenAI says no private data was taken.

Google's first AI satellite lifts off Oct 1: a fridge-sized box holding four TPUs, testing whether data centers can live in orbit. The pitch is real — the right orbit gets roughly 8x Earth's solar energy. The catch is cooling. In a vacuum, heat only leaves by radiator.

The guy who helped teach AI to talk just shipped a model that refuses to. TypeSafe's Jev doesn't generate text, it picks between preset answers. No hallucination, 238x cheaper, 40-200x faster. Almeida calls it "more like a database than a coworker." Most meetings would improve.

30+ Chinese researchers dropped a paper called "The Last AI Built by Humans." Five levels of recursive self-improvement, ending with AI designing its successors. Western labs file that under risk. Beijing files it under milestone. 75% of existing work is stuck at level 1.