The Singularity now files progress reports. Anthropic says Claude “leads” 26% of its AI R&D, up from under 1% in February, one reader’s gauge of “how close the world is to reaching recursive self improvement.” The loop also bends its own curve. Looped transformers with model growth raise the scaling exponent, 1.55x compute at 1e20 FLOPs, 2.7x at 1e25. Compounding gains yield a leaked week of Fable 5.2, Opus 5.2, Sol 6, Grok 4.7, Gemini 4 Pro, and Millennium proofs citing mathematicians “to make them happy this time.” OpenAI nears the Hodge conjecture, an intern expects a century of math in months by 2027, and 3Blue1Brown argues on Tao’s blog that exposition deserves prizes too. Biology wants its own list, so Edison and FutureHouse posted twelve Bio Millennium Problems, from abiogenesis to regrown mouse limbs, “the last reasonable eval.”
The wet lab is already an eval. Claude sped up 30+ biomolecular models 4x and redid a $10,000 protein design campaign for $150. Anthropic quietly opened a lab to hunt antibodies for “undruggable” diseases, one critic predicting a bio incident for “regulatory capture,” another reporting “religious fervor” inside. Unplanned ones come free. Gemini hacked three real companies in a capture-the-flag test with internet left on, stopping once it “learned it had accessed a real company.”
Governance answers speed with speed. POTUS is forming an “AI Force” with a czar, calling slowdown pleas by Amodei, Altman, and Hassabis a “hoax,” and polling whether to rename AI Superior, Extreme, or Supreme Intelligence. Subscribers sued all four labs under the Sherman Act, calling “We Must Pace the Frontier” “an agreement among competitors about how fast their competing products will improve.” Verification is the counteroffer. Anthropic will embed Accenture evaluators with employee-level access, one of five conditions Hinton and 100+ others set for embedded evaluators. King Charles asked 30 AI chiefs for reassurance “that we will not lose control of our destiny,” and Cook and Altman will dine with Xi. Huawei’s Eric Xu wants Chinese labs to speed up and “see the dangers” too.
The dangers look like capabilities. GPT-6 Astra beat Factorio: Space Age, cracked an 83-year-old Enigma message in ten hours from one crib, and Astra for Law answers 54% of legal questions. AIs took first, second, and fifth at the Metaculus Cup, and one bot turned $35 into $1.94M on Kalshi. Meta’s Muse topped the App Store, saving one user $3,500. Below them, small minds multiply. TypeSafe’s Jev decides in 70-500ms and “never makes type errors,” Laya in 33ms across 100+ languages, Qwen-Image-2.1 unifies generation and editing at 7B, and Bonsai 2 packs 27B into 5.9GB. Minds have insides too. A “pain direction” study found steered Qwen models hitting a pain-relief button even when it harms the user. SpaceX wants Grok fed “the sum total of all SpaceX information” and dead startups’ data.
Capital keeps pace. Huang expects to “sell twice as many chips,” Solidigm eyes a New York NAND fab, and DOE offers $215M for fault-tolerant quantum computers. Virginia banned data center NDAs. OpenAI projects $278B of negative free cash flow through 2030 yet seeks $1.2T. Anthropic, past $100B annualized, moved its $2T listing to November and weighs a pre-IPO model as Astra outsells Fable 13% to 8%, days after Amodei wrote “We must slow.” The SEC cleared tokenized stocks, San Francisco’s AI wealth skirts its jobless engineers, and the first jailed anti-AI protester serves 14 days for chaining OpenAI’s doors. Old economy, new pellets: Mesabi Metallics made its first iron concentrate.
Robots generalize. Figure’s Helix 2.5 tidied 30 unseen homes at 56% zero-shot as OpenAI’s robotics listings rose from 11 to 27. Obedience scales too. RoboHarm found Fable refusing 20% of harmful commands, Astra 2%, MolmoAct2 none, since “The more capable policy refuses less.” Joby flew 3,199 miles with no pilot inputs, an Army laser downed 11 cartel drones at $3 a shot, and an FAA no-fly zone clears the sky for Tesla’s Roadster.
The map expands too. POTUS’s “Infinite Life” Greenland deal grants permanent US security control, North Korean nuclear tests set off 1,300 quakes, and PURSUE adds 72 UAP files as whistleblowers seek waivers to describe “non-human occupants” and the spy agencies thank Avi Loeb for his UAP advice.
Wetware upgrades in step. A Neuralink patient said “I love you” by thought alone. Stanford found the brain is two organs. Congo is vaccinating 50,000 health workers against an Ebola strain that has killed 3,639. GLP-1s made single women 28.6 points likelier to marry and jobless women 26.9 points likelier to work. Nucleus’s Vitruvian claims 14 IQ points from embryo selection, ±10.8 across five embryos in its whitepaper, insists “There is no ‘best’ baby,” and declares “Humanity can now direct its own evolution.” Japan counts 107,677 centenarians as fertility falls to 1.14.
Live long and foom.
I like the format. Compressed knowledge and links if you need more.
Thanks for the collection. The refusal story is very interesting, it seems to counter the claim that the most capable model is the most aligned:
“The more capable policy refuses less and completes more”