OpenAI 用约一万个 Agent 在 88 小时内推进数学证明
OpenAI 展示一个跑在 Astra 之上的内部模型数字实验室:约一万个 Agent、27 亿条消息、1300 亿输出 token,88 小时拼出证明路径,Astra 再用约 17 小时在 Lean 中形式化。
openai agent labs make girl humanoids feel inevitable ngl once you watch ten thousand agents compress ~90 years of open math tension into 88 hours a soft girl on a dock stops looking like a toy she starts looking like the missing hands under the same factory → what openai just showed not one chatbot genius cracking navier-stokes in a long chat a digital lab ~10,000 agents on an internal model above gpt-6 astra parallel research 2.7 million messages across the swarm ~130 billion output tokens burned through the run proof path assembled in 88 hours of continuous agent work then gpt-6 astra spent ~17 hours formalizing it in lean millions in compute no millennium prize claim planned yet independent mathematicians still have to adjudicate but the shape of discovery flipped in public swarm finds astra formalizes humans still judge → why that makes girl humanoids feel next science just got an agent factory with a stop condition and a proof target planner / worker / critic loops at scale message graphs token volume parallel search lean verify the scarce unit is no longer a lonely paper draft it is the lab graph itself once that graph exists the next bottleneck is obvious agents can research order draft verify and click they still cannot walk the kitchen lift the bag set the product in frame or finish the last ugly centimeter girl humanoids become the effector layer under the same architecture cameras for eyes soft hands for grasp dock for overnight charge policy link so the swarm’s plan becomes meters not another notification → what that stack becomes at home same pattern as the digital lab only the target is life not pde one agent shops one writes the shot list one critic-checks brand safety one soft girl executes the physical beat you approve like a lab lead not a joystick babysitter ugc chores returns research-plus-fetch all of it reads like agentic science with a knit suit finite-time blowup was the math punch finite-time “done” is the house punch → where this leads openai just demoed a new model for scientific discovery every lab will copy the agent graph then every serious home stack will want a body on the other end of it teleop becomes exception handling the dock becomes producer credit the soft girl becomes rented staff for an agent factory not a mermaid reel and not a silent mannequin also the leash a swarm that can burn 130b tokens on fluids can also over-buy over-film and over-act in a living room frontier agents plus a home body need stop conditions as hard as the proof target → my take openai agent labs make girl humanoids feel inevitable because after decades collapse into days you do not ask for another chat window you ask for hands swarm finds astra formalizes the soft body finishes the room humans still adjudicate bookmark the graph before every brand copies it into a kitchen so back
这条是引用着另一条写的:x.com/rimtoln/status/2097692697939087549