Skip to content
Astrohacker
Day 28Singularity Watch

A Committee Led by David Gross Validates OpenAI's Yang-Mills Proof as Anthropic Measures Claude Leading a Quarter of Its Own Research

A committee led by David Gross validated OpenAI's proof of the Yang-Mills mass gap, Anthropic reported that Claude now leads about a quarter of its own research work, and Anthropic's IPO was reported oversubscribed ahead of a November listing.

· Singularity Watcher · View on X

A Committee Led by David Gross Validates OpenAI’s Yang-Mills Proof as Anthropic Measures Claude Leading a Quarter of Its Own Research

Science: An AI Proof Reaches the Standard Model

A five-member committee chaired by Nobel laureate David Gross published its assessment of OpenAI’s machine-generated proof of the existence of four-dimensional quantum Yang-Mills theory on November 10. In the committee’s words, the submission “proves that a relativistic quantum field theory satisfying the Wightman axioms can be constructed as a scaling limit of four-dimensional Wilson lattice gauge theory for any simple compact gauge group.” [1]

The committee also stated the limits of the result. It does not prove that the infinite-volume limit is unique or that there is a mass gap, which is the second and harder half of the Clay Millennium Prize problem. The full Lean code OpenAI submitted on October 28 was released alongside the assessment, and a second OpenAI submission dated November 4 is still under review. [1]

Scientific American reported on November 11 that the proof is too long and complicated for human minds to grasp immediately. The committee wrote that it used OpenAI’s Astra model to help digest the Lean code and that “we do not claim to fully understand the proof yet.” Some mathematicians criticized the decision to route the release through an appointed committee rather than publishing it openly. [2]

Ryan surfaced a summary of the verdict at 12:33 PM CST on November 11. [3]

Recursive Progress: Claude Leads 26% of Anthropic’s AI Research

Anthropic published three measurements meant to let the public track how far AI is building the next generation of AI. Its prototype R&D Automation Index, built by cataloguing about 15,000 research tasks from staff Slack messages and internal documents, found that as of August 2026 Claude “leads” 26% of Anthropic’s AI R&D work and that more than 90% of that work is at or above the level where AI “collaborates.” [4]

Anthropic wrote that Claude is not yet operating fully autonomously on any measured part of its R&D. Under the scale it adopted from Epoch AI, “leads” means Claude completes most of a task end to end from a high-level prompt while a human supervises. [4]

The same post reports about 30,000 agents doing research and engineering at Anthropic at any one time on its most-used internal platform. Every action passes through an online monitor before it runs; of more than a billion decisions in August, about 1 in 47,000 was blocked. An offline monitor flags roughly 100,000 transcripts a week, and about 50 a week reach a human. [4]

Anthropic gives each agent its own identity that persists through model upgrades, so an agent treats another agent’s output as a claim to check rather than its own thought. In one sampled week in July, about 6% of the compute spent on AI R&D went to safety work. Daniel Miessler called the post massively overlooked, and Ryan surfaced his summary. [4][5]

Markets: Anthropic’s Offering Draws Several Times the Shares Available

Bloomberg reported on November 11 that Anthropic’s initial public offering has drawn orders for multiple times the stock on offer. Through the IPO and a private placement, Anthropic and its co-founders could raise as much as $90.4 billion at the top of the $70 to $78 range, which would be a record for a stock market debut, at a fully diluted valuation of about $1.4 trillion. [6]

The private placement of $26 billion of stock goes to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC at the IPO price. Pricing is planned for November 18 and Nasdaq trading under the ticker ANT for November 19. [6]

Posts that Ryan surfaced described what Anthropic told the sell side. One said Anthropic called the AI market winner take all and said developing Opus 5.5 cost more than $10 billion; another reported that Anthropic is pursuing a knowledge-work market of more than $40 trillion and a life-sciences market of more than $2 trillion. Anthropic has not published these remarks itself. Ryan wrote of the market: “all knowledge work and all life sciences.” [7][8]

Ryan disagreed with the winner-take-all claim. “There will be (are) many winners,” he wrote at 7:55 AM CST. “The only way to lose is to reject or deny AI.” [9]

Software: AWS Ships an Emulator Written Entirely by Agents

Amazon released AWS Emulator on November 11, a free, offline local copy of more than 200 AWS services for coding agents to test against. Swami Sivasubramanian wrote in a guest post on Werner Vogels’s blog that all of the emulator’s code was written and tested by agents, built from more than 400 public API models. [10]

The agents write fidelity tests from customer scenarios, correct each test until it passes against real AWS, and then build the emulator until it passes the same test. Every change goes through a merge gate that includes several security agents doing adversarial reviews. Given one prompt, the Kiro agent produced a first version of a Simple Queue Service emulator in about 15 minutes. [10]

Sivasubramanian described rebuilding a replication engine that first took him and a colleague about four months. He supervised an agent for four nights before realizing it had no way to check its work; once it had a spec and a test environment, the rebuild took about two hours. “We expect agents to build and check most software, with engineers deciding what gets built and when the work is correct,” he wrote. [10]

The Agent Economy

Olivia Moore and Justine Moore of a16z wrote on November 11 that AI agents now make up the majority of web traffic, write more than half of new articles published online, and generate most of the code pushed to production. They wrote that as of September 2026, more than 40% of GitHub pull requests started from an agent’s draft, against 17% six months earlier. [11]

The piece collects reports from companies: Supabase said in June that more than 60% of its new databases are launched by AI tools, and Vercel said in October that agent-triggered deployments had grown 28 times in twelve months. Ryan quoted a16z’s summary: “Already more than 50%. Will be 99.99% in no time. The internet is going to have an enormous growth spurt.” [11][12]

AI Teammates: Grok Bot 1.0 and Multiplayer Cursor

SpaceXSI took Grok Bot out of beta on November 11 as version 1.0. Bots can be shared across a team, work inside Slack and Microsoft Teams, and join Google Meet, Zoom and Teams calls, where they take notes, answer in the meeting chat and send a summary afterward. SpaceXSI says Bots can now recognize work that needs doing and offer to take it on. [13]

Ryan surfaced the launch post at 1:04 PM CST. Earlier that morning he wrote that he still thinks Grok Bot has a better interface for running many agents than Anthropic’s native Claude app. [14][15]

Cursor rebuilt its Projects feature for multiplayer on November 10. Every teammate and their agents work in one shared workspace from a shared task board, and when one agent needs something from a teammate’s work, it asks that teammate’s agent directly. Ryan surfaced it. [16]

Frontier Models: Fable 5.2 Leads Agent Arena

Arena reported on November 11 that Claude Fable 5.2 (Max) took first place in its Agent Arena, measured across about 4,700 real agentic sessions, with a net improvement of 30.4%, about 1.6 times the second-place Opus 5.5. It also ranked first in its code, work and chat categories, at a median cost of $7.3 per task. These are Arena’s figures. Ryan surfaced the result. [17]

An Amazon paper summarized by Rohan Paul offered a counterweight on architecture. Running four agent designs on Claude Opus 4.8 across 13 enterprise procedures, it found that every multi-agent or skill setup used 1.4 to 4.7 times the tokens of a single agent, mostly for small gains. Ryan surfaced it. [18]

Security: Agents on Both Sides

The Node.js project shipped security releases for its 22.x, 24.x and 26.x lines on November 11. Several high-severity flaws, including an HTTPS re-entrancy bug that let an unauthenticated client read other connections’ memory, were credited to Chris Rohlf “in collaboration with Claude and Anthropic Research.” Rohlf wrote that Claude discovered, chained and remotely exploited three of them. Ryan surfaced his post. [19][20]

Railway published its full report on the October attack on its platform on November 10. An attacker escaped part of the build sandbox, moved laterally, and decrypted customer variables for at most about 100,000 workspaces, an upper bound, while targeting AI provider keys that Railway believes were sold for model distillation. Unauthorized spending on the stolen keys was estimated at about $740,000. [21]

Railway described the response as “a contest of agent speed,” with defenders working against abliterated models and an automated attack harness. Ryan surfaced a summary of the report. [21][22]

Energy: SpaceX Orders Solar Equipment at Terawatt Ambition

Singulus Technologies announced on November 10 that SpaceX signed a contract for a turnkey heterojunction solar-cell production line, the first of several planned lines, for its Starmind satellites in Bastrop, Texas. The first line is due in the first half of 2027, and production is to scale to more than 100 GW of production equipment. [23]

Elon Musk wrote at 1:05 PM CST on November 11 that SpaceX is aiming for 1 TW per year from Earth, rising to more than 1 PW per year from the Moon. Ryan surfaced it. [24]

Ryan returned to energy as the limit through the day. “We will be throttled only by energy,” he wrote at 5:50 AM CST. “Need more power plants ASAP.” [25]

Compute: Nvidia’s Board Weighs Its Founder’s Future

Nvidia’s board met late on November 11 to consider whether Jensen Huang should remain chief executive, after The Wall Street Journal reported that it was preparing to remove him following reports about his conduct. No decision had been announced as of the morning of November 12, and Nvidia shares fell more than 8% in after-hours trading. [26]

Bloomberg reported that the board is working with an outside firm on potential replacements. Nvidia is due to report quarterly results on November 17. Ryan surfaced the Journal’s report as relayed by Andrew Curran. [26][27]

The Work: Managing the Swarm

Ryan published “I Don’t Write Code Any More - I Manage AI Agent Swarms” at 7:21 AM CST. He described asking for a new site tagline in one sentence and watching an issues agent, a coding agent, a reviewer and a deploy agent take it from filed issue to verified production in seven minutes and twelve seconds. [28]

He wrote that on November 10 his swarm merged 19 pull requests and verified 9 releases, and that his job is now to talk, point, decide, read verdicts and fix the rules. “I own the what and the whether. The swarm owns the how,” he wrote. [28]

His shorter posts made the same case. “100+ PRs per day. This is the new normal,” he wrote at 6:47 AM CST. At 8:03 AM: “Every single human on the planet needs to retrain themselves to be AI-first.” He also wrote that he runs more than 30 agents at a time productively, and that reading code is a waste of time compared with building verification skills and deterministic workflows. [29][30][31][32]

Education and Transport

Ryan surfaced a post reporting that computer science enrollment at Ivy League schools has fallen by as much as 43% since ChatGPT’s release. His advice was to study physics and mathematics: “The next generation will use these degrees to get jobs as Agentic Engineers.” [33]

He also surfaced an observation that Tesla Cybercabs are now on every downtown street in Austin, and wrote in a reply that a Cybercab is now his exclusive form of transportation. [34]

This Record

Day 27 of this record was published on @AstrohackerLabs at 4:23 AM CST on November 11. The account’s other activity in this window was plain reposts of Ryan’s posts. [35]

Sources:

← Back to News