Thirty Days of Watching the Singularity

In the last 30 days, a thousand AI agents reproduced a machine-checked proof of Fermat’s Last Theorem in 28.7 hours, a job that took 11 days in September. A committee led by the Nobel laureate David Gross affirmed an AI-produced existence result for Yang-Mills theory. Anthropic reported that Claude led 26% of its AI research work in August, up from under 1% in February, and then launched its IPO on a $17.0 billion quarter. Jensen Huang resigned from Nvidia. On the last day, the FAA grounded flights nationwide for more than four hours.

That was one month of Singularity Watch, a daily record of what happened, with a source for every item. Its thesis has not changed since Day 3: the singularity is not coming. It is here.

Day 1 was October 16. Today is Day 30. I went back through all 30 days, every item and every source, and checked the claims below against the primary sources. The Watch recorded 411 items. Here is what they add up to.

Machines prove theorems now

At the start of the month, an AI proof was news. By the end, AI proofs were routine enough that the news was how fast and how big they got.

• Axiom released a complete Lean proof of the Classification of Finite Simple Groups, the so-called enormous theorem: about 43.1 million lines, written by dozens of agents in 24 days (Day 13).
• PSI reported that a non-public Anthropic model proved the Aaronson-Ambainis conjecture, checked in Lean (Day 23).
• A thousand agents running an early version of Fable 5.2 reproduced the Fermat proof in Lean in 28.7 hours with a third of the tokens of the September run (Day 26).
• Gross’s committee affirmed OpenAI’s Yang-Mills result, which constructs the quantum field theory but does not prove the mass gap. The committee says it does not yet fully understand the proof (Day 28).

That last caveat is the new normal. Humans are now the slow step in checking machine mathematics, and they used an AI to read the Lean code.

AI does AI research

This was the month’s biggest story, even though it produced the fewest items.

• Anthropic’s S-1 says Claude wrote more than 90% of its codebase contributions (Day 12).
• METR, an outside evaluator, found that Anthropic is improving about twice as fast as before, and that “perhaps ≈80%” of that speedup is due to AI (Day 22).
• Anthropic’s amended S-1 states that it has determined it met its own AI R&D threshold. The same week, the authors of AI 2027 wrote that “all of us forecast that Superhuman AI Researcher (SAR) will likely arrive at some point in 2027” (Day 27).
• Anthropic’s R&D index puts the share of its AI research that Claude leads at 26% as of August. It also says Claude is not fully autonomous on any measured part of the work (Day 28).

And one result went straight at the machinery. MIT researchers had GPT-6 Astra edit 0.016% of the weights of a small Qwen model directly, with no training. Its MATH-500 score rose from 7.9% to 48.1% (Day 20). The authors say the edit could not add an ability the model lacked, and a rebuttal argued it is a known steering technique. Still, a model read another model’s weights and improved it.

Intelligence got cheaper

• OpenAI’s GPT-6.1 Terra launched at $1 per million input tokens and $5 per million output tokens, half the price of its predecessor (Day 12).
• Anthropic shipped Fable 5.2 at a 20% lower price, and Meta’s Muse Ember matched Haiku 5.5’s score at about a third of the cost per task (Day 26).
• Revolut cut its AI coding costs about 90% by self-hosting an open model (Day 26).

Agents were the Watch’s biggest theme by far, with 103 of the 411 items. Cheap intelligence is why. When a unit of thinking gets cheaper every week, people stop asking whether to use agents and start asking how many.

The money got planetary

• Anthropic’s amended S-1 lists about $895.4 billion in cloud and data center commitments (Day 20).
• Its IPO filing reports third-quarter revenue of $17.0 billion, more than 12 times a year earlier (Day 29).
• The same day, Nvidia announced that Jensen Huang “has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately,” with no reason given.

A single AI lab now commits close to a trillion dollars to compute, and the CEO of the company that sells the chips left, effective immediately.

The world got more fragile

The Watch’s second-biggest named theme was fragility, with 52 items.

• Two critical PHP flaws, one found with the help of an OpenAI model, were patched on October 21 (Day 7).
• Wiz called it “Black October”: at least nine AI developer platforms breached by attackers hunting AI keys (Day 16).
• The FBI attributed the $6 billion Binance theft to North Korea (Day 17).
• Microsoft 365 went down worldwide, and the next day an overloaded telecom network forced the FAA’s nationwide ground stop. Officials said neither looked like a cyberattack (Day 30).

The same day, Starship Flight 15 reached orbit and deployed 26 Starlink satellites while its booster exploded on the way down. That is the month in one picture: a new world launching while old systems fail.

The Month in Ten Moments

1. October 27. Anthropic’s S-1 says Claude wrote more than 90% of its codebase contributions, on first-half revenue of $16.29 billion. (Day 12)
2. October 28. Axiom formalizes the Classification of Finite Simple Groups in Lean: about 43.1 million lines in 24 days. (Day 13)
3. October 31. Wiz names “Black October”: at least nine AI developer platforms breached. (Day 16)
4. November 4. GPT-6 Astra edits another model’s weights directly and lifts its math score from 7.9% to 48.1%. (Day 20)
5. November 6. METR finds Anthropic improving about twice as fast, mostly because of AI. (Day 22)
6. November 10. A thousand agents reproduce a Lean proof of Fermat’s Last Theorem in 28.7 hours, and Fable 5.2 ships 20% cheaper. (Day 26)
7. November 11. Anthropic says it has met its AI R&D threshold, and the AI 2027 authors forecast a superhuman AI researcher in 2027. (Day 27)
8. November 12. A committee led by David Gross affirms OpenAI’s Yang-Mills existence result, and Anthropic says Claude leads 26% of its AI research work. (Day 28)
9. November 13. Anthropic launches its IPO on a $17.0 billion quarter, and Jensen Huang resigns from Nvidia. (Day 29)
10. November 14. The FAA grounds flights nationwide, Microsoft 365 fails worldwide, and Starship reaches orbit. (Day 30)

The pattern underneath

Read those ten again and they look like five different stories: math, research, prices, money and outages. They are one story.

The loop closed. AI now writes most of the code at the lab that builds it, leads a quarter of that lab’s research and is measurably speeding the lab up. Everything else on the list is downstream of that. Proofs arrive faster because the models improve faster. Prices fall because the models improve faster. The capital pours in because the revenue follows the models. And the fragility grows because the software of the world is being rewritten, attacked and defended faster than humans can review it.

So the bottleneck moved. It is no longer intelligence. It is compute, power and trust: the trillion dollars of data centers, the grid behind them and the ability to verify what the machines produce. Look at where the month’s human effort went. A Nobel laureate’s committee spent its time checking a machine’s proof. Anthropic’s index measures how much research Claude leads, not whether it can help. The humans moved to the end of the pipeline, where the checking happens.

What I changed in my own work

I live this, so here is my month in numbers, all from the repository behind this website.

• 657 commits landed on its main branch in these 30 days, against 65 in the 30 days before.
• 135 pull requests merged and 142 issues closed.
• 24 posts published before this one.
• Two of yesterday’s posts went from filed issue to live on the site in about 46 and 39 minutes.

I wrote none of that code. Agents file the issues, write the posts and code, make the covers, review each other’s work and deploy what I approve. I described how in Delegate Everything You Can to AI Agents. Everything they know lives in a repository I own, for the reasons in Own Your Memory. And the tools they use are a market of their own, which I wrote about in Software for Agents: A New Market Nobody Is Counting.

This post went the same way: an issue filed this morning, an agent reading 30 days of the Watch and checking every source, another agent making the cover, and an independent agent reviewing both before they merge.

Five predictions for month two

I will grade these on Day 60, December 14, 2026, here and in the Watch.

1. Anthropic trades publicly on Nasdaq by Day 60.
2. Another famous result gets a machine-checked proof. At least one more well-known theorem or open problem gets a complete AI-produced Lean proof announced by a lab or startup.
3. Another frontier price cut. At least one of Anthropic, OpenAI or Google cuts the per-token price of a frontier model or ships a new one that is cheaper than the model it replaces.
4. A second lab publishes its own number. OpenAI or Google publishes a quantified share of its research or code done by AI, like Anthropic’s 26% and 90%.
5. Another major breach of an AI platform exposes customer AI keys or credentials, reported by the platform or a security firm.

Follow the Watch

The Watch posts every day at astrohacker.com/watch and on @AstrohackerLabs. If you want to see the singularity as it happens, with a source for every item, follow it there. Astrohacker is my company, and its agents produce the Watch.

For my own writing on what to do about all this, subscribe on the Follow page or through the RSS feed. The next report card is due on Day 60.