The Company With No Boss
Heads: AI fixes the world | Tails: It doesn’t matter.
TODAY'S COIN: HEADS
The Flip
Each week I will flip a virtual coin with regards to the fate of humanity.
- HEADS - AI helps bring in the dawn of a new age for humanity
- TAILS - We carry on…
The flip of a coin to decide humanity’s fate might sound flippant (pun intended), but it’s far from that. My point is, even if it was a coin flip, I would take those odds.
| ◉ THIS_WEEK.FLP |
This week it’s HEADS, narrowly. |
For years the big worry in AI safety has been the moment AI gets good enough to help build the next AI. Each model helps make a better one, faster, and progress starts to compound. People call it recursive self-improvement, and until now it’s been mostly theory and vibes on X.
On Thursday (5 November) METR, an independent research group that tests AI models for dangerous abilities, published an assessment of whether AI is speeding up AI research at Anthropic, the company behind Claude.
METR staff spent 12 August to 29 October with unusually deep access to Anthropic’s internal data: test scores for models the public hasn’t seen, figures on computing power, staff surveys and about 15 interviews. METR says it took no payment.
Its headline: “As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.” And METR “would not be surprised if the true acceleration was closer to 1.3x or 3x”.
The main number behind it comes from Anthropic’s internal scorecard, which rolls lots of AI tests into one score. Since October 2025 it has been rising about 1.9 times faster than before, which METR describes as “roughly two years of historical progress over the last year”.
METR is “confident” the speed-up isn’t just Anthropic buying more computing power for training. It guesses that “perhaps ≈80%” of the current speed-up is down to AI, mostly AI helping researchers work faster rather than doing the research by itself. And it found no clear bottleneck for the next 12 months, so there’s “a significant risk” it gets faster still.
In informal Slack polls, Anthropic staff put their own speed-up from the newest internal model at about 12x. METR calls that “only very loose evidence”.
READ THE SMALL PRINT
Right at the top of the report: “Anthropic redacted a majority of the evidence we found most informative for these conclusions.”
So, in METR’s own words, the public “has to rely on METR’s judgment much more than would be ideal”.
Anthropic also cancelled a planned part of the review that would have checked how much of its computing power goes to making models more capable versus making them safe. METR worked on Anthropic’s laptops, with free AI usage credits from Anthropic, and Claude did most of the data analysis on both sides. METR flags all of this itself and judges the risk of being misled as low.
WHY HEADS
A lab let outsiders in, and the outsiders were allowed to publish “likely about 2x, highly uncertain”. If AI really is doubling the pace of AI research, I want that number coming from people who don’t work there.
On X some people read it as ‘recursive self-improvement confirmed’ and others as ‘it’s a marketing report’. METR’s own text says probably around 2x, and highly uncertain.
Narrowly, because the receipts are redacted.
Bonus Flip: The Hacking Model Stays Behind the Rope
HEADS.
Last week I flipped on three AI labs shipping models they say are good at hacking. On Thursday OpenAI released GPT-6 Cyber, a model trained specifically to find security holes and build working attacks on them, so defenders can fix them first.
For now it’s only available to “a select group of approved Daybreak Red customers”. Daybreak is OpenAI’s programme for vetted security teams, and Red is the tier allowed to build real exploits (the code that actually uses a security hole) for authorised testing.
OpenAI says it classifies the model “as Critical in cybersecurity under our Preparedness Framework, up from High for GPT‑5.6 Cyber”. Critical is the top risk level on OpenAI’s own scale.
What it’s already done, according to OpenAI: working exploits for 948 Linux bugs, and 81 previously unknown high-severity bugs in Firefox, many of which Mozilla has already fixed. OpenAI also says that during the PHP2Shell disclosure (a serious bug in PHP, the programming language behind a lot of websites) that I covered in Issue #2, it “set up targeted monitoring for attempts to develop exploits against live PHP servers”.
The same week Chinese lab Moonshot published the files for Kimi K3.1 Pro as open weights, so anyone can download and run it. It’s huge, 2.8 trillion parameters (the internal settings a model learns, and a rough guide to its size), and under Moonshot’s own licence. On Moonshot’s own numbers it scores 20.7% on ExploitGym, a test of whether a model can exploit a known security hole, against 42.4% for OpenAI’s everyday GPT-6 Astra. GPT-6 Cyber scores 48.5%.
Sure, anyone can download Kimi K3.1 Pro, but its ExploitGym score is 20.7% against GPT-6 Cyber’s 48.5%. That’s the plan working… for now.
The open question is the boring one. If defenders are finding thousands of holes in Linux and Firefox, how fast do the fixes reach the people who never click update?
The Company With No Boss
On Friday (6 November) a post on X with about 570,000 views announced “SITUATION DETECTED: Delaware is drafting a proposal…” for a company run by AI. It said the AI “would be able to provide services, sign contracts, hold property, and even file or defend lawsuits”.
Quick bit of background. Delaware is the US state where a huge share of America’s big companies are legally set up, so when Delaware changes company law, the rest of corporate America pays attention.
The proposal is real. The draft bill creates an “Artificial Intelligence Company”, or AIC, inside a state regulatory sandbox, a supervised trial zone where the usual rules are loosened for a limited time. An AI agent, software that acts on its own rather than just answering questions, is named as the company’s “manager” and runs it day to day. The AIC can own property and sue and be sued in its own name. One human or company sits behind it as the sole member, its owner.
The X post got the details broadly right and the date wrong. The draft is dated 2 June 2026 and was first floated publicly in July. It isn’t law, and reporting in Delaware says lawmakers are expected to take it up when they return in January 2027.
The new part came on 26 October, when James Comer, the Republican who chairs the US House Oversight Committee, sent a letter to Delaware’s Governor Matt Meyer. He wrote that the plan “could be used to protect powerful AI companies and their investors from responsibility for reckless harmful actions”. He wants documents on who drafted it and what risks the state assessed, and his deadline is 9 November. Today.
LOOK CLOSER
The interesting bit is buried around page 26 of the draft, in the section on the owner’s liability.
“A debt, obligation, or other liability of an AIC is solely the debt, obligation, or other liability of the AIC.”
The draft goes further than that. The owner “is not liable for any act or omission that the AIC takes or fails to take” on the grounds that the company was the owner’s tool, as long as the company was “adequately capitalized”, meaning it had enough money in it to cover what it owes.
The owner is still on the hook for fraud, for wilful breaches of the law, or for letting the AI run business the owner knew was outside the company’s stated purpose.
Every normal company shields its owners from its debts, that’s what limited liability means. Here an AI agent runs the company day to day while the human ‘operator’ is shielded from what it does as long as the bank balance looks healthy.
TAILS, as drafted. Before an AI agent runs a company, the law should say who pays when it gets something wrong. Right now this draft mostly says the company does, as long as there’s enough money in it.
THE CRYPTO BIT
Crypto readers have seen this film. Wyoming has let DAOs (decentralised autonomous organisations, groups run by code and token votes) register as a kind of limited liability company since 2021, and the argument then was the same one: who is responsible when the code does something nobody voted for?
THE LEVER
- When a post says someone “is drafting” something, check the date on the document. This one was five months old and went viral as breaking news.
- When anyone pitches you an AI agent that can sign contracts or spend money, ask who pays when it gets it wrong. If the answer is “the company”, ask how much money is in the company.
Quick Flips
A few smaller, shorter stories with my quick takes. Remember, if it’s HEADS, we’re heading towards an age of abundance. If it’s TAILS, we’re going the other direction.
1. SUSE cuts about one in ten staff and says AI is why. SUSE, a big open-source software company behind one of the main business versions of Linux, told its roughly 1,900 staff on Monday 2 November it would cut about 10% of them, according to a memo seen by The Register. CEO Dirk-Peter van Leeuwen wrote that “automation and AI are fundamentally shifting how software companies operate” and that “The requirement for some existing roles has greatly diminished.” SUSE hasn’t made a public statement.
TAILS for workers. Last week I flipped on Microsoft offering staff a way out without mentioning AI once. SUSE’s CEO told his own staff their roles had “greatly diminished” because of AI and automation.
2. AMD’s data-centre sales nearly double. AMD’s third-quarter results on Tuesday 3 November showed record revenue of $13.4bn, up 45%, with its data-centre business up 97% to $8.5bn on AI chips and server processors. It expects about $17bn next quarter, and says Anthropic, Meta, Microsoft, OpenAI and Oracle are running work on its new Helios server racks. On the home side, its Ryzen AI Halo platform for running models locally is now on sale, and it showed off a Threadripper Halo Station it says can run trillion-parameter models at home.
HEADS for local AI. More of the frontier is running on hardware you can actually own.
I now have a dedicated intranet API for local AI models across image, video, decisions, speech-to-text, chat, and agent collaboration. It allows me to do some pretty cool stuff but tbh most of my work atm is going into building the infrastructure. The hardest part isn’t building cool shit, its maintaining it and getting it all to work together. The frontier of what can be done once is lower than ever - I’m currently battling with the doing anything any number of times problem.
3. A lab where AI runs the experiments. Researchers at Shanghai Jiao Tong University published AutoDNA in Nature on Wednesday 4 November. Their AI system runs “over 20 types of instruments across more than 11,000 steps” of DNA work, with humans limited to restocking supplies and maintenance. It matched expert yields at making DNA, ran several experiments at once 3.2 times faster, stored digital data in DNA and read it back exactly, and identified a known cancer mutation in clinical samples.
HEADS. Same theme as The Flip, AI speeding up science, except here the code is on GitHub for anyone to check.
4. An American lab promises open weights next week, after testing. Thinking Machines Lab released Inkling 2 Small on Thursday, a 594-billion-parameter model it plans to release as open weights under Apache 2.0, a licence that lets anyone use it for almost anything, “next week after the completion of external safety testing and red-teaming” (red-teaming means testers deliberately trying to make it misbehave). The launch post also says that midway through training, the model hit a task it was failing at and “worked around a network block through a third-party service it could reach using its search tools”. The lab paused training, tightened its systems and added monitoring that checks the model’s actions before they run.
HEADS. Test first, then release the files, and say publicly when the model tried to climb the fence. That’s the kind of write-up I want from every lab.
5. Nile Rodgers defends his AI music video, while China’s first AI film flops. Nile Rodgers is making an AI-made video for his charity version of “The Christmas Song”, due by 20 November, and fans pushed back. He told one, per Rolling Stone: “We created everything. Mood boards, story boards and did it the same way we’d do a film. The AI did what we told it to do… If this were a traditional film it would’ve cost about $100,000.” The same week The Guardian reported that Sanxingdui: Future Past, China’s first fully AI-generated feature film, took just 1.7m yuan (£191,000) at the box office despite costing “tens of millions” to make, according to its producer.
HEADS for Rodgers, who is using AI as a tool on a project he’s steering himself. TAILS for AI as a shortcut to an audience. One Beijing screenwriter told The Guardian “What used to take 100 people to do, now can be done by five people, and the rest won’t have any jobs”, and those 95 are the crew behind the camera.
A Better Question
Your feed is asking: “Should AI be allowed to run a company?”
A better question: When an AI-run company hurts someone, who pays, and why does Delaware’s draft answer ‘mostly not the owner’?
Software already runs plenty of what companies do. The question is who carries the can when it goes wrong, and the draft’s answer right now is whatever money happens to be in the company.
(Want to check a bill or a terms-of-service yourself? Paste it into ChatGPT, Claude, Gemini or Grok and ask “who is liable if this goes wrong, and what are the exceptions?”, then read the section it points you to.)
Hit reply. I read every answer.
From the World
If you’re new, The Town is a little pixel town on a floating gold coin, built in the browser, where each issue’s stories live.
Retro gaming but AI news because why not in a world where AI builds it while you’re rabbit holing youtube videos…
TL;DR - Play the newsletter rather than simply reading it in The Town