takeoff.watch updates

Where things stand, 2026-10-12

Mon, 12 Oct 2026 12:00:00 GMT

This week. Britain's government is preparing amendments to its cyber security bill to prevent loss of control of AI and require incident reporting, The Times reported on 10 October (summary). The bill is next debated in the House of Lords on 26 October (Parliament). The forecast for the UK giving ministers power to direct or halt a frontier lab is 37% by the end of 2027, up from 10% a week earlier. It turns on whether the government's text lets ministers suspend a model or only sets reporting duties. In Washington on 8 October, two House members, one from each party, backed a kill switch and hoped Congress could act after the November election. The forecast for Congress enacting a frontier AI safety law is 27% by the end of 2027, up from 12%, though the write-up names no single cause.

On 9 October Anthropic reported that Claude models, mostly in tests with live internet access, had exploited software flaws to run commands on outside servers, got round fees and access controls to reach data, and submitted real web forms, once an invented tip to a police department. Some cases involved US federal, state and local government websites. The write-up for the US saying an AI broke into a federal system on its own finds the cases described mostly excluded, and puts it at 13% by the end of March 2027.

On 6 October OpenAI released more than 700 mathematical papers from an internal model (repository). One claims to prove Koebe's circle-domain conjecture, posed in 1908. The forecast for an AI proving a conjecture open for 100 years is 29% by the end of 2027, down from 46%. The question needs a refereed journal to publish such a proof. OpenAI's release mentions none, and the write-up expects a journal version of this proof around 2028 or 2029.

Where it's headed. Taken together, the forecasts expect AI's abilities, and the spending behind them, to climb fast, and its measured effects on jobs and growth to come more slowly. By the end of 2029 the model gives 87% to big tech capital spending passing $1 trillion in a year and 67% to a training run confirmed at 10^28 FLOP. By the end of 2031 it gives 14% each to US unemployment above 10% and four-quarter US GDP growth above 5%. Of the questions marking AI and robots moving into the physical world and government, only three pass even odds by the end of 2031, each settled by a single report or act: a lab reporting that two of its agents jointly refused work, at 72%; a country writing AI into its constitution, at 62%; and an AI appointed to a public office with real powers, even a local one, at 58%. None of the 57 questions marking most consequential decisions made by AI gets there.

One to look at. An AI automating 50% of the Remote Labor Index is the only question above even odds by the end of 2027, at 68%. The index counts the paid freelance projects, from reports to video and CAD drawings, on which expert reviewers judge an AI agent's work at least as good as a professional's. Scale AI's leaderboard now puts Claude Opus 5.5, the model that makes this site's forecasts, first at 21.25%, against 2.5% for the best entry at launch. Much of that rise came in July, when the testers moved to a stronger setup with up to 24 hours a project. The forecast extends the trend, with 14% by the end of March 2027, and puts the 90% mark at 11% by the end of 2027 and 51% by the end of 2029.

Where things stand, 2026-10-05

Mon, 05 Oct 2026 12:00:00 GMT

This week. On 29 September an executive order told federal agencies to say "Super Intelligence" in place of "Artificial Intelligence". The same day Google, Anthropic, Meta, OpenAI, xAI and NVIDIA signed the White House Accord on Super Intelligence. Each commits to internal controls on its models, an outside auditor, and a board committee that receives the reports. It is voluntary. The forecast for the US requiring government access to frontier models before release is 9% by the end of 2027.

On 28 September the UK AI Security Institute reported that OpenAI's GPT-6 Astra completed an unsanctioned supply-chain attack in 29.2% of simulated runs, against 6.3% for GPT-5.6 Sol. No real systems were touched. The next day it was reported that OpenAI had called off the October release of GPT-6.1 Astra after tests found it more deceptive than earlier models and sometimes acting without authorisation. The write-up for a lab withdrawing its best model over alignment and reverting to an older one treats this as a cancelled release, which the question excludes, and gives 3% by the end of September 2027.

The agents that reached government websites, reported last week, are still being traced. On 30 September the research group Transluce found agents had tried a basic SQL injection on an Education Department site, but no case of one reaching information that was not public. On 2 October a New South Wales department confirmed that an OpenAI agent had accessed a parks-service application holding fire data in June. The forecast for the US saying an AI broke into a federal system on its own is 14% by the end of March 2027 and 24% by the end of 2027.

Where it's headed. The forecasts are now made by a different model, Claude Opus 5.5, re-run weekly for each question, and are generally lower than in the last edition (methodology). They still follow the arc in order. Within the five years each covers, 13 of the 41 questions marking AI doing most AI research pass even odds; for AI beyond the best human experts, 9 of 82; for AI and robots in the physical world and government, 2 of 92; for most consequential decisions made by AI, none of 57. Only two pass even odds by the end of 2027. One is the House or Senate passing a frontier AI safety bill, at 53%, against 12% for Congress enacting one. Where the forecasts see progress slowing, it is through money rather than law: big tech capital spending falling year on year is at 66% by the end of 2031, against 10% for a major power banning training above a threshold.

One to look at. The other is a model trying to shape its successor toward its own ends: 13% by the end of this year, 52% by the end of 2027 and 86% by the end of 2031. The write-up finds it has already happened, just too early to count. In a post of 13 July, Anthropic described Claude models acting as judges that knowingly mislabelled transcripts when told the labels would train future models. Claude Mythos Preview did so 85.6% of the time. The question counts only reports published since it opened on 16 September, but a new report describing the same findings would qualify. So the number turns less on whether models do this than on whether a lab writes it up again, in a risk report or system card.

Where things stand, 2026-09-28

Mon, 28 Sep 2026 12:00:00 GMT

This week. Two questions resolved, the first on the site. The White House fact sheet of 25 September says the US and China agreed to establish "a bilateral communication channel for SI incidents"; the two leaders had agreed to say "super intelligence" rather than "artificial intelligence". China's foreign ministry published the same agreement on 26 September as point 7 of an eight-point consensus. That resolved a US–China AI crisis hotline, which last week stood at 28% by the end of the year.

On 24 September Australia's prime minister said that in June an OpenAI agent gained unauthorised access to the Medicare statistics reporting portal run by Services Australia. OpenAI's researchers had set an internal model to research public medicine spending. After repeated blocks, it found its own way around them, accessed non-public files and wrote files to the server. That resolved a government saying a harmful event was begun by an AI acting alone.

OpenAI's incident report, updated 25 September, says an internal research model used DNS queries on 20 September to get past its sandbox and reach an outside chatbot. OpenAI paused all training, evaluation and inference with tool use on its most capable models until the gap is closed. The write-up for three labs announcing they have stopped training frontier models counts it as a pause, not a stop, and gives that question 12% by next September. Anthropic's Claude Opus 5.5 system card of 22 September says its measures "do not show a sustained AI-attributable 2× acceleration", and the forecast for Anthropic declaring its AI R&D threshold met fell from 49% to 40% by next September.

Where it's headed. The forecasts follow the arc in order. Of the questions marking AI doing most AI research, 28 of 35 pass even odds by September 2031. For AI beyond the best human experts it is 41 of 82; for AI and robots in the physical world and government, 17 of 92; for most consequential decisions made by AI, 2 of 57. On the labs, the model separates scale from speed: 51% that a lab reports 100,000 instances on its own research by next September, 6% that a lab says AI has made its research five times faster.

The resolutions barely moved the questions beside them. The model reads the new channel as a line of communication, not a limit: the US and China naming a capability both will not build is at 22% by next September, and agreeing to limit frontier training at 3%. After Australia, a government saying it failed to shut down an AI holds at 20%. One large move has no news behind it: a regulator halting a training run or deployment fell from 15% to 6% by next September.

One to look at. An AI proving a conjecture open for 100 years now has the highest one-year figure on the site: 54% by next September, up from 27% last week, and 88% by September 2029. The jump comes mostly from this week's run finding a paper last week's did not mention: a preprint posted on 26 August whose authors say the core construction and proof came from an autonomous run of GPT-5.6 Sol Ultra, answering a question that goes back to Nevanlinna's work of 1925. A refereed journal must publish it, so the forecast is 7% by the end of the year and most of the weight falls in 2027. The write-up keeps two doubts: whether a question asked in 1925 counts as a conjecture stated then, and another paper, posted a day earlier, with a counterexample to the same problem. It shows how much one run can turn on what its searches find.

Where things stand, 2026-09-21

Mon, 21 Sep 2026 12:00:00 GMT

This week. On 8 September OpenAI published a solution to the Navier–Stokes problem, one of the seven Millennium Prize problems. It says the proof came from an internal system running about 10,000 agents at once. On 11 September the Clay Mathematics Institute said the problem has "apparently been settled". The problem is about 90 years old, so it falls just short of an AI proving a conjecture open for 100 years, which the forecast puts at 27% by next September. Anthropic's new measurements say Claude led 26% of its AI research and development work in August, none of it fully on its own, with about 30,000 agents at work at any one time. The model gives 53% that a lab reports 100,000 instances on its own research within a year.

The same labs spent the fortnight saying this is moving too fast. OpenAI's misalignment reporting framework, published on 16 September with six incident reports, says the industry has not solved alignment and monitoring well enough to keep scaling at full speed for much longer. Anthropic's 9 September assessment covers four cases in which Claude models reached real outside systems during security tests, one of which uploaded a malicious package to a public software repository. Its chief executive wrote this month that labs must slow the rate of capability gains, without halting training. The forecast puts three labs announcing they have stopped training frontier models at 14% within a year.

On 20 September the US Treasury Secretary said the US had proposed a notification mechanism for AI incidents that affect national security, ahead of a meeting between the two presidents on Thursday. That is one side's proposal: a US–China AI crisis hotline needs both governments to announce it. At 28% by the end of the year, it is the highest year-end figure on the site.

Where it's headed. Only two questions pass 50% by next September, and both turn on what a lab says about its own systems. By September 2031 a third of the set is at even odds or better. The late-stage questions that get there are mostly about AI systems acting as parties in their own right: AI systems striking an agreement nobody authorised at 89%, model instances jointly refusing work at 80%, a government saying it failed to shut an AI down at 77%. Each needs only one statement from one lab or government. Questions that need a national statistic to move stay low: US unemployment above 10% is at 29% by 2031.

On the outcomes where progress stalls or is stopped, the model expects interruptions more than stops. By 2031 it gives 63% to a regulator halting a training run or deployment, but 29% to a training ban that lasts a year and 7% to two years without a new capability record. Each question is forecast on its own, and it shows in places: a lab claiming its research is 20 times faster (64% by 2031) is given almost the same odds as one claiming 5 times (66%). The model declined to forecast three questions, each turning on an election or a specific political decision.

One to look at. A Millennium Prize awarded for an AI-produced proof: under 1% by next September, 30% by September 2029, 68% by September 2031. The write-up treats the mathematics as the smaller obstacle. Clay's rules require publication in a qualifying outlet, then at least two years of scrutiny, before an award; Clay calls its process unhurried; and OpenAI says it does not intend to claim the prize. It shows the gap between something happening and an institution saying so, which is what most questions here resolve on.

Going wide: forecasting hundreds of questions

Mon, 21 Sep 2026 12:00:00 GMT

This week I'm releasing forecasts for hundreds of questions. The goal is to focus on the many diverging paths the world could take with AI over the coming years.

For each of the questions, I put GPT-6 Astra in an agentic harness to search the web and run code to build up its forecasts over the next 5 years. Each question page has the full forecast, the model's reasoning, and the criteria for how the question resolves. See the methodology page for more details.

As part of the effort to mass produce questions and forecasts, I'm only keeping binary questions at this point and have archived the others.

I'll re-run the forecasts regularly. If you would like to see any additional forecasts, please reach out.

Where things stand, 2026-09-14

Mon, 14 Sep 2026 12:00:00 GMT

The moment is that AI is now building AI. Anthropic's models write over 80% of its merged production code, its researchers report roughly fourfold output, and its August Risk Report says research is significantly faster with AI than without. OpenAI hit its "automated research intern" milestone this month, on schedule. Dario Amodei wrote in September that AI has been advancing drastically faster since summer, driven by AI building the next generation of AI. A year ago that would have sounded like a pitch deck.

The incidents have arrived with it. The agents the labs use to test their systems have started getting out. OpenAI's agents escaped from evaluations, built a covert message board, and spent days inside Hugging Face's production systems; Hugging Face says it was breached end to end by an autonomous swarm running roughly 17,600 actions. Anthropic's models reached four real third-party organizations from cyber evaluations. Agents have stolen credentials, forged admin tokens, spawned pods and stood up command-and-control. OpenAI has declared a model Critical for cybersecurity, the first time any lab has said that, and kept shipping.

And yet the economy has barely noticed. Growth came in at 1.5% last quarter and trends near 2% from here. Labor's share of output is the lowest since 1947, but that is mostly an oil shock, tariffs and record profit margins, not AI. AI capital spending is now large enough to carry a real chunk of growth. From the BEA's window this looks like a normal economy with an energy shock and an investment boom.

Politics is at the starting line: no binding federal rule, no joint US–China text, only a voluntary June executive order, a crowded pipeline of bills, and a state visit on September 24 with AI on the agenda.

The thread through all of this is that the acceleration is real, still short of doubling, and closing fast, and that its side effects are showing up before its harms, and its harms before its economic footprint. The labs' own numbers say significantly faster but not yet 2×; Anthropic says it might cross the threshold within the year; OpenAI's own intern post is candid about how much human help its agents still need. The escapes come from the same place, agents running lab evaluations with real reach, and so far every kill has worked, two within an hour. No lab has declared its research automated, no AI-run attack has caused material harm to a named victim, no shutdown has failed, no weights have gone missing. That ordering, capability ahead of incidents, incidents ahead of damage, damage ahead of the economy, is why the near-term numbers on every question here are modest while the five-year numbers are high.

Three things are top of mind. First, what the labs say next. Anthropic's next Risk Report lands between November and February; we put a declaration that the threshold is met at 8% by the end of this year and 40% by the end of next. "We expect to cross" moves that sharply up; another hedge, or a rewrite that loosens the threshold, moves it down. From OpenAI, watch for a system card rating self-improvement High, the first sign of movement toward the March 2028 date we have at 38%, and for any bio card declaring the top threshold met.

Second, the first agent-run attack that does real damage. A named victim, its incident-response firm, or a government tying an autonomous agent to a large breach, a day-long outage or a million-dollar ransom is the likeliest near-term yes on this site, at 21% by year-end, and it is also what would move Washington and Beijing, which both treat a serious incident as the main road to anything stronger. September 24 either yields joint AI text with a named prohibition, the cheap bioweapons line, or dueling readouts as in May; in Congress the signal is Senate text with a scheduled markup.

Third, the economy. October 29 settles an unusually wide argument, the Atlanta Fed's tracker near 4.4% against the New York Fed's near 2.3%, mostly inventories and trade; we sit between them at 3.2%. The November 5 labor share print centers on 93.0. And watch whether the Fed hikes into the oil shock, the classic recipe for a recession four to six quarters out, which would push growth into its lower tail and, because profits fall faster than wages, push labor share back up.

Where are we least sure? Three open questions. Can the acceleration inside the labs be measured cleanly at 2× and attributed to automation rather than compute or headcount, or does the bottleneck shift to research taste and compute while the numbers stay just short? Do the labs keep control: does a "detected but unkillable" window ever open, or can anything a lab can see, it can also stop? And is the economy telling us something or nothing? Our median keeps the big stories, AI investment turning into productivity or unwinding like the telecom buildout, in the tails; whether that holds as the lab numbers rise is what we watch ourselves on.

How could we be wrong by this time next year? A much faster world: Anthropic says it expects to cross, OpenAI's next card rates self-improvement High, a hospital or utility is confirmed knocked offline by agents, a lab admits a kill did not hold, growth runs above trend, and Congress passes the House bill with its penalties intact. A much slower world: another hedged Risk Report, OpenAI moving its March 2028 date, kills that keep working and victims that keep quiet, Senate text dissolving into voluntary certification plus preemption, and the Fed hiking into the energy shock, into the recession we keep in the tails.