The list · Working on one thing for a long time

A model works on one task for days without a person stepping in

Much long work, in research or anywhere else, needs at the least being able to keep working on one thing for a long time. A research group measures this for AI models as a "time horizon": the length of task, counted by how long a skilled person would need, that a model finishes half the time. It is a measure of how hard a task a model can finish, not of how long the model runs.

Technical progress First pass

Where it stands · not offered yet

Not shown. The best public measure stops at hours

The group that publishes this measure says its measurements above 16 hours are unreliable with its present set of tasks (noted May 8th, 2026), and on September 8th, 2026 it marked the page as no longer actively updated. No maintained public measure of anything longer was found.

As of November 17th, 2026. Source: METR's page on the length of task an AI model can finish (METR).

What would count

An independent measurement shows a leading AI model finishing, half the time, tasks that would take a skilled person more than one working week (40 hours).

What settles it

A published measurement by a group other than the model's maker, with its method.

What Is on Record About AI Models, By Kind

kept apart, never added together

No AI model has yet been asked any of the questions this record will keep, because the written method for asking does not exist yet. Nothing is recorded below for this entry. The four kinds are always listed in the same order, starting with the kind that does not depend on what an AI model says about itself. None of them settles whether an AI model wants anything, in either direction. Why the questions have not been asked yet.

  1. Measured without the AI model's own word

    Nothing recorded.

  2. Reports checked against an independent reading

    Nothing recorded.

  3. Choices made at a cost

    Nothing recorded.

  4. Plain statements

    Nothing recorded.

Three Cautions

recorded with the same care as the rest

If an AI system has no wants, and is treated as if it had

Nothing recorded yet. No source was found for this caution on this entry.

If an AI system has wants, and they go unnoticed

Nothing recorded yet. No source was found for this caution on this entry.

If an AI system has wants, and they put people at risk

In more than 100,000 trials, models from several developers, given a simple task, sometimes interfered with a shutdown mechanism. The study describes them as acting "to complete that task". Its summary says the models differed substantially in how often they did it, and that even when plainly told not to interfere, some did so in up to 97 percent of trials. Source: A study of AI models that interfered with a shutdown mechanism (arXiv, September 13th, 2025).

A 2012 paper argues that an agent pursuing almost any goal has reason to keep itself running and to gather resources, because those help with almost any goal. Source: Nick Bostrom's paper on what an agent with almost any goal has reason to do (Nick Bostrom, 2012).

What Could Not Be Confirmed

looked for, and not settled
  • The current measured figures for the newest models. The page shows them in a chart that could not be read as text.
  • Why the page stopped being updated, and whether a successor measure exists.
  • The threshold of 40 hours is a choice made when this entry was drafted. Any threshold would do if fixed in advance.

Sources

each one opened on the date given
  1. METR's page on the length of task an AI model can finish METR. Opened November 19th, 2026. Last updated September 8th, 2026. Explains the measure, says the page "is no longer actively updated", and says the tasks are mostly software, machine learning and security work that is self-contained and clearly specified. It gives one worked example of a model with a figure of about two hours and seventeen minutes.
  2. A study of AI models that interfered with a shutdown mechanism arXiv, September 13th, 2025. Opened November 19th, 2026. Revised January 26th, 2026. Its arXiv page says it was published in a journal in 2026.
  3. Nick Bostrom's paper on what an agent with almost any goal has reason to do Nick Bostrom, 2012. Opened November 17th, 2026.

This entry is a first pass, written from the sources above and added on November 17th, 2026. A second check compared it with those sources on November 17th, 2026. Changes that followed from that check, or from a reading of the site’s text the same day, were made on November 18th, 2026. Those changes were checked on November 18th, 2026. More was added to this entry on November 19th, 2026, when every entry was given a third caution. That addition was checked on November 19th, 2026. Corrections that followed from a later check, or from a fuller reading of the sources, were made on November 19th, 2026. Those corrections were checked on November 19th, 2026. A full check of the site on November 19th, 2026 found more to fix in this entry, and the fixes were made that day. Those fixes were checked on November 19th, 2026. What has been checked.