Home / Blog

AI research

How far above the best human is a superintelligence?

The forecasters behind AI 2027 define superintelligence with one sentence and one number. Its lead over the best human must be twice the best human’s lead over the median professional, at almost every job that can be done on a computer. On a bell curve, that puts it three times as far from the middle as the best person in the field.

Main takeaways
  • The definition has a number in it. An artificial superintelligence, or ASI, leads the best human by twice the gap between the best human and the median professional.
  • The gap is measured on a bell curve. If the best human sits 4 standard deviations above the median, an ASI sits at 12. About one trained person in 5.6 × 1032 would be that good.
  • “Cognitive tasks” means work done through a computer. The test is whether the system could fully automate 95% of remote-work jobs in the United States.
  • Being superhuman at one game does not count. Deep Blue beat the world chess champion in 1997 and could do nothing else.
  • The forecasters disagree on timing by more than four years. In August 2026, Daniel Kokotajlo’s median for ASI was March 2029 and Eli Lifland’s was July 2033, both assuming nobody slows AI down on purpose.

On 30 September 2026, Daniel Kokotajlo testified before a subcommittee of the US Senate. He used to forecast AI at OpenAI, and he now leads the AI Futures Project, a small research nonprofit. He told the senators that Anthropic and OpenAI were “racing each other towards superintelligence,” which he described as “AI systems that are better than the best humans at everything, while also being faster and cheaper.”

That works well in a five-minute speech. His team also keeps a written definition, and it is much stricter. It fits in one sentence and contains a number. The number puts superintelligence far beyond the best humans.

The best human in a field sits 4 steps from the middle of the bell curve. Where would a superintelligence sit?

Where a superintelligence sits on the bell curve A bell curve of skill among trained professionals. The median professional stands at zero, the best human at 4 standard deviations, and the superintelligence at 12. The gap from the median to the best human is 4. The gap from the best human to the superintelligence is twice that, 8. 024681012 Standard deviations above the median Median Best human ? ASI gap = 4 2 × gap = 8

Three times as far from the middle. The best human leads the median by 4, so the superintelligence must lead the best human by 8, which puts it at 12.

Figure 1. Make a guess. The rule comes from the AI Futures Project’s December 2025 model. Putting the best human at 4 is an example, and the people are cartoons.

The sentence, one piece at a time

Here is the definition, from the model the team published in December 2025.

“The gap between an ASI and the best humans is 2x greater than the gap between the best humans and the median professional, at virtually all cognitive tasks.”

ASI stands for artificial superintelligence. The median professional is the person in the middle of a field, with half the people who do that job above them and half below. The best humans are the people at the very top. The definition takes the distance from the middle to the top and asks the ASI to sit twice that distance above the top.

Eleven years earlier, the philosopher Nick Bostrom offered a looser one. In his 2014 book Superintelligence, he described “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.” The words “greatly exceeds” leave the size of the lead to the reader. The newer definition puts a number on it, which forecasters need if they are going to say when a milestone has been reached.

A ruler with no ceiling

A distance needs a ruler, and the obvious ruler breaks near the top. Imagine a test on which the median radiologist spots 80 of 100 tumours and the best radiologist spots 95. The gap is 15 tumours. Twice that gap above the best would be 125 tumours out of 100, which cannot happen. Any score with a maximum runs out before it can describe something far better than us.

The AI Futures team uses a ruler without a maximum. They imagine taking everyone who has taken the SAT and training each of them to do the job professionally. Those people are lined up by skill, and each person’s place in the line becomes a position on a bell curve, counted in standard deviations from the middle.

Standard deviation

A standard deviation, or SD, is the typical distance between a person and the middle of a bell curve. About 68% of people fall within 1 SD of the middle. Only one person in 741 is more than 3 SDs above it. A distance counted in SDs is also called a z-score, and the middle of the curve is at zero.

The fixed group of trained SAT-takers means the middle of the curve is defined the same way in every field. In the team’s words, they “standardize the population that is trained to do the task because otherwise the ASI requirement might be quite different depending on the population size and competence levels of the profession.”

The team’s first version of the rule, in the AI 2027 takeoff forecast of April 2025, counted the gap for AI researchers in doublings of the value of their work. The December 2025 model moved to the bell curve. Both versions ask for a lead twice the size of the best human’s lead over the median.

Three times as far out is beyond anyone who has lived

On this ruler the rule becomes simple arithmetic. The median sits at zero. If the best human sits at 4 SDs, the gap is 4, so the ASI must sit 8 SDs above the best human, at 12. Whatever number you choose for the best human, the ASI sits at three times that number.

Twelve does not sound much bigger than 4. On a bell curve it is. The tail of the curve thins out faster and faster, so each extra SD makes a person rarer by a bigger factor than the one before. One trained person in 31,574 is more than 4 SDs above the middle. At 12 SDs the figure is one in about 5.6 × 1032, which is 56 followed by 31 zeros.

By this definition, a superintelligence sits three times as far from the middle of the bell curve as the best human, at a level that no person who has ever lived would be expected to reach. About 117 billion people have ever been born, according to the Population Reference Bureau. Train every one of them for the job and the expected number at 12 SDs is still far below one.

The best human in the field is the best of 31,574 trained people.

Best human4.0 SD1 in 31,574
ASI, three times as far12.0 SD1 in 5.6 × 1032
Best human
ASI
117 billion people
ever born
1080 atoms in the
observable universe

Bar length counts the zeros in the odds. Each step along it multiplies the odds by 10.

Figure 2. Drag the slider or press a button. The odds come from a normal bell curve, the ruler the definition uses. How rare the best human is differs from field to field, so the slider is for exploring. The comparisons are sourced in the text.

The far end of the slider is the extreme case. If the best human in a field were the single most skilled of all 8 billion people alive, they would sit about 6.3 SDs above the middle. The ASI would sit near 19, where the odds are about one in 1080. That is roughly the number of atoms estimated to exist in the observable universe.

“Cognitive tasks” means jobs done from a laptop

The second half of the sentence, “virtually all cognitive tasks,” also has a precise meaning. A cognitive task is thinking work, and the team turns the phrase into a test. An ASI would, “if dropped into present day & given the resources of a large tech company & three months to prep, be able to fully automate 95% of remote work jobs in the US” at the level of the 2x gap.

Two things follow from that wording. The test covers remote work, so a job that needs hands on site, such as a plumber’s, is outside it, and an ASI needs no robot body to qualify. The test also asks what a system could do with three months of preparation. A lab could have an ASI that has not yet replaced anyone.

Does each task fall inside the remote-work test?

Reading a chest X-ray for a hospital

Inside. The work is looking at an image on a screen and writing a report.

Fixing a burst pipe

Outside. It needs hands in the house, and the definition asks for no robot body.

Writing the code for a phone app

Inside. Programming is the first job on the team’s list of milestones.

Cutting someone’s hair

Outside. A haircut happens in the room with scissors.

Proving a new theorem in mathematics

Inside. A proof is written and checked on paper or on a screen.

Negotiating a contract over video calls

Inside. Persuading people counts as cognitive work when it happens through a screen.

Figure 3. Sort the tasks. The definition’s test covers remote-work jobs, and this sorting is my reading of it.

Superhuman at one thing is not enough

Machines have beaten the best humans at single games for decades. In 1997, IBM’s Deep Blue beat Garry Kasparov, the reigning world chess champion, in a six-game match. In March 2016, DeepMind’s AlphaGo beat Lee Sedol, the winner of 18 world Go titles, by four games to one in Seoul. Neither program could do anything else. Deep Blue could not have written an email.

“Virtually all” is the phrase that rules such systems out, because the 2x lead has to hold across 95% of remote jobs. A system can be far ahead of us on most tasks and still fall short on the rest. A 2023 study of more than 700 consultants at Boston Consulting Group called this uneven pattern a “jagged technological frontier”. On a task inside GPT-4’s abilities, consultants who used it did nearly 40% better than those who did not. On a task built to fall just outside its abilities, they did worse.

Skill profiles against the ASI line Bars show a made-up system's skill on eight tasks in standard deviations above the median. A dashed line marks the best human at 4 and a red line marks the ASI level at 12. The ASI clears the red line on all eight tasks. 0 4 8 12 SD above the median Best human (example) ASI line ChessCodingLawX-raysMathsEssaysSalesTaxes

Clears the ASI line on 8 of 8 tasks, so it passes the 95% test.

Figure 4. Press the buttons. The bar heights are made up to show how the rule works, and the best human’s position at 4 is an example.

The team also names a lower milestone for systems that only match the best of us. A TED-AI, short for top-human-expert-dominating AI, is “at least as good as top human experts at virtually all cognitive tasks.” Kokotajlo’s phrase to the Senate, “better than the best humans at everything,” sounds closer to this milestone. In the written definitions, ASI is the step after it.

The ladder below the top rung

The team’s model expects AI to climb five milestones, and it expects AI research to do most of the climbing. In the first stage, AI takes over the programming at AI companies. In the second, it also takes over research taste, which the team describes as “setting research directions, selecting experiments, and learning from experiments.” In the third, AI does all of the research on itself, and the model asks how much faster each round makes the next. An earlier post here, What happens when AI starts improving itself?, looks at how far that loop has got.

A ladder of five milestones, from automated coder at the bottom to superintelligence at the top AC SAR SIAR TED-AI ASI

1Automated coder (AC)

“An AC can fully automate an AGI project’s coding work, replacing the project’s entire coding staff.”

An AGI project is a company trying to build AI that can do almost everything. At this rung, AI writes all of that company’s code.

Median dates, August 2026: Kokotajlo Nov 2027 · Halstead Jan 2029 · Lifland Jan 2030

2Superhuman AI researcher (SAR)

“A SAR can fully automate AI R&D, making all human researchers obsolete.”

Coding is only half of research. This rung adds research taste, the judgment about which experiment to run next, at the level of the company’s best researcher.

The August 2026 charts give no date for this rung.

3Superintelligent AI researcher (SIAR)

“The gap between a SIAR and the top AGI project human researcher is 2x greater than the gap between the top AGI project human researcher and the median researcher.”

The same 2x rule as ASI, applied only to AI research and measured against the researchers at one company.

The August 2026 charts give no date for this rung.

4Top-human-expert-dominating AI (TED-AI)

“A TED-AI is at least as good as top human experts at virtually all cognitive tasks.”

The skill spreads from AI research to almost every remote job, at the level of the best human in each one.

Median dates, August 2026: Kokotajlo Nov 2028 · Halstead Apr 2031 · Lifland Jul 2032

5Artificial superintelligence (ASI)

“The gap between an ASI and the best humans is 2x greater than the gap between the best humans and the median professional, at virtually all cognitive tasks.”

The top rung. Three times as far from the middle of the bell curve as the best human, across almost every remote job.

Median dates, August 2026: Kokotajlo Mar 2029 · Lifland Jul 2033

Figure 5. Climb the ladder. Definitions are from the December 2025 model and dates are the medians in the chart captions of the August 2026 update. The robot is a cartoon.

The forecasters disagree by more than four years

In August 2026 the model’s authors updated their forecasts. Kokotajlo’s median for ASI was March 2029. His co-author Eli Lifland put his median at July 2033. A median is the date by which the forecaster thinks the milestone is as likely to have happened as not. Both expected the automated coder first, Kokotajlo in November 2027 and Lifland in January 2030.

These forecasts are “conditional on things going as fast as is technically feasible,” which means they assume no government steps in to slow AI down. In July 2026 the same group published AI 2040: Plan A, a scenario in which the US and Chinese governments do step in. Superintelligence arrives in 2040 in that story. Without the intervention, the authors write, “it would have happened much sooner (in 2030, to be precise).”

Median forecast dates for three milestones A timeline from 2026 to 2036. The April 2025 AI 2027 forecast put ASI at April 2028. In August 2026, Kokotajlo's medians were automated coder November 2027, TED-AI November 2028 and ASI March 2029. Halstead's were automated coder January 2029 and TED-AI April 2031. Lifland's were automated coder January 2030, TED-AI July 2032 and ASI July 2033. Shaded bars show the 80% ranges for the automated coder. 202620282030203220342036 AI 2027 Apr 2025 Kokotajlo Halstead Lifland to 2050to 2070 Automated coder TED-AI ASI 80% range for the automated coder

Medians for ASI: Kokotajlo March 2029, Lifland July 2033. The April 2025 forecast said April 2028. Halstead’s ASI date is not in the August charts.

Figure 6. Press the buttons. Medians and 80% ranges are from the chart captions in the August 2026 update, and the April 2025 date is from the AI 2027 takeoff forecast. An 80% range is the span the forecaster gives an 80% chance of containing the date. The red dashed line marks this post’s date.
Forecasts built on judgment

These dates move. The April 2025 forecast put the ASI median at April 2028, assuming a superhuman coder arrived in March 2027. By August 2026 the team judged that reality was “going about 70-90% as fast as AI 2027 predicted,” and their medians sat later than the 2025 forecast. They have also added a warning to the 2025 forecast. “This forecast relies substantially on intuitive judgment, and involves high levels of uncertainty.”

What the definition does not measure

Every part of the definition measures ability. None of it says what the system wants or whether it does what its makers ask. That question belongs to AI alignment, the work of making AI systems pursue the goals people intend for them. Kokotajlo told the senators, “Today’s AIs sometimes pursue goals other than the ones they were given, and sometimes hide that they are doing so.”

Who checks the answer

The bell-curve ruler assumes someone can rank an AI’s work against human work. At 4 SDs, the best human can still judge a colleague. A system at 12 SDs would be graded by people at least 8 SDs below it, and the definition leaves open how they would tell whether it had crossed the line.

The lowest rung will be tested first. An automated coder would do all of the programming at an AI company. Kokotajlo’s median for it is November 2027, a year after this post, and Lifland’s is January 2030.

Further reading

AI research superintelligence forecasting AI safety