How close are we to the milestones people use when they talk about superintelligence? This page does not make that call itself. Each milestone below takes its status only from the SiPoly claims linked to it and their published verdicts, using the rule under “How we decide”. Where no tracked claim covers a milestone, it says so.
11 milestones: 4 reached, 3 partly, 1 not yet, 3 with only open claims, 0 with no tracked claims yet.
- Reached
AI matches top human experts on hard tests
An AI system performs at the level of top human experts on a hard, recognised test of expertise (for example gold-medal level at the International Mathematical Olympiad, or PhD-level expert questions).
Latest evidence: · 8 linked claims
Linked claims
- Wrong 8.6%: AI gold-level performance at the IMO by 2025 XPT domain experts · says it happens · points to reached
- Wrong 2.3%: AI gold-level performance at the IMO by 2025 XPT superforecasters · says it happens · points to reached
- Wrong 8% that AI gets IMO gold by 2025 Paul Christiano · says it happens · points to reached
- Wrong >16% that AI has IMO-gold capability by end of 2025 Eliezer Yudkowsky · says it happens · points to reached
- Correct 4% that AI solves the hardest IMO problem by 2025 Paul Christiano · says it happens · points to not yet
- Correct PhD-level AI for specific tasks in "a year and a half" Mira Murati · says it happens · points to reached
- Open At least one model will match human experts across many industries before end-2026 Julian Schrittwieser · says it happens · open, not yet graded
- Open Bet: AI will do at least 8 of Marcus's 10 hard tasks by the end of 2027 Miles Brundage · says it happens · open, not yet graded
- Partly
AI agents complete multi-day tasks autonomously
AI agents reliably complete tasks that take a skilled person several working days, without step-by-step human help. A full working day (8 hours) counts as partway.
Latest evidence: · 4 linked claims
Linked claims
- Correct Models will work autonomously for full 8-hour days by mid-2026 Julian Schrittwieser · a step toward it · points to partly
- Open METR 50% time horizons of about two weeks around the start of 2028 Ryan Greenblatt · says it happens · open, not yet graded
- Open ~25%: 80%-reliability one-month task horizons by the start of 2028 Ryan Greenblatt · says it happens · open, not yet graded
- Open Within a decade, AI agents will independently do many software tasks that take humans days or weeks METR · says it happens · open, not yet graded
- Not yet (only open claims)
AI makes novel scientific discoveries
AI systems produce new scientific or mathematical results that experts recognise as genuine discoveries, not just assistance.
5 linked claims
Linked claims
- Open In 2026 AI will be able to make "very small discoveries" OpenAI · says it happens · open, not yet graded
- Open 2026 will likely see AI systems that figure out novel insights Sam Altman · says it happens · open, not yet graded
- Open "Pretty confident" of AI that makes more significant discoveries in 2028 and beyond OpenAI · says it happens · open, not yet graded
- Open By 2030 AI will outcompete most professional mathematicians at research Jacob Steinhardt · says it happens · open, not yet graded
- Open By 2030 many scientific fields will have AI assistants comparable to today's coding assistants Epoch AI · a step toward it · open, not yet graded
- Not yet (only open claims)
Software engineering is fully automated
AI systems can do essentially all the work of professional software engineers, end to end.
3 linked claims
Linked claims
- Open Median 2029 for full coding automation (as of Feb 2026) Daniel Kokotajlo · says it happens · open, not yet graded
- Open Median 2030 for full coding automation (AI Futures Model) Eli Lifland · says it happens · open, not yet graded
- Open By 2030 AI will autonomously fix issues, implement features and solve hard scientific programming problems Epoch AI · says it happens · open, not yet graded
- Partly
AI automates AI research (recursive self-improvement)
AI systems do essentially all of the research and engineering needed to build better AI, so that AI improves AI with little human input. An automated research assistant counts as partway.
Latest evidence: · 8 linked claims
Linked claims
- Kept An automated "intern-level research assistant" by September 2026 OpenAI · a step toward it · points to partly
- Open "Probably" by Dec 2026 the recursive self-improvement loop on algorithms will be closed David "davidad" Dalrymple · says it happens · open, not yet graded
- Open "Strikingly plausible" that models do the work of an AI researcher/engineer by 2027 Leopold Aschenbrenner · says it happens · open, not yet graded
- Open An autonomous AI team beats human-led model research on equal time and compute, setting its agenda (within ~12 months) Nathan Benaich · says it happens · open, not yet graded
- Pending A fully automated "legitimate AI researcher" by 2028 OpenAI · says it happens · open, not yet graded
- Open About a 50% chance that AI companies automate the whole AI research process by end of 2028 Daniel Kokotajlo · says it happens · open, not yet graded
- Open ~15%: AI capable of fully automating AI R&D by the start of 2029 Ryan Greenblatt · says it happens · open, not yet graded
- Open 45%: AI capable of fully automating AI R&D by the start of 2033 Ryan Greenblatt · says it happens · open, not yet graded
- Not yet (only open claims)
A frontier training run of 10^27 FLOP or more
A single AI training run uses at least 10^27 floating-point operations.
1 linked claim
Linked claims
- Open Frontier training clusters will cost over $100B by 2030, enabling ~1e29 FLOP runs Epoch AI · says it happens · open, not yet graded
- Partly
Humanoid robots in real deployment
Humanoid robots do useful paid work outside demos and pilots, in real workplaces or homes, in more than token numbers. Building or starting production of humanoids counts as partway.
Latest evidence: · 6 linked claims
Linked claims
- Broken "Several thousand" Optimus robots built in 2025 Tesla · a step toward it · points to not yet
- Broken at deadline Optimus 3 production to start "at the beginning of next year" (2026) Tesla · a step toward it · points to partly
- Correct 2025: humanoid hype, but nothing "remotely as capable as Rosie the Robot" Gary Marcus · says it will not happen · points to not yet
- Pending Optimus "production design 2" to launch in 2026 Tesla · a step toward it · open, not yet graded
- Open Home humanoid robots "all demo and very little product" in 2026 Gary Marcus · says it will not happen · open, not yet graded
- Open 2027 may see robots that can do tasks in the real world Sam Altman · a step toward it · open, not yet graded
- Reached
Driverless commercial fleets in several major cities
Members of the public can pay for rides (or freight) in vehicles with no safety driver, as a commercial service, in several major cities.
Latest evidence: · 8 linked claims
Linked claims
- Correct 80%: average person can hail a self-driving car in at least one US city by 2023 Scott Alexander · says it happens · points to reached
- Correct 80%: Waymo opens public driverless rides in a new city in 2024 Vox Future Perfect · says it happens · points to reached
- Wrong Driverless taxi with arbitrary pick-up and drop-off in a major US city: not before 2032 Rodney Brooks · says it will not happen · points to reached
- Kept Commercial driverless trucking in Texas in April 2025 Aurora Innovation · says it happens · points to reached
- Kept Zoox to start public robotaxi rides in Las Vegas later in 2025 Zoox · says it happens · points to reached
- Broken Tesla robotaxi in "eight to ten metro areas" by end of 2025 Tesla · says it happens · points to partly
- Kept Waymo One open to riders in Miami in 2026 Waymo · says it happens · points to reached
- Open Driverless taxi service in 50 of the 100 biggest US cities: not before 2028 Rodney Brooks · says it will not happen · open, not yet graded
- Reached
Binding AI-specific law in a major economy
A major economy has binding, AI-specific legal obligations in force (not only guidelines or voluntary commitments).
Latest evidence: · 3 linked claims
Linked claims
- Kept China to have initial AI laws, ethics norms and safety assessment by 2025 China State Council · says it happens · points to reached
- Partly kept AI Act to apply generally from 2 August 2026 European Union (AI Act) · says it happens · points to partly
- Pending Binding regulation on the most powerful AI model developers UK Labour Party / UK Government · says it happens · open, not yet graded
- Reached
A dangerous-capability threshold is declared
A frontier developer or government evaluator publicly says a released model has crossed a threshold for meaningfully enabling serious harm (for example biological or cyber misuse).
Latest evidence: · 1 linked claim
Linked claims
- Correct Capabilities able to cause a very serious incident arrive "this year" (2025) Miles Brundage · says it happens · points to reached
- Not yet
AGI declared or widely agreed
A major AI lab formally announces it has achieved AGI, or AI that can do most human jobs is widely agreed to exist.
Latest evidence: · 8 linked claims
Linked claims
- Correct 20% chance of AGI (replacing the majority of human jobs) by end of 2025 Andrew Critch · says it happens · points to not yet
- Correct 30%: a major AI lab will formally claim AGI in 2025 Vox Future Perfect · says it happens · points to not yet
- Correct No AGI in 2025 Gary Marcus · says it will not happen · points to not yet
- Wrong AGI, defined as smarter than the smartest human, "within two years" Elon Musk · says it happens · points to not yet
- Open By end of 2026 OpenAI will have an internal system Altman would call AGI Sam Altman · says it happens · open, not yet graded
- Open An AI company will claim AGI "probably in 2026 or 2027" Kevin Roose · says it happens · open, not yet graded
- Open No AGI in 2026 (or 2027) Gary Marcus · says it will not happen · open, not yet graded
- Open AGI in 2027 (State of AI Report ninth prediction) Nathan Benaich · says it happens · open, not yet graded
How we decide
A milestone’s status comes only from the SiPoly claims linked to it and their published verdicts. For each claim we ask what its grading says about the milestone event: a claim that said it would happen and was graded Correct (or came true late) counts as reached; one graded Wrong counts as not yet; Partly counts as partly. For a claim that said it would not happen, it is the other way round. For a forecast with a stated probability, we use whether the event actually happened, not whether the forecast was good. A claim linked as a step toward the milestone can only show “partly”. The milestone is Reached if any linked claim shows it reached, otherwise Partly, otherwise Not yet. If its claims are all still open, it shows “Not yet (only open claims)”; with no linked claims it shows “No tracked claims yet”. We do not judge milestones directly.
Milestone definitions and links are an editorial choice and can be corrected; tell us. JSON