Superintelligence: definitions, evidence and timelines

The common thread is AI that greatly exceeds human cognitive performance. The differences are in the comparison: one clever person, every human, or the best humans working with AI. These bars are far apart, so a claim that superintelligence has arrived means little until it names its definition. See Superintelligence has many definitions and no test.

On bounded tasks with a checkable answer, often yes. On long, unsupervised work and in the physical world, no. The pattern is uneven rather than a single line being crossed. Details and caveats are in Where AI already beats the best humans in 2026.

Not from today's benchmarks alone. Epoch AI found answer-changing errors in 46% of sampled Humanity's Last Exam questions, FrontierMath's hardest tier is fully solved, and METR says its time-horizon suite cannot measure above 16 hours. A credible test would report the weakest domain rather than the average, compare against the best humans with AI tools, give cost and reliability curves rather than one score at 50% success, require long tasks with little supervision, use fresh audited items with the harness disclosed, count outcomes in the world, and be run by someone other than the developer.

Forecasts moved earlier through 2026. The AI Futures Project, authors of AI 2027, now expect a superhuman AI researcher to “likely arrive at some point in 2027” and give 35–65% to AI that dominates top experts before 2028. The largest survey of AI researchers (2023) put a 50% chance on machines beating humans at every task by 2047. The main new evidence is METR's November 2026 finding that Anthropic is improving about 1.9× faster than a year earlier, mostly because of AI; most of the evidence behind it is redacted. See The case for a 2027 intelligence explosion and Anthropic's AI R&D threshold.

Power and build times: Epoch expects no data centre to clearly exceed Colossus 2 before the second half of 2027, and the grid sets the near-term ceiling (our analysis). Learning: deployed models do not learn on the job (agents that learn on the job). And verification: METR says the measured acceleration may be limited to tasks that grade themselves, which is where optimisation finds the grader's blind spots.

Our work measures the two failure points a self-improving system has to get past. On verification, Thurstone bounds Best-of-N shows that once a verifier can be fooled at any rate, more samples first help and then hurt in theory, and constructed exploits beat a verifier far more often than sampled ones. A pre-registered test on three real judges found no decline up to 32 samples (Genesys program). On selection and memory, forgetting as selection and the findings ledger track what models lose when they are updated. The threshold audit toolchecks a lab's capability threshold against the evidence offered for it.

• Research Desk • Nov 10, 2026Superintelligence has many definitions and no test. Here is what a test would need.Labs now say superintelligence is in sight, but their definitions compare AI with different humans, and none comes with a way to check. Benchmarks saturate, ceilings are set by label errors, and the best long-task metric stops working at 16 hours.
• Research Desk • Nov 10, 2026Where AI already beats the best humans in 2026, and where it still doesn'tAI now matches or beats top humans on olympiad maths, contest programming and most reasoning benchmarks, and has disproved a famous conjecture. It still fails on long reliable work, learning after deployment and the physical world, and many of the strongest claims are graded by the companies themselves.
• Research Desk • Nov 10, 2026The case for a 2027 intelligence explosion now rests on evidence the public can't seeForecasters moved their dates for automated AI research back to 2027 this week, on the strength of METR's finding that Anthropic is improving about twice as fast as a year ago. Most of the evidence behind that finding is redacted, and the gains may be limited to tasks that grade themselves.
• Research Desk • Nov 09, 2026Anthropic told the SEC it may soon cross its own AI R&D line. The evidence is mostly its own.Anthropic's amended IPO filing says it “may imminently meet” its policy's definition of AI-driven acceleration. The numbers that would settle it are internal, redacted and lagging.
• Research Desk • Nov 09, 2026Investors are paying for agents that learn on the job. The hard part is knowing what they forgot.Trajectory just raised $45M at a $300M valuation to make deployed models keep learning. But on CL-Bench, the best system gained only 25.4% over a stateless baseline, and plain long context beat dedicated memory.
• Research Desk • Nov 09, 2026AI's 2027 ceiling is set by the grid, not the chip. Optics buys back only a few percent.PJM is 6.8 GW short, turbines take six years, and FERC just pushed the backstop to 2027. Interconnect is a small slice of facility power, so photonics can't fix this.
• Paper • Preprint draft • Jul 27, 2026Thurstone bounds Best-of-N: a 1927 identity, calibrated for modern verifiersThe noise floor of Best-of-N selection is Thurstone's 1927 Case V identity, not a new law. Contributed: blind-spot rates ε ≈ 0.002–0.05 from inverting published overoptimization curves, an exact consensus threshold p/(1−p), and a budget rule N*(ε) with a working tool. Heavy-tailed score noise breaks the identity, so estimate ε directly.
• Paper • Research note • Jul 30, 2026When the exploit is constructed: ranking collapses, counting survivesAt a matched exploit rate, Best-of-8192 accuracy is 0.862 with sampled exploits and 0.000 with exploits constructed against the verifier — while plurality voting among high scorers holds at 0.999. The consensus threshold missed a dominant-wrong-answer regime and was corrected; the two-regime framing is prior art (Dwork et al. 2015).
• Paper • Preprint draft • Jun 27, 2026Catastrophic forgetting in class-incremental learning is a verifier-exploitability problemFreeze one near-perfect expert per task and all of class-incremental difficulty becomes selection — which expert to trust. Accuracy factorizes exactly as selection accuracy × within-task accuracy, and the selection term is pinned by the selector's exploitability ε. A November 2026 follow-up shows that selecting by a statistical model of shared features instead of expert confidence lifts Split-MNIST to 0.926.
• Paper • Synthesis • Jul 29, 2026What survived: a claims register with falsifiers for a pre-registered AI research programEvery surviving claim from a pre-registered, run-once program, with its evidence and its falsifier. Alive: pool-relative verifier exploitability, test-time recurrent depth, the AIMC window-policy rule, narrow optical interconnect. Dead: wave/complex substrates, optical compute for inference, frontier-datacenter AIMC, and the ε 'universality class'.
• Research tool • In-browser · Pythonthreshold-auditBefore you register a threshold: its standard error, the instrument's ceiling and the noise band of range statistics. After the run: was the miss malformed, underpowered or substantive?