UofAi · Capability Observatory
The honest read on AI progress
Three transparently-sourced dials. Every number shows its source, retrieval date, and data class.
Index snapshots as of 2026-07-02 (quarterly). Signal Feed latest issue 2026-08-28. Next Signal refresh 2026-11-06.
UofAi Compute Pulse
as of 2026-07-02
Components & sources (4)
- Largest known training run500,000,000,000,000,000,000,000,000FLOP
Largest publicly estimated run (Grok 4). Newer frontier runs likely exceed this but are undisclosed. Frontier training compute grows 4–5× per year.
- Global AI chip compute stock growth3.4× per year
Doubling roughly every 6.8 months. Largest single AI data center ≈ 800,000 H100-equivalents.
- AI chip performance per dollar37% per year improvement
Inference cost at fixed capability has been halving roughly every 2 months.
- Frontier training power draw100MW+ per frontier run
Growing 2.2–2.9× per year; single runs projected to reach 4–16 GW by 2030.
band 51.0–60.0
Superintelligence Horizon Index
as of 2026-07-02
Components & sources (3)
- Expert knowledge frontier (Humanity's Last Exam)59.2
59.2 is not the raw HLE score. It is (53.3 ÷ 90) × 100 — Scale/CAIS SOTA against the ~90% human-expert bar. Leaderboards disagreed 45.7–53.3% as of 2026-06-30; Scale/CAIS is canonical.
- Novel-skill acquisition (ARC-AGI-2 & 3)33.3
ARC-AGI-2 SOTA 54% (verified Dec 2025); ARC-AGI-3 best purpose-built agent 12.58% at March 2026 launch, frontier LLMs <1%. Human baseline 100% on both.
- Autonomous task horizon (METR 50%-reliability)74.5
Frontier 50%-horizon ≈ 12–16 hours (May 2026); METR notes measurements above 16 hours are unreliable with the current task suite. Target for SHI=100 on this subscale: one work-month (167 hours). Doubling time ≈ 4 months in 2024–2026, up from 7 months historically.
The expert-knowledge subscale is HLE SOTA scaled to the ~90% human-expert bar — (53.3 ÷ 90) × 100 = 59.2 — not the raw percentage. Read the methodology.
Human Parity Tracker
as of 2026-07-02
Human Parity ledger
Landmark human capabilities AI has reached, is approaching, or has yet to reach — each with falsifiable criteria and a citation.
Coding & Software
- Elite competitive-programming parity2025 · OpenAI reasoning modelsAchieved
Gold-medal / top-rank performance at IOI/ICPC-level contests.
- Autonomous resolution of most real-world software issuesFrontier coding agentsFrontier
Majority resolution on verified real-repository issue suites with human-level acceptance; strong but not yet reliable across arbitrary codebases.
Economic & Professional
- Majority of remote-work tasks at expert qualityOpen
Independent evaluation shows AI completing >50% of a representative remote-task basket at expert quality and cost parity.
- Award-competitive long-form creative work, judged blindFrontier
AI-authored long-form work places in blind professional judging; strong short-form results exist, long-form remains contested.
Games & Strategy
- Superhuman chess1997 · IBM Deep BlueAchieved
Defeats reigning human world champion in match play.
- Superhuman Go2016 · DeepMind AlphaGoAchieved
Defeats top-ranked human professionals in even matches.
- Superhuman multiplayer no-limit poker2019 · Pluribus (CMU/Meta)Achieved
Beats elite professionals in six-player no-limit Hold'em.
Language & Knowledge
- Broad academic exam parity (MMLU-class)2023 · GPT-4-class modelsAchieved
Exceeds expert-human accuracy on broad multiple-choice academic suites; now saturated above 88–90%.
- Passes controlled Turing-test studies2025 · GPT-4.5 (UC San Diego study)Achieved
Judged human more often than actual humans in a controlled three-party Turing test.
- Expert parity on Humanity's Last ExamOpen
SOTA reaches ~90% (human domain-expert average); current SOTA 45.7–53.3%.
Long-Horizon Agency
- One-workday autonomous task horizon2026 · Frontier agents (METR-measured)Achieved
50%-reliability time horizon ≥ 8 hours on METR methodology; frontier ≈ 12–16 hours as of May 2026.
- One-workweek autonomous task horizonOpen
50%-reliability horizon ≥ 40 hours; METR's current suite cannot yet measure reliably beyond 16 hours.
- One-workmonth autonomous task horizonOpen
50%-reliability horizon ≥ 167 hours (the SHI autonomy target).
Mathematics
- International Math Olympiad gold-medal performance2025 · OpenAI & Google DeepMind systemsAchieved
Official gold-medal-level score on IMO problems under competition conditions.
- Research-level mathematics (FrontierMath majority)Open
Solves a majority of FrontierMath research-tier problems.
Novel-Skill Acquisition
- Human parity on ARC-AGI-2Open
≥85% on the semi-private set at human-comparable efficiency; SOTA 54%.
- Human parity on ARC-AGI-3 interactive reasoningOpen
Matches the 100% human baseline on unseen interactive environments; best system 12.58% at launch, frontier LLMs <1%.
Perception
- Human parity in image classification2015 · ResNet-era modelsAchieved
Surpasses human error rate on ImageNet.
- Human parity in conversational speech recognition2017 · Microsoft ResearchAchieved
Matches human transcription error rate on Switchboard.
Physical & Embodied
- Broad driverless deploymentWaymo and peersFrontier
Fully driverless ride-hailing at scale in multiple metros; not yet generalized across all conditions/geographies.
- General household manipulationOpen
A commercially available robot reliably performs varied household chores in unmodified homes.
- General-purpose humanoid in commercial workOpen
Humanoid robots performing economically useful multi-task labor at scale without teleoperation.
Science & Research
- PhD-expert parity on Google-proof science Q&A2024 · Frontier reasoning modelsAchieved
Exceeds the 65% PhD-expert baseline on GPQA Diamond; SOTA now >90%, approaching saturation.
- Autonomous novel scientific discoveryMultiple lab systemsFrontier
An AI system independently generates and validates a novel, peer-review-published scientific finding end-to-end; current systems co-author and accelerate but require human direction.