The AI flood has reached the mountaintops and the last dry land is the physical world

Commonsense puzzles: underwater since 2019. The Winograd Schema Challenge, a set of sentences that need real-world knowledge to untangle, fell years ago: "By 2019, a number of AI systems... achieved better than 90% accuracy" 3.

Theorem proving: underwater. In July 2025, at the International Mathematical Olympiad (IMO), an advanced version of Google DeepMind's Gemini "solved five out of the six IMO problems perfectly, earning 35 total points," a gold-medal score confirmed by the IMO's president 4. In 2026 several companies claimed perfect scores, though those claims came from the companies rather than the IMO 5.

Programming: underwater, with an asterisk. On SWE-bench Verified, a set of real bug-fixing tasks, the top score was 97% as of September 2026 6, although OpenAI stopped using that test in February 2026 because models had memorized parts of it 7. A better measure comes from the research group METR (Model Evaluation and Threat Research), which found that the length of tasks AI agents can complete on their own "has been doubling approximately every 7 months for the last 6 years" 8. By May 2026 its estimate for one model was at least 16 hours of skilled human work, near the top of what its tests can even measure 9.

Science: the water is rising fast. The 2024 Nobel Prize in Chemistry went partly to Demis Hassabis and John Jumper "for protein structure prediction" with AlphaFold 10. In 2026, Google's AI "Co-Scientist" was reported in Nature to have "discovered new drug targets for liver fibrosis," with the caveat that "rigorous preclinical and clinical assessment would be required" 11.

Driving: going under. Waymo "surpassed 500,000 fully autonomous rides per week" in April 2026, across "11 major U.S. cities" 12.

Art and writing: contested, but wet. Google's Veo 3 video model can now "generate videos with audio," including "dialogue between characters" 13. In one 2025 study, trained writers strongly preferred human prose over prompted AI text, but once the AI was fine-tuned on an author's own books, they preferred the AI 14.

AI design: the peak that matters most. Google DeepMind's AlphaEvolve found a way to speed up a key part of Gemini's own architecture "by 23%, leading to a 1% reduction in Gemini's training time" 15, and in September 2026 OpenAI said it had built an automated "research intern" that can do "well-defined research tasks under human direction" 16.

The clearest sign of how fast this moves is ARC-AGI-3 (the AGI in its name stands for artificial general intelligence), a puzzle test built in 2026 specifically to be easy for people and hard for AI. At launch in March, "Humans score 100%. Frontier AI scores 0.51%" 17. By September, one model scored 99.9% with a specialized setup 18.

The test's own makers stress that maxing it out "would not represent 'proof of achieving AGI'" 18, which is fair, but it's a striking six months, and it fits the longer trend METR found, where the length of task an AI agent can finish doubles about every 7 months, which sounds modest until you run it forward.

Here's the twist Moravec saw coming, and honestly it's the part that surprised me most. He also noticed that "it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility" 19. That's now called Moravec's paradox, and it has held up beautifully.

A February 2026 review by Epoch AI found that "Navigation is deployed commercially, while most industrial and household tasks are not," and that folding basic laundry is "validated in lab" and still short of everyday use 20. Hard open math is dry too: on a set of open problems posed by the mathematician Paul Erdős, the best system solved 3% 21, and on Humanity's Last Exam, a test of expert-level questions, the top score is around 55% 22.

So the last dry land is a strange mix of the hardest frontier research and the things a toddler does without thinking.