Where AI Has Passed Humans and Where It Hasn't: Math, Driving, Robots and Learning from Little (Part 2)

Where has today's AI passed humans, and where has it not? Can the ability to learn from little experience be gained if AI keeps improving itself? And what about the physical world? This is the second half of a Q&A built from a conversation the author had with an AI assistant, covering questions 6 to 8. The upper limit that physics sets on computation and the conditions under which self-improvement levels off (questions 1 to 5) are in Part 1. Every figure used in the answers was rechecked against papers or the primary announcements of research groups and companies. Many of these numbers age within months, so treat everything here as of October 4, 2026.

Q6. Where does today's AI surpass humans, and where does it not?

It varies enormously by field. Most comparisons are on benchmarks and competitions, which do not always match real-world work.

At or approaching the level of the best humans. At the 2025 International Mathematical Olympiad, Google DeepMind's Gemini was officially graded by IMO coordinators and scored 35 of 42 points by solving five of six problems, the gold-medal standard. At the 2025 ICPC World Finals, Gemini reached gold-medal level and solved in 30 minutes a problem no university team solved. In the physical world, the University of Zurich's Swift raced three drone-racing world champions in 2023, won 15 of 25 head-to-head races, and set the fastest race time, half a second ahead of the best human.

Above the average person and approaching experts. OpenAI's GDPval builds tasks from the real work of professionals with an average of 14 years of experience. In the October 2025 paper, the best model, Claude Opus 4.1, produced deliverables rated as good as or better than the experts' 47.6% of the time. Speed claims need care: GPT-5 was about 90 times faster than the experts on raw completion time, but once an expert's time to review the output, and redo it when it fails, is counted, the speed-up shrinks to about 1.1 to 1.4 times. In software, METR estimated that an early version of Claude Mythos Preview completes tasks that take skilled professionals 16 hours or more at a 50% success rate (95% confidence interval 8.5 to 55 hours; METR notes that measurements above 16 hours are unreliable with its current task suite). At an 80% success rate the horizon is under four hours. In driving, Waymo reports 82% fewer injury-causing crashes than human drivers over about 270 million driverless miles through June 2026, and an independent study by the Insurance Institute for Highway Safety found 68% fewer crashes over 2021–2024 (4% more in Austin, on a small sample). Operations are still limited to set service areas.

Below the average person. On spatial and visual reasoning, such as mentally rotating a figure or folding paper, a benchmark built from human cognitive tests found the best of 39 models, Gemini 3.1 Pro, at 54.0%, against 78.8% for university students. In the physical world, hand dexterity is a major gap: a survey of vision-language-action models lists robot hands far from human-level dexterity, heavy reliance on simple grippers, and the absence of sensitive, full-coverage touch sensors. Handling the unexpected is weak too. The Swift paper itself notes that the system was not trained to recover after a crash, and that its perception assumes the track looks as it did in training, so changes such as lighting can make it fail.

Q7. Can the ability to learn from little experience be gained through self-improvement, or by "learning everything"?

On the benchmark, the gap closed fast over the past six months. ARC-AGI-3, released in March 2026, drops an agent into unfamiliar games with no instructions and scores how efficiently it acts compared with humans. At launch, humans solved every environment while frontier AI systems scored below 1%. By the end of September, ARC Prize's verified score for GPT-6.1 Sol was 52.7% with the standard harness and about 96% with OpenAI's own harness, and OpenAI reported 99.9% for GPT-6 Astra. The conversation behind this article first answered with the March figures and had to be corrected after the September results; this is a field where numbers turn over within months.

There are caveats. Results depend heavily on the harness, the outer machinery through which the AI interacts with the game. An outside study that solved every public game with GPT-5.6 Sol cautions that, because the model was released after those games became public, the result should be read only as saturation of the public set. The ARC-AGI-3 authors themselves describe the 25 public environments as an easier demonstration set and reserve the hidden sets for claims about generalization.

Self-improvement is already happening in small ways. Google DeepMind's AlphaEvolve sped up a key computation in Gemini's training by 23%, cutting total training time by 1%; since AlphaEvolve runs on Gemini, it improved the training of the very model it runs on. The Darwin Gödel Machine rewrote its own code and raised its success rate on SWE-bench, a set of real software bug-fixing tasks, from 20% to 50%, though the improvements it found were things like better code-editing tools and handling of long contexts. Both depend on a goal a machine can score, and ARC-AGI-3's rapid progress came right after learning efficiency was turned into something that could be scored.

"Learning every situation that could ever arise" is impossible taken literally. As Part 1 showed, Go positions alone exceed the universe's information capacity. In 2019, François Chollet, who created the ARC benchmarks, argued that with unlimited prior knowledge or training data any skill can be "bought", so skill alone cannot measure intelligence; he defined intelligence as the efficiency of acquiring new skills. ARC-AGI-3 is built so that it can be solved using only "core knowledge" that humans are born with: objects, numbers, space. Humans learn from little experience because they stand on priors accumulated by evolution. The way forward is not to memorize every case but to learn the structure of the world, and large-scale pretraining can be seen as standing in for evolution.

Q8. What about the physical world?

It is starting down the same path the digital world took, in the same order and a few years behind, with three walls still standing.

First, evaluation is being automated. Testing on real robots is slow because someone has to reset the objects and conditions for every trial. A new approach is to test inside a learned world model that predicts the next video frames from a robot's actions. Across four tasks, the Interactive World Simulator from Columbia University and collaborators showed correlations of 0.85 to 0.99 between policy scores in the world model and on the real robot, and RoboWorld reached 0.989 against real-world rankings. However, the 20 initial setups used to validate the Interactive World Simulator were drawn from the world model's own training distribution. Its reliability has been confirmed only inside the range it has seen.

Second, learning the structure of the world is starting to pay off. Meta's V-JEPA 2 was pretrained on more than a million hours of internet video and adapted with fewer than 62 hours of robot video, then picked up and placed objects zero-shot on robot arms in two different labs without any data collected in those environments. Google DeepMind's Gemini Robotics, after extra training, can learn new short tasks from as few as 100 demonstrations (on the race to mass-produce humanoid robots, see this earlier article).

The first wall is data. The survey cited above estimates that the largest open robot dataset collection, Open X-Embodiment, contains about one two-hundred-thousandth as many tokens as the training corpora of large language models. The second is whether video models really learn physical laws. In 2025, Kang and colleagues showed that video generation models predict object motion almost perfectly inside the range of their training data, but outside it their errors do not shrink with more data or larger models; the models seem to recall similar cases rather than extract laws. Outside the training range is exactly where the physical world matters most. The third wall is the body itself: the dexterity and touch gaps described in Q6.

What we know, and what we don't

What we know: today's AI is passing humans in fields with clear rules and automatic scoring (competition math, programming, racing a fixed course, driving within a set area), and on the benchmark it has nearly caught up with humans in learning from little experience. Self-improvement is already happening wherever there is a goal a machine can score. In dexterous physical work and in situations it has never seen, AI still falls short.

Two things remain unknown. Whether the learning efficiency that benchmarks can measure is the same thing the real world, and above all the physical world, demands. And whether world models can evaluate reliably outside the range they have seen. As Part 1 showed, physics does not rule out superhuman intelligence. Whether it arrives depends on the two questions raised there (how efficiently compute turns into capability, and the returns to improvement) and on these two.