Research program
Reasoning and generalisation
The question
What makes a system learn new tasks as cheaply and reliably as people do?
Generalisation tests have fallen fast: ARC-AGI-3 went from under 1% to 99.9% in 2026, but only with a harness that preserves the system's memory and at a cost far above a person's. We study where today's reasoning still breaks, why the system around a model matters so much, and how to generalise cheaply.
Our predictions
Our view, with how sure we are. Each one links to the evidence it rests on.
On the road to AGI
AI will match people at learning genuinely new tasks from a few examples, at a similar cost, which we see as one of the last clear gaps before AGI.
What changes
AI systems will run whole cycles of research in narrow fields, from hypothesis to experiment design to analysis, with scientists approving each step.
Progress around the world
Public results from other teams that shape our view.
· ARC Prize Foundation (testing OpenAI's model)
GPT-6 Astra scores 99.9% on ARC-AGI-3, with a caveat
ARC Prize verified 62.7% with its standard harness and 99.9% with a harness that keeps the model's reasoning state between steps; in one setting it used fewer actions than the median person on 96% of levels. ARC Prize states it is not claiming this is AGI.
What it means for AGI: the gap on learning new tasks closed within six months of the test's launch. The jump from 62.7% to 99.9% came from how the system keeps memory, not from a new model, which supports the view that AGI is a system, not one model. The run cost about $18,800, far more than a person would.
· ARC Prize Foundation
ARC-AGI-3 launches: people score 100%, frontier AI under 1%
A new interactive benchmark of game-like environments with no instructions, rules or stated goals. At launch people scored 100% and the best frontier AI 0.51%.
What it means for AGI: it tests exploring and learning in a new situation, the skill most clearly missing from AI at the start of 2026.
· Google DeepMind (graded by IMO coordinators)
AI reaches officially graded gold-medal level at the Mathematical Olympiad
An advanced Gemini Deep Think model solved five of six problems for 35 of 42 points, working in natural language within the 4.5-hour limit. IMO coordinators graded it with the same criteria as students.
What it means for AGI: hard, creative reasoning in plain language, once thought far off, reached top human level.
· Insilico Medicine and collaborators (published in Nature Medicine)
An AI-discovered drug shows results in a randomised phase 2a trial
Rentosertib, a drug whose target and molecule were found with generative AI, was tested in 71 patients with idiopathic pulmonary fibrosis. Safety was similar to placebo, and the highest dose showed a lung-function gain, though the trial was small.
What it means for AGI: AI-designed medicines have reached patient trials. Approval still requires larger trials.
Source: nature.com(opens external site)
· Google DeepMind
AlphaEvolve: a coding agent that discovers better algorithms
An evolutionary agent built on Gemini models found a way to multiply 4×4 complex matrices with 48 multiplications, improving on Strassen's 1969 result, and improved on known answers for a share of over 50 open maths problems.
What it means for AGI: a loop of models proposing, testing and keeping the best ideas can produce new knowledge, an early form of self-improvement.
· Google DeepMind (published in Nature)
AI weather model beats the leading physics-based ensemble forecast
GenCast, a machine-learning ensemble model, was more accurate than ECMWF's ENS on 97.2% of 1,320 targets tested, including extreme weather and cyclone tracks, and produces a 15-day forecast in minutes.
What it means for AGI: learned models can outperform decades of hand-built simulation in a complex physical system.
Source: nature.com(opens external site)
· ARC Prize Foundation
ARC Prize: generalisation is still hard for AI
A public competition built on the ARC-AGI puzzles, which most people solve easily but AI systems find hard, to reward progress on learning new skills from few examples.
What it meant for AGI in 2024: strong exam scores did not mean general intelligence; learning new tasks efficiently was the clearest open gap. See ARC-AGI-3 in 2026 for how that changed.
· Google DeepMind (published in Nature)
AI finds 2.2 million new crystal structures
Graph networks for materials exploration (GNoME) proposed about 2.2 million new crystals, of which about 380,000 were predicted to be stable.
What it means for AGI: AI can search spaces far larger than people can. Lab confirmation is still the slow step.
Source: nature.com(opens external site)
· Google DeepMind (published in Nature)
AI discovers faster sorting algorithms used in a standard C++ library
A reinforcement-learning agent found sorting routines faster than human-written ones for short sequences; they were added to the LLVM libc++ library.
What it means for AGI: AI can improve the basic tools computing is built on, a small step toward systems that improve themselves.
Source: nature.com(opens external site)
· Google DeepMind (published in Nature)
AlphaFold predicts protein structures at near-experimental accuracy
A deep-learning system predicted the 3D shape of proteins with accuracy close to lab experiments, solving a problem biology had worked on for about 50 years.
What it means for AGI: AI can make real scientific progress in a narrow field. The open question is whether the same can happen across many fields at once.
Source: nature.com(opens external site)
Publications
No publications yet.
Collaborations
Open to collaboration.