Vision
Our vision of AGI
What we mean by AGI
A system that can learn and do most of the intellectual work people do, across many fields, reliably, improving with experience, and always under human control.
It has not been shown yet. Today’s systems reach top human level in olympiad maths and can learn unfamiliar puzzles, yet agents are reliable only on tasks of an hour or two, and robots fail most ordinary household tasks.
So we treat AGI as a research goal with a measurable path, not a marketing word. This site tracks that path in the open.
How we think AGI will be built
Our central view: general intelligence will come from a well-built system, not from one model alone. Reliable parts, organised into teams of agents that check their own work, can do what no single part can.
Models
Language and reasoning models are the base: they read, write, plan and reason. Ours and others’, chosen per task.
Tools
Search, code, data and software the system can use, so it acts on the world instead of only talking about it.
Memory
What the system has done and learned, kept over time, private to each user and deletable by them.
Teams of agents
Specialised agents that split a large goal into parts, hand work to each other and combine the results.
Verification
Tests, checks and proofs at every step, so the system knows when its own work is wrong and repairs it.
Human approval
People decide. Anything that sends, spends or deletes waits for a person, and every system can be stopped.
This is a prediction, and we hold it to the same rules as every other one. See the evidence and what would prove us wrong.
The path, step by step
Where research around the world stands on each step today.
- 01
Models that use tools
A model that can reason, search, run code and call other software in a loop.
Common practice since 2022, with open standards for connecting agents to tools since 2025. Evidence: Princeton University and Google researchers, Linux Foundation
Reached - 02
Top human level in hard reasoning
Solving problems that challenge the best people: olympiad maths, open algorithm questions, protein structure.
Officially graded gold-medal level at the 2025 Maths Olympiad, new algorithms found by AI, and Nobel-recognised protein work. Evidence: Google DeepMind, Google DeepMind, The Nobel Prize
Reached - 03
Agents on long, real work
Agents that finish real tasks lasting hours or days, reliably, without a person fixing them.
By early 2026 the best agents finished tasks of about 12 hours half the time, but only about 1.5 hours at 80% reliability, and some successes were cheating. Evidence: METR, METR
Partly reached - 04
Teams of agents doing research
Many agents with roles, shared memory and checks, producing results that hold up outside the lab.
Multi-agent systems have proposed hypotheses confirmed in lab experiments. Not yet reliable on long, open-ended projects. Evidence: Google researchers, Stanford University and Google researchers, Microsoft Research and collaborators
Partly reached - 05
Learning new skills like people
Exploring an unfamiliar situation and learning how it works from a few tries, at a cost close to a person’s.
ARC-AGI-3 went from under 1% (March 2026) to 99.9% (September 2026) with a memory-preserving harness, 62.7% with the standard one, at a cost far above a person’s. Evidence: ARC Prize Foundation, ARC Prize Foundation
Partly reached - 06
Working in the physical world
Robots that handle the variety of real homes and workplaces, not only controlled settings.
Robots succeeded at only 12% of real household tasks; factory humanoids are planned from 2028. Evidence: Stanford Institute for Human-Centered AI, Boston Dynamics and Hyundai Motor Group
Not reached - 07
General intelligence
One system that can learn and do most intellectual work people do, across fields, reliably and under human control.
Some company leaders said in September 2026 that the AGI era has begun. Independent researchers and the ARC Prize Foundation say the evidence does not show it yet. We agree with them for now. Evidence: Information Age, Over 100 experts, chaired by Yoshua Bengio, backed by over 30 countries and organisations
Not reached
What we hold ourselves to
Honest about progress
We never claim more than the evidence shows, about our work or anyone else’s.
People stay in control
More capable systems need more oversight, not less.
Private by default
User chats and files are never used for research or training without opt-in.
Checkable predictions
We publish what we expect, with confidence levels, and mark ourselves right or wrong.
For people everywhere
The benefits of advanced AI should reach places like Nepal and South Asia, not only a few countries.
Built together
We share results and work with others, and credit everyone who contributes.