Understand AGI › Syllabus › Module 3
Who has AGI in their stated mission, how each of them is structured, and what else is acting on the people making the claims.
It should not matter who makes a claim. The evidence should settle it. In practice, for AGI specifically, it matters a great deal - because most claims about frontier capability come from the organisation that built the system, evaluated it, wrote the report and benefits commercially from the conclusion.
That is not an accusation. It is a structural fact about a field where the systems are expensive enough that only a handful of organisations can build them, and where those organisations are also the ones with the access required to evaluate them properly.
So the question is not whether these organisations are honest. It is what pressures are on them, and which claims would look the same whether or not they were true.
The landscape section below describes the position as at September 2026. Organisational structure and stated goals change; where a specific current claim matters, it is reported with a name and a date on our sister site, AGI Made Clear.
Stated goal
What would they have to say either way?
Infrastructure
Who else can build, because of them?
Closing the gap
How quickly did that gap close?
The labs with AGI in the mission
A small number of organisations state AGI as an explicit goal rather than an eventual possibility - OpenAI, Google DeepMind and Anthropic most prominently. Each publishes research, each has a written position on what AGI would mean, and each has a different corporate structure attached to that position: a capped-profit arrangement under a non-profit in one case, a research organisation inside a large public company in another, a public benefit corporation in a third.
The large technology companies
Meta, Microsoft, Amazon, Nvidia and others are central to the field through models, compute, distribution or all three, without necessarily framing their work as an AGI project. Meta’s release of open-weight models has been the most consequential single decision here, because it changed who else can build.
The newer and non-US entrants
Organisations including xAI, Mistral and several Chinese labs have narrowed the gap to the frontier faster than most expected, and a meaningful share of them release open weights. Capability is considerably less concentrated than it was, which matters for every argument that assumes a small number of actors can be coordinated.
Everyone else
University groups, national institutes and independent evaluation organisations mostly cannot train frontier systems, and have moved towards the work that does not require it: evaluation, interpretability, measurement and critique. Some of the most useful recent results come from here, because these groups have no commercial stake in the answer.
Capital
Frontier training runs are extraordinarily expensive and the money is raised against a story about where the technology is going. A confident public timeline is not merely an opinion; it is an input to fundraising. This cuts both ways - it is also a reason a lab might be cautious, since a missed timeline is expensive too.
Talent
The pool of people who can do this work is small and mobile. Mission statements are recruitment instruments, and an organisation that describes itself as building AGI is competing for people who want to work on AGI.
Definitional latitude
Module 1 showed that the thresholds differ. An organisation with a commercial interest in having reached a threshold has some freedom over which threshold it is talking about. Watch for the definition moving between the technical report and the announcement.
Regulation
Claims about capability are also claims made to governments. There are incentives both to emphasise capability - it justifies a seat at the table - and to downplay it, when the alternative is a compliance regime.
Safety positioning
Several organisations argue that the technology is dangerous and that this is precisely why they should build it. Whatever you make of that argument, note that it requires the technology to be powerful, which aligns the safety case and the commercial case in an uncomfortable way.
Not everything an organisation publishes is the same kind of document.
Technical reports and system cards are generally the most useful. They contain the evaluation procedures, the caveats and often the failure cases, and they are written by people who have to live with being wrong.
Peer-reviewed and preprint research is where methods can be checked, though in this field the preprint is usually the first and last version.
Model cards and documentation tell you the intended use and the known limitations, which is frequently the most honest paragraph a company publishes.
Launch posts, keynotes and interviews are marketing. That is not a criticism - they are doing their job - but they are the layer where numbers lose their procedure and definitions lose their footnotes.
The rule of thumb: when a headline claim excites you, go and find the corresponding paragraph in the technical report. The gap between the two is usually the most informative thing available.
This is deliberately not a league table. Who is “ahead” changes faster than a page like this can honestly track, depends entirely on which capability you measure, and is exactly the kind of claim that needs a name and a date attached.
That work belongs to AGI Made Clear, which does it properly and dates it. What this module gives you is the structure to read that reporting with.
Last reviewed
Enrol early
We will email you when each module publishes. You will also get AGI, Made Clear, keeping you up to date on AGI and ASI developments in a field that moves faster than anyone can follow.
Double opt-in. One-click unsubscribe. We never share your address.
AGI Made Clear covers what is happening, the evidence behind it and the claims that are not supported. This site teaches the subject. That one tracks it.
Visit AGI Made Clear →