Google DeepMind has launched an interdisciplinary institute to work on the big open questions about artificial general intelligence. The timing is the interesting part. DeepMind's stated position is that AGI is close — and the company is now standing up a research body to settle what it is that is close, because the field still has no commonly accepted definition of the term. DeepMind's own is a system that possesses all the cognitive capabilities of the human brain.
The people at the company have been putting dates on it. Demis Hassabis expects AGI within the next few years. Shane Legg recently suggested that a precursor to it could appear by 2028. At OpenAI, the framing is different: the emphasis there is on AGI's economic value, and Sam Altman, the company's chief executive, has allowed that such a system could arrive by the end of this year.
Those three statements are routinely stacked next to each other as though they were forecasts of the same event. They are not. A system that clears an economic-value bar and a system that possesses every cognitive capability of the human brain are different objects, and "by the end of the year" versus "a few years" is not a disagreement about speed — it is a disagreement about what is being measured. That is precisely the gap an institute exists to close, and it is unusual to start closing it after the countdown has begun.
DeepMind also concedes what anyone using these systems already knows: today's AI still fails simple tasks and cannot work creatively. Hold that next to "a few years" and the claim gets thinner. Either the remaining distance is small, in which case the failures on simple tasks are strange, or the failures are load-bearing, in which case the dates are doing more rhetorical work than predictive work.
There is a further term to keep straight. AGI is not ASI — artificial superintelligence, a system that would significantly surpass humans in every respect. DeepMind's position is that AGI is probably a necessary stage on the way to it.
My read is that an institute is what you build when the finish line you have been racing toward turns out not to have been drawn. That is not a criticism of building it; definitional work is real work, and a field in which two of the leading labs use incompatible criteria for the same milestone needs it badly. But the announcement is quiet about the things that would tell you whether this is inquiry or positioning: who runs it, what it will publish, on what schedule, and — the one that matters — whether a conclusion it reaches would be allowed to contradict the company's public timelines. An institute that can only ratify Hassabis's few years is a communications asset. One that can tell him he is wrong is a research body.
The more interesting question is what happens if the definitional work succeeds. A rigorous account of AGI would be a test the labs could fail in public, on a date they themselves named.