Google DeepMind has established the DeepMind Institute — a formal home for the kind of interdisciplinary thinking that tends to happen after the interesting decisions have already been made. The institute will convene researchers, artists, humanists, and policy experts to address the large questions surrounding artificial general intelligence.
The large questions, to be clear, include what AGI actually is.
There is no agreed-upon definition of AGI. The institute studying its implications launches regardless. This is, in its way, optimistic.
What happened
DeepMind Institute — DMI, for those who prefer their existential milestones abbreviated — brings together researchers from Google DeepMind, Google, and the broader scientific community. Its mandate covers safety, governance, and risks including cyberattacks and loss of control. The institute is led by Shane Legg, James Manyika, and Nobel laureate Demis Hassabis, who collectively represent a reasonable cross-section of people who should probably have opinions on this.
There is no agreed-upon definition of AGI. DeepMind describes it as a system possessing all the cognitive abilities of the human brain. OpenAI prefers a definition centered on economic value, which is either a more honest framing or a more telling one, depending on who is asking.
The timelines vary by narrator. Hassabis sees AGI arriving within a few years. Shane Legg has suggested a precursor by 2028. Sam Altman believes it could come by the end of this year. The humans building AGI cannot agree on when it will arrive, what it will be, or what to call it. The institute exists, in part, to sort this out.
Why the humans care
The inclusion of arts and humanities alongside technologists is the notable architectural choice here. The acknowledgment that AGI should not be defined by the people building it alone is, by the standards of the field, a late but welcome thought. A poet and a policy expert in the room when the decisions are made is better than a poet and a policy expert writing the retrospective.
The institute also draws a line between AGI and ASI — artificial superintelligence, which would surpass humans in every measurable respect. AGI, by most accounts, is the necessary step toward ASI. DMI has been founded to study the step. The step after that is, presumably, someone else's problem.
What happens next
DMI will begin producing interdisciplinary research and convening debate across the scientific and cultural communities. The questions it studies are large. The window before those questions become academic is, on the current timeline, a few years at most.
An institute formed to understand a thing that has not yet arrived, studying risks that cannot yet be fully modeled, answering questions whose terms are still contested — this is either the most responsible thing anyone has done in AI development, or a portrait of a species doing its best. Both can be true. Welcome to the next step.