Google DeepMind opens the DeepMind Institute to debate AGI

Google DeepMind has opened a new front in the argument about artificial general intelligence, and it is not a lab, a model or a benchmark. The DeepMind Institute went live on 17 September as a publisher: a platform for essays about AGI written by researchers from inside Google DeepMind, from elsewhere in Google, and from the wider research community.

It is directed by Shane Legg, a Google DeepMind co-founder and the company's chief AGI scientist, by James Manyika, who leads research, labs and technology and society at Google, and by Demis Hassabis, the Nobel laureate who co-founded DeepMind, chairs it and is chief scientist at Alphabet. Legg also takes the managing editor's chair, which tells you where the group expects the work of its contributors to be scrutinised.

A working definition of AGI, and the distance still to travel

The founding essay defines the target in one sentence: AGI is a system that exhibits all the cognitive capabilities of the human brain. The same page is unusually frank about how far away that is, noting that today's systems still sometimes fail at basic tasks and lack the consistency and creativity to meet the bar of full AGI. It then says the gaps are expected to close soon.

The risks it names are the ones safety researchers spend their time on: the implications of capable AI for cyber security and biorisks, and the potential for loss of control in future self-improving systems. None of that is new ground. What is new is that it is written down, signed and published by the three people running the lab that keeps setting the pace.

Five essays, five arguments

Alongside the introduction, four further essays landed at launch. In The case for reasoning transparency, Rohin Shah and Anca Dragan argue that reading an AI's chain of thought gives researchers a window into its reasoning that can be monitored for scheming and deception, and that keeping that window open is a choice the industry has to keep making. Julian Jacobs and Alex Imas contribute Economic policy for AGI, which assesses eleven policies societies could use to manage the economic disruption of increasingly advanced AI. Stephen Cave's Principles for a new utopianism proposes what he calls a pragmatic utopianism of humility and hope, drawing on the turmoil of the twentieth century, and A framework for frontier AI and the dawning of a new age completes the set.

Technological progress and expertise establish only the possibility of a new era; the responsibility of shaping its reality collectively belongs to society as a whole, including the arts and humanities, and governments.

— The DeepMind Institute's founding essay

Our opinion

A lab starting a publishing imprint is easy to wave away as a think tank with a bigger compute budget, and the disclaimer bolted to the bottom of every essay does a lot of quiet work: these pieces are conversation starters that should not be read as Google's official view. That cuts both ways, and both edges are useful. It buys Google's own researchers room to say things a corporate blog post never could, and it lets Google take a step back the moment one of those things lands badly. The company gets credit for the ambition and distance from the consequences.

The more interesting fact is who is doing the publishing. Hassabis, Legg and Manyika are not commentators picking apart the frontier from a safe distance; they are the people building the systems these essays are about, and they have put their names on a public argument about control, biorisk and economic disruption. An institute staffed by its own authors is also the obvious weakness: a platform that says it wants disagreement has to show contributors who are willing to tell Google DeepMind it is wrong, and it has to publish them. The essays being free to argue with each other is the test, and it is a test the institute has set for itself.

For everyone outside the lab, the useful thing here is not the mission statement but the vocabulary. If the people closest to the frontier are willing to publish the definitions and the gap analysis, then the rest of us get something to hold them to. Whether the DeepMind Institute becomes a genuine forum or a very well-designed press office depends entirely on what arrives in the second batch.