Shane Legg on DeepMind’s Founding Mission, Fifteen Years Later

I never met a man with a mission who could show me the mission, printed and audited, on a single sheet. DeepMind was founded in 2010 with a statement of intent that was, at the time, close to unfundable: solve intelligence, then use it to solve everything else. The company’s chief scientist has been the most consistent public voice for what that mission actually requires.

Fifteen years later, the mission statement remains, and the field has changed around it in ways that make it considerably harder to evaluate.

What the mission demands

The first clause is a research programme. The second is a claim about consequences, and it is the part that generates most of the difficulty. A laboratory inside a large advertising-funded company has to demonstrate that its work serves a public interest, while its parent company’s incentives point elsewhere. That is not a scandal; it is arithmetic.

• Publishing research in a field where publication can be competitively costly.
• Taking on problems with no near-term commercial application.
• Maintaining scientific credibility while being owned by a platform.

The measurement problem

There is no agreed metric for whether a lab has used intelligence well. Citations, prizes and published results measure the first clause effectively. The second is assessed by anecdote and argument, which is why the same evidence supports both a generous and a sceptical reading of any large lab’s record. I have watched two men read the same river gauge and reach opposite conclusions, and both of them had boats to lose.

The honest version

Legg’s public posture has been to treat general intelligence as a long-term engineering project with a serious risk profile — a position more modest than some of the field’s rhetoric but considerably more defensible than the claim that the problem is solved. Modesty is not a popular commodity in a boom, and it is the one I would buy.

The safety research that came out of it

Those results spread quickly because they were published and because they addressed problems every lab had. That is a different model of influence from the one where a company keeps a safety breakthrough private.

That is a different model of influence from the one where a company keeps a safety breakthrough private.

The independence question

The recurring tension is between the lab’s scientific identity and its position inside a company with commercial objectives. Research priorities are set through a negotiation that is not public, and the outcomes are visible mainly through what gets published and what does not. When the minutes are missing, a man learns to read the table instead.

Legg has generally declined to discuss internal disagreements, which is understandable and also means that external observers have to infer the state of that negotiation from the record of outputs. I do not blame a man for keeping his own counsel.