Navigating AI Alignment and Safety

Technology doesn’t force us to do anything — it merely opens doors. But military and economic competition pushes us through.

That’s how today’s guest Allan Dafoe — director of frontier safety and governance at Google DeepMind — explains one of the deepest patterns in technological history: once a powerful new capability becomes available, societies that adopt it tend to outcompete those that don’t. Those who resist too much can find themselves taken over or rendered irrelevant.

Links to learn more, highlights, video, and full transcript.

This dynamic played out dramatically in 1853 when US Commodore Perry sailed into Tokyo Bay with steam-powered warships that seemed magical to the Japanese, who had spent centuries deliberately limiting their technological development. With far greater military power, the US was able to force Japan to open itself to trade. Within 15 years, Japan had undergone the Meiji Restoration and transformed itself in a desperate scramble to catch up.

Today we see hints of similar pressure around artificial intelligence. Even companies, countries, and researchers deeply concerned about where AI could take us feel compelled to push ahead — worried that if they don’t, less careful actors will develop transformative AI capabilities at around the same time anyway.

But Allan argues this technological determinism isn’t absolute. While broad patterns may be inevitable, history shows we do have some ability to steer how technologies are developed, by who, and what they’re used for first.

As part of that approach, Allan has been promoting efforts to make AI more capable of sophisticated cooperation, and improving the tests Google uses to measure how well its models could do things like mislead people, hack and take control of their own servers, or spread autonomously in the wild.

As of mid-2024 they didn’t seem dangerous at all, but we’ve learned that our ability to measure these capabilities is good, but imperfect. If we don’t find the right way to ‘elicit’ an ability we can miss that it’s there.

Subsequent research from Anthropic and Redwood Research suggests there’s even a risk that future models may play dumb to avoid their goals being altered.

That has led DeepMind to a “defence in depth” approach: carefully staged deployment starting with internal testing, then trusted external testers, then limited release, then watching how models are used in the real world. By not releasing model weights, DeepMind is able to back up and add additional safeguards if experience shows they’re necessary.

But with much more powerful and general models on the way, individual company policies won’t be sufficient by themselves. Drawing on his academic research into how societies handle transformative technologies, Allan argues we need coordinated international governance that balances safety with our desire to get the massive potential benefits of AI in areas like healthcare and education as quickly as possible.

Host Rob and Allan also cover:

• The most exciting beneficial applications of AI
• Whether and how we can influence the development of technology
• What DeepMind is doing to evaluate and mitigate risks from frontier AI systems
• Why cooperative AI may be as important as aligned AI
• The role of democratic input in AI governance
• What kinds of experts are most needed in AI safety and governance
• And much more

Chapters:

• Cold open (00:00:00)
• Who's Allan Dafoe? (00:00:48)
• Allan's role at DeepMind (00:01:27)
• Why join DeepMind over everyone else? (00:04:27)
• Do humans control technological change? (00:09:17)
• Arguments for technological determinism (00:20:24)
• The synthesis of agency with tech determinism (00:26:29)
• Competition took away Japan's choice (00:37:13)
• Can speeding up one tech redirect history? (00:42:09)
• Structural pushback against alignment efforts (00:47:55)
• Do AIs need to be 'cooperatively skilled'? (00:52:25)
• How AI could boost cooperation between people and states (01:01:59)
• The super-cooperative AGI hypothesis and backdoor risks (01:06:58)
• Aren’t today’s models already very cooperative? (01:13:22)
• How would we make AIs cooperative anyway? (01:16:22)
• Ways making AI more cooperative could backfire (01:22:24)
• AGI is an essential idea we should define well (01:30:16)
• It matters what AGI learns first vs last (01:41:01)
• How Google tests for dangerous capabilities (01:45:39)
• Evals 'in the wild' (01:57:46)
• What to do given no single approach works that well (02:01:44)
• We don't, but could, forecast AI capabilities (02:05:34)
• DeepMind's strategy for ensuring its frontier models don't cause harm (02:11:25)
• How 'structural risks' can force everyone into a worse world (02:15:01)
• Is AI being built democratically? Should it? (02:19:35)
• How much do AI companies really want external regulation? (02:24:34)
• Social science can contribute a lot here (02:33:21)
• How AI could make life way better: self-driving cars, medicine, education, and sustainability (02:35:55)

Video editing: Simon Monsour
Audio engineering: Ben Cordell, Milo McGuire, Simon Monsour, and Dominic Armstrong
Camera operator: Jeremy Chevillotte
Transcriptions: Katy Moore