Behind the Scenes of Demis Hassabis’ AI Safety Body Push

[00:00:03] Welcome to Tech News Briefing. It's Friday, August 14th. I'm Belle Lin, a reporter for the Wall Street Journal Leadership Institute. Health obsessives are no strangers to collecting data. In fact, they often have a constant read on their workouts, step counts, sleep metrics, heart rates, and nutrition data. Now that AI is in the mix, they're using that data to build super-powered health coaches.

[00:00:30] Then, one of the pioneers in machine learning has been spending more and more of his time advancing the idea of an organization that would create safety guardrails and best practices for developing artificial general intelligence, or AGI. We've got an exclusive report on what that might look like and what it could mean for the tech industry.

[00:00:53] But first, health-conscious techies are combining their data wizardry and love of spreadsheets with the power of AI chatbots to build hyper-personalized health coaches who can be used for everything from understanding a poor night's sleep to revamping a diet or training for a marathon.

[00:01:16] WSJ's Natalie Kaufman spoke with our colleague Katie Dayton to explain what's behind this recent data-maxing trend. Natalie, can you start by giving us an example from one data-maxer to show us how this all works? What kind of data did they share with the chatbot and what insights did they get back? Yeah, so we have a former competitive runner who now trains on his own.

[00:01:39] And when he went independent after quitting the track team, he really wanted his fitness trackers to help him keep, you know, a structured regiment for training. But he quickly realized that his fitness bands were no replacement for his coach. So what he did was he turned to AI and he used Anthropics Clawed chatbot to help him analyze his sleep and running data,

[00:02:02] as well as some biometrics from his doctor's visits, so things like blood work and height, weight, all that stuff. So what he does is he inputs all that data into Clawed and asks it, how should I be adapting my training to cohere with everything that's going on in my body? And so Clawed generates him some suggestions on how he should work out better. Wow, so like a real personal trainer there. What were some of the more surprising or unexpected use cases you came across?

[00:02:31] Someone we interviewed actually hooks up his Whoop fitness band to his Google calendar to see which co-workers at work stress him out the most. So Whoop does measure things like your stress and your heart rate and such. So from there, he was able to pinpoint which co-worker was really giving him some of that anxiety when he went to the office. That was pretty funny. I wonder what kinds of questions are people asking these chatbots that traditional health apps just can't answer right now?

[00:02:58] A lot of it has to do with not really answering one question, but having this way to aggregate all of your data in one place. You can get deeper insights across all of your trackers. And how reliable are these AI coaches? Obviously, there's not a huge take up of all of this right now, but from what you can tell, how accurate are they? There's no really way to pinpoint if something's accurate in terms of, oh, this training plan really got me to have a faster 5K or something like that. So it's really variable based on the person.

[00:03:28] But I would say, and the people I spoke with say the same thing, kind of take it with a grain of salt because this isn't a doctor. It isn't a personal trainer. It's not a coach that has been experienced in this kind of thing. And AI can make hallucinations. So it's important to take it as sort of a guide point or guidepost for what you do with your body, but not as like this is exactly what you should be doing to maximize your health. I can definitely see that this should by no means replace doctors.

[00:03:57] I wondered if, you know, your sources and people you've spoken to, have they given the indication that they might replace human coaches though? Some people really talked about the money saving part of this. You know, a personal trainer can cost tons of money per session. But instead of having to pay for that, you might either pay for a chatbot or not pay at all and use the free tier and get the insights you're still kind of looking for. And then they also talk about this idea of not having to bother their coach at weird hours of the day.

[00:04:26] So sort of an on demand tool that you can use. These tools only work because people are handing over incredibly personal information. How are the people you spoke with thinking about privacy? A lot of them say that the upside outweighs the risk involved with inputting all of this personal health data into these LLMs. So it really comes down to personal comfort.

[00:04:50] But these so-called data maxers are really interested in optimizing every corner of their lives, including their fitness, their schedules and everything like that. So they're not as worried about privacy and they're willing to take kind of some of the tradeoffs that come along with it. That was the WSJ's Natalie Kaufman speaking with our colleague Katie Dayton. How would you data max your life? If you're a listener on Spotify, leave us a comment with your answer.

[00:05:17] Coming up, regulating and overseeing AI safety is top of mind for AI leaders, as well as government officials around the world. And Google DeepMind's Demis Asabis has been mulling over the topic too. We'll get into our exclusive reporting after the break.

[00:05:43] Last week, machine learning pioneer Demis Asabis stepped down as CEO at Google DeepMind, taking on a new role as chief scientist at Elphabet. In the weeks before he relinquished the job, he was in discussions with peers from AI labs and high-ranking Trump administration officials about a new idea, the formation of a new independent industry safety entity to shape the future of AGI.

[00:06:11] WSJ reporter Amrith Ramkumar joins us now to break down Asabis' discussions, what they tell us about the state of regulating the tech, and Google's position in the AI race. So, Amrith, what can you tell us about the group that Demis just discussed putting together? Yeah, it's pretty complicated. It has a few different aspects. And the idea is to create AI standards and consistency across the leading AI companies to prevent bad things from happening.

[00:06:39] And there would also be a global component, and Demis Asabis has described it as essentially something like the IAEA, which coordinates on global nonproliferation for nuclear weapons. So, the idea would be that the U.S. would lead a coalition of other countries that would sort of align on standards and best practices to prevent cyber attacks, bioweapons, things like that. So, you see Demis playing this role as a global ambassador for AI, essentially,

[00:07:07] and pitching other AI executives and U.S. government officials on this plan in recent weeks. But why would he do this? What's Demis' goal, do you think? All the top AI executives are very worried. They see these mounting threats from the technology they're creating, and they see a huge public backlash against AI. And they also see the U.S. government and other governments around the world largely not doing enough to rein in some of these risks.

[00:07:34] So, the goal is to have industry play a big role in setting the standards. And they want to do that so that they don't get saddled with some really burdensome regulation down the road. Another part of it is each of these AI executives is trying to curry favor with leaders around the world and also show that they are uniquely qualified to lead the industry moving forward. You've seen Sam Altman at OpenAI, Dario Amadei at Anthropic, and now Demis, all doing this sort of thing.

[00:08:01] What they're trying to do is cultivate this image as a global AI leader because they know AI and chips are going to drive geopolitics moving forward. So, those are the main reasons. And another is also just that Google is trying to seem as forward-leaning on AI as possible because they're behind in the AI race. So, the more they can be part of things like standard setting, it's all helpful for branching their image in the AI race. Interesting. Some optics going on there.

[00:08:30] When you talked about reining in some of the risks, what would that look like potentially? Nobody knows, frankly. That's a big part of why people are sort of doing these brainstorming exercises and proposals. Demis and others are proposing companies would voluntarily at the beginning submit their models to this group of experts, and the experts would run tests on them, see how powerful they are, engage the risks. And if there were issues, they would be flagged, and then the companies could potentially hold the model back in consultation with the government

[00:08:59] or figure out some alternative changes to make it safer. And then over time, that voluntary testing would transition to become sort of mandatory, where this group of experts would have so much credibility that the only way that the public would trust a model essentially would be if it had that stamp of approval. What kind of progress did Demis make on putting this group together?

[00:09:21] So in June, at the Group of Seven summit in France, he and other executives spoke to President Trump and world leaders about the need for a U.S.-led coalition to coordinate standards globally on AI. And then around that time and in the subsequent weeks, he met with U.S. government officials, including Treasury Secretary Scott Bessent and Michael Kratios, one of the president's top tech advisors. And he's also spoken to executives at other AI companies about this as well and spoken about the need, again, for global coordination.

[00:09:49] And that's so important because we know China is making great advances in this field and worrying a lot of other AI executives who see these models getting more powerful and maybe not having to abide by the same rules or guardrails that Western labs have to follow. What did Google or Demis have to say about your reporting? The main thing they said was that he's just as committed as ever to Google and they're doing everything they can essentially to stay apace in the AI race.

[00:10:15] They didn't really deny that he's been having these conversations and Demis didn't respond to a request for comment. All right. And if we kind of step back a little and take everything that we just talked about in a broader context, what does it tell us about the state of the AI world and the industry right now? The state of model regulation is pure chaos right now.

[00:10:39] Companies have no idea how they're going to get their models released publicly in a safe way with U.S. government oversight. The Trump administration is in the middle of sort of overhauling how it's done AI regulation. And they've instituted a process where leading companies like OpenAI, Anthropic and Google will submit their models voluntarily again, maybe voluntarily in air quotes because the government is asking them to do it. But then those models will be tested by security experts and the government could have a say in who gets access.

[00:11:08] And that's been a trend over the last few months with Anthropics Mythos and other powerful models coming out. But these AI executives are very worried because this model review period can lead companies to delay the rollout. And even rolling out a model a few weeks late can mean big dollars in terms of users adopting other tools or going to competitors. That was WSJ reporter Amrith Ramkumar. And that's it for Tech News Briefing. If you're a listener on Spotify, be sure to leave us a comment.

[00:11:38] Today's show was produced by Julie Chang. Jessica Fenton and Michael LaValle wrote our theme music. Our supervising producer is Katie Ferguson. Our development producer is Aisha El-Muslim. And Chris Sinsley is the deputy editor of audio for The Wall Street Journal. We'll be back later this morning with TNB Tech Minute. Logging off, I'm Belle Lin, a reporter for The Wall Street Journal Leadership Institute. Thanks for listening.