Learn. Share. Evolve…

0:00:06 Welcome to the Next Wave Podcast. I’m Matt Wolfe, and I could not be more excited to
0:00:11 share today’s episode with you. So we’ve gone from AI that can chat with you to AI that
0:00:18 can work for you. And the difference? Well, this new AI can actually think through problems,
0:00:24 catch its own mistakes, and complete complex tasks from start to finish, just like a human
0:00:30 employee would. This is what everyone in the AI world calls AI agents. You’ve probably heard the
0:00:36 term. But here’s why this breakthrough changes everything for regular people. If you’re someone
0:00:41 who codes, there’s no more debugging AI hallucinations. It can actually check its own
0:00:48 work. If you run a business, these AI agents can actually plan out and finish complex tasks,
0:00:53 just like one of your employees might. And the implications for humanity? Well,
0:00:59 with these new tools, drug discoveries, testing, and real-world trials can now take weeks instead
0:01:05 of decades. In fact, Isometric Labs is already gearing up for human trials of AI-discovered drugs
0:01:11 right now. And we’re also already getting stories about how AI has successfully diagnosed human
0:01:19 illnesses when human doctors couldn’t. These things are accelerating insanely fast. But this isn’t just
0:01:26 about better chatbots. We’re talking about AI that understands the physical world, plans weeks ahead,
0:01:31 and even works while you’re asleep. And the company that’s leading the charge in all of this
0:01:38 is Google DeepMind. They’ve already used this thinking AI to predict protein structures that used to take
0:01:45 years. Now it just takes seconds. It’s called AlphaFold. They’ve also invented AI that can invent new
0:01:53 algorithms, including AI algorithms. It’s called AlphaEvolve. It’s insane stuff. Two million researchers
0:02:00 worldwide are using their tools right now. But with this power comes some pretty big questions. Can we
0:02:07 trust it? What happens to privacy? What about our jobs? Can we trust Google with our data?
0:02:13 So I sat down with Google DeepMind CEO Demis Hassabis to get answers straight from the source
0:02:22 about how we got from autocomplete to actual thinking and what comes next. He’s a Nobel laureate,
0:02:28 a knight, and one of the most influential pioneers in the world of AI. And somehow,
0:02:34 I managed to get him to sit down and chat with me about all of this. What he told me will change
0:02:42 how you think about AI forever. So without further ado, here’s my conversation with Sir Demis Hassabis.
0:02:52 Cutting your sales cycle in half sounds pretty impossible, but that’s exactly what Sandler
0:02:57 Training did with HubSpot. They used Breeze, HubSpot’s AI tools, to tailor every customer interaction
0:03:01 without losing their personal touch. And the results were pretty incredible. Click-through
0:03:08 rates jumped 25%, and qualified leads quadrupled, and people spent three times longer on their
0:03:13 landing pages. Go to HubSpot.com to see how Breeze can help your business grow.
0:03:21 Hey Demis, great to see you again. So my first question for you is, can you sort of describe
0:03:26 what’s happening under the hood with an LLM? Like what’s kind of going on? Can we sort of demystify
0:03:33 it for people a little bit? I can try. So I mean, at the basic level, what these LLM systems are trying
0:03:38 to do is very simple in a way. They’re just trying to predict the next word. And they do that by obviously
0:03:46 looking at a vast training set of language. But the trick is not just to regurgitate what it’s already seen, but actually
0:03:52 generalize to something novel that you are now asking it. And it seems like, you know, we’ve managed with the
0:03:55 modern day systems is to get that generalization to work.
0:04:01 Gotcha. So at IO, you announced the new DeepThink, right, which is so much more powerful, and it’s topping all of the
0:04:06 benchmarks for things like coding and math and all that. What happened under the hood that caused that new leap?
0:04:12 Well, new techniques have been brought into the foundational model space where there’s called
0:04:16 pre-training, where you sort of train the initial base model based on, you know, all the training
0:04:21 corpus. Then you try and fine tune it with a bit of reinforcement learning feedback. And now there’s this
0:04:27 third part of the training, which is we sometimes call inference time training or thinking, where you’ve
0:04:34 got the model, and you give it many cycles to sort of go over itself and go over its answer, maybe do use
0:04:40 some tools. For example, it could fact check with search, something like that, before it outputs the
0:04:45 answer to the user. So it gets a chance to sort of correct itself and adjust what it’s going to
0:04:50 output. And of course, if you do that, you get a much better answer. And then what Deep thinks about
0:04:56 is actually taking that to the maximum and giving it loads more time to think and actually even doing
0:05:01 parallel thoughts and then choosing the best one. And it turns out it works really well. And, you know,
0:05:07 we pioneered that kind of work in the past, actually nearly a decade ago now with AlphaGo and our games
0:05:12 playing programs, because in order to be good at games, you need to do that kind of planning and
0:05:14 thinking. And now we’re trying to do it in a more general way here.
0:05:20 Right, right. So it almost kind of thinks of a whole bunch of potential responses and then goes
0:05:23 through reviews all the potential responses and then figures out what the best response from those
0:05:28 potential responses. Exactly. And it can go over and correct some parts of it and use tools to check
0:05:33 some aspects of it. So, you know, especially in certain areas like maths and the coding, it really
0:05:39 improves the answers. Amazing. Very cool. So you’ve mentioned that the long term goal is to sort of
0:05:45 let these AIs have like a world model. Right. So can you sort of explain what you mean by a world
0:05:51 model and what does that open up to us? Well, so we’re all familiar with large language models now,
0:05:57 but of course, we have five sensors and we operate in the real world. And language is only one aspect,
0:06:04 very important aspect of our world and human civilization, but only one aspect. And so I think for a model,
0:06:09 what we mean by a world model is a model, sometimes we call it a multimodal model that can understand
0:06:16 not just language, but also audio, images, video, all sorts of input, any input, and then potentially
0:06:22 also output any kind of token as well. And the reason that’s important is if you want a system to be a good
0:06:27 assistant, it needs to understand the physical context around you. Or if you want robotics to work
0:06:32 in the real world, the robot needs to understand the physical environment. Right. So in order to do that,
0:06:35 you have to have what we sometimes like to call a world model.
0:06:40 Cool. So what sort of new things do you think that’ll open up to people once they have that ability?
0:06:45 I think robotics is one of the major areas. I think that’s what’s holding back robotics today. It’s
0:06:49 not so much the hardware, it’s actually the software intelligence. You know, the robots need to understand
0:06:54 the physical environment. But I think that that’s also what will make today’s sort of nascent
0:06:58 assistant technology and things like you saw with Project Astra that we show in Gemini Live.
0:07:04 For that to work really robustly, you want as accurate as world models you can. And then the
0:07:10 other thing is, if you want to do planning in the real world, you need to sort of plan multiple
0:07:15 steps with your world model. So in order for that to be good for long range planning, your world model
0:07:20 has to be very accurate as well, which is pretty hard when you’re talking about real world situations.
0:07:29 My First Million hosted by Sam Parr and Sean Puri is brought to you by the HubSpot Podcast Network,
0:07:34 the audio destination for business professionals. My First Million features famous guests like Alex
0:07:40 Hromozzi, Sofia Amoroso, and Hasan Minhaj sharing their secrets for how they made their first million and
0:07:45 how to apply their learnings to capitalize on today’s business trends and opportunities.
0:07:52 They recently had a fascinating episode about how you can scale a profitable agency with zero employees
0:07:57 using AI agents. Listen to My First Million wherever you get your podcasts.
0:08:05 So you’ve mentioned things like AI will be able to, most likely in the future, solve things like
0:08:10 room temperature superconductors and, you know, more energy efficiency and curing diseases.
0:08:15 Out of the sort of things that are out there that it could potentially solve, what do you think the
0:08:20 sort of closest on the horizon is? Well, as you say, we’re very interested and we actually work on
0:08:25 many of those topics, right? Whether they’re mathematics or things like material science, like
0:08:30 superconductors, you know, we work on fusion, renewable energy, climate modeling. But I think
0:08:35 the closest if you think about and probably most near term is building on an alpha fold work.
0:08:41 And we spun out a company called Isomorphic Labs to do drug discovery. We think that we sort of
0:08:46 the whole drug discovery process from first principles with AI. And normally, you know,
0:08:52 it takes the rule of thumb is around a decade for a drug to go from sort of identifying why a disease
0:08:57 is being caused to actually coming up with a cure for it. And then finally being available to patients.
0:09:02 It’s a very laborious, very hard, painstaking and expensive process. And I would love to be able
0:09:09 to speed that up to a matter of months, maybe even weeks one day and cure hundreds of diseases like that.
0:09:14 And I think that’s potentially in reach. It sounds maybe a bit science fiction like today,
0:09:19 but that’s what protein structure prediction was like, you know, five or six years ago before we
0:09:24 came up with alpha fold and used to take years to find painstakingly with experimental techniques,
0:09:29 the structure of one protein. And now we can do it in a matter of seconds with these computational
0:09:34 methods. So I think that sort of potential is there. And it’s really exciting to try and make that happen.
0:09:40 Amazing. So you guys just announced Alpha Evolve recently, which looks amazing, right? It’s an AI
0:09:46 that essentially can help you come up with new algorithms, right? So how close are we to AIs that
0:09:51 are sort of designing new AIs to improve the AIs and then we start entering the cycle?
0:09:56 Yes. I mean, it’s a baby step in that direction. I think it’s really cool breakthrough piece of work
0:10:02 where we’re combining kind of, in this case, evolutionary methods with LLMs to try and get
0:10:08 them to sort of invent something new. And I think there’s going to be a lot of promising work actually
0:10:13 combining different methods in computer science together with these foundation models like Gemini
0:10:18 that we have today. So I think it’s a great, very promising path to explore just to reassure everyone
0:10:22 when it still has humans in the loop, scientists in the loop to kind of, it’s not directly improving
0:10:28 Gemini. It’s using these techniques to improve the AI ecosystem around it, slightly better algorithms,
0:10:32 better chips that the system’s trained on versus it’s the algorithm that it’s using itself.
0:10:38 Right, right. So AI agents, they’ve been sort of a big talk in the AI community recently.
0:10:43 And this week at IO, we saw Project Mariner, which can go and open up 10 different browsers and
0:10:48 go and do a whole bunch of things on your behalf. How far off do you think we are to
0:10:52 being able to give an agent like a week’s worth of work and then goes and executes that for us?
0:10:57 Yeah, I mean, I think that’s the dream to kind of offload some of our mundane admin work and
0:11:01 also to make things like much more enjoyable for us. You know, you have maybe have a trip to
0:11:06 Europe or Italy or something, and you want the most amazing itinerary sort of built out for you and
0:11:11 then booked. I love our assistance to be able to do that. You know, I hope we’re maybe a year away or
0:11:16 something from that. I think we still need a bit more reliability in the tool use. And again,
0:11:21 the planning and the reasoning of these systems, but they’re rapidly improving. So as you saw with
0:11:26 the latest Project Mariner, and so it’d be great for that to come together with some of the other
0:11:29 advances we’re making with the Gemini Live and the Astro technology.
0:11:34 Yeah. What do you think the biggest bottleneck is right now to sort of getting that long-term agent?
0:11:38 I think it’s just the reliability of the reasoning processes and the tool
0:11:45 And making sure because each one, if it has a slight chance of an error, if you’re doing like
0:11:50 a hundred steps, even a 1% error doesn’t sound like very much, but it can compound to something
0:11:55 pretty significant over a hundred, you know, 50 or a hundred steps. And a lot of the really interesting
0:12:00 tasks you might want these systems to help you with will probably need multi-step planning and action.
0:12:05 Gotcha. So I want to shift gears a little bit here and talk a little bit about some of the
0:12:09 sort of fears and concerns that have come up in like my YouTube comments and things like that.
0:12:14 You know, people are worried about things like privacy and losing their jobs to AI and all of
0:12:20 that kind of stuff. And so I’m curious, how does a company like DeepMind build the trust of the general
0:12:23 public that you can trust them with this kind of technology?
0:12:26 Yeah. Well, look, I think we’ve tried to be, and I think we try to be responsible role
0:12:31 models actually with these frontier technologies. Partly that’s showing what AI can be used for,
0:12:36 for good, you know, like medicine and biology. I mean, what better use could there be for AI than
0:12:42 to cure, you know, terrible diseases. So there’s always been my number one thought there, but there’s
0:12:45 other things, you know, where it can help with the climate, energy and so on that we’ve discussed.
0:12:50 But I think we’ve got to, you know, companies is incumbent on them to behave thoughtfully
0:12:55 and responsibly with this powerful technology. We take privacy extremely seriously at Google,
0:12:59 always have done. And I think, you know, most of the things we’ve been discussing with the
0:13:03 assistants, they would be opt-in. They’ll make the universal assistant much more useful for you,
0:13:08 but you would be, you know, intentionally opting into that very clearly with all the transparency
0:13:13 around that. And what I want us to get to is a place where the assistant feels like it’s working
0:13:19 for you. It’s your AI, right? Your personal AI, and it’s working on your behalf. And I think that’s
0:13:22 that’s the mode, you know, that’s the, at least the vision that we have and that we want to deliver
0:13:25 and that we think users and consumers will want.
0:13:31 So all of those are incumbent. And actually I would say to your viewers as well, you have a lot of
0:13:36 say in this in the sense of like, you should exercise your consumer choices and buy services
0:13:41 and products from companies that you feel are acting responsibly and the leadership is acting
0:13:45 responsibly and you like the type of work that they’re doing. Because now we’re entering this sort
0:13:51 of commercialization, productization era of AI now. Right. And you know, I think your viewers
0:13:55 and everyone has a big say in that. Right, right. So one of the things that you guys also demoed at
0:14:01 IO that I got a chance to actually test out a little bit earlier was the Android XR glasses. And those were
0:14:06 absolutely mind blowing when I tried them the first time. And so I guess the flip side of this sort of
0:14:12 privacy thing is if everybody’s sort of walking around wearing glasses that have microphones and cameras on
0:14:18 them, how do we ensure that the sort of privacy of the other people around us is secure?
0:14:21 I think it’s a great question. I mean, first thing is to make it very obvious that you’re
0:14:26 it’s on or off in these types of things, you know, in terms of the user interfaces and the form factors.
0:14:30 I think that’s number one. But I also think this is the sort of thing where we’ll need
0:14:36 sort of a societal agreement and norms about how do we do we all want if we have these devices,
0:14:41 they’re popular and they’re useful. What are the guardrails around that? And I think that’s why we’re
0:14:45 only entrusted tester at the moment is partly the technology is still developing, but also we need
0:14:50 to think about the societal impacts like that ahead of time, you know, not just with the technology,
0:14:55 but also society in general and civil society kind of inputting into what might be the right way to
0:14:59 handle that type of world. Right. So I’ve got one last question here. It’s kind of a two-parter
0:15:05 question. So what excites you most about what you can do with AI today? And what excites you most about
0:15:10 what we’ll be able to do in the very near future? Cool. Well, today, I think it’s the AI for science work is my,
0:15:15 you know, always been my passion. And I’m really proud of what AlphaFold and things like
0:15:18 that have empowered. They’ve become a standard tool now in biology and medical research, you know,
0:15:23 over two million researchers around the world use it in their incredible work and vital work. So that’s
0:15:30 fantastic to me. In the future, you know, I’d love a system to basically enrich your life and work for
0:15:36 you on your behalf to protect your mind space and your own thinking space from all of the digital world
0:15:41 that’s bombarding you the whole time. And I think actually one of the answers to that is that we’re
0:15:45 all feeling in the modern world with social media and all these things is maybe a digital assistant
0:15:50 working on your behalf that only at the times that you want surfaces the information rather than
0:15:54 interrupting you at all times of the day. Amazing. Well, thank you so much, Demers. This has been
0:16:10 absolutely fascinating. I really, really appreciate the time that you spent with me today. So thank you
0:16:18 We’ve got a major announcement. HubSpot is the first CRM to launch a deep research connector with
0:16:24 ChatGPT. Customers can now bring their customer context into the HubSpot deep research connector
0:16:30 and take action on those insights. Now you can do truly remarkable things for your business. Customer
0:16:36 success teams can quickly surface inactive companies, identify expansion opportunities and receive targeted
0:16:43 plays to reengage pipelines, then take those actions in the customer success workspace in HubSpot to drive
0:16:49 retention. Support teams can analyze seasonal patterns and ticket volume by category to forecast staffing needs
0:16:56 for the upcoming quarter and activate breeze customer agents to handle spikes and support tickets. This truly
0:17:04 is a game changer for the first time ever. Get the power of ChatGPT fueled by your CRM data with no complex setup.
0:17:10 The HubSpot deep research connector will automatically be available to all HubSpot accounts across all
0:17:17 tiers that have a ChatGPT team enterprise or Edu subscription. Turn on the HubSpot deep research connector
0:17:26 in ChatGPT to get powerful PhD level insights from your customer data. Now let’s get back to the show.