Speaker A
There's a lot of topics to go through. Let me start with DeepSeek and the impact, Demis. To you, what have you and the team found?
DeepMind and Google discuss AI model efficiency, investment, regulation, and productivity at the Paris AI Summit.
Ask about this video. Answers come from its transcript only — with the timestamp, so you can check them.
Generated from the transcript and can be wrong — check the timestamp.
Key Takeaways
- DeepMind sees DeepSeek's cost claims as underestimated and not revolutionary in efficiency.
- Massive investments in AI infrastructure by Alphabet and the US government highlight AI's strategic importance.
- AI productivity gains require widespread adoption and practical application in business.
- Regulation must balance risk mitigation with fostering innovation and global coherence.
- Ethical concerns remain critical, especially regarding misuse and future advanced AI capabilities.
What the video covers
- DeepMind critiques DeepSeek's cost claims and efficiency compared to Gemini and ChatGPT.
- Alphabet plans to invest $75 billion in AI infrastructure, including data centers and chips.
- Compute remains critical for AI innovation, training, and inference, with demand growing for advanced models like Gemini 2.0.
- Google emphasizes building its own AI infrastructure to avoid overreliance on external suppliers.
- Broad-based productivity gains from AI depend on adoption, focused use cases, and empowering workers.
- Balanced AI regulation is necessary to manage risks while enabling innovation and productivity.
- The EU AI Act is progressing while the US federal government is scaling back AI regulations.
- Risks include misuse by bad actors and challenges related to advanced autonomous AI systems (AGI risk).
- Google and DeepMind commit to ethical AI policies and product guidelines.
- AI's potential impact on economy, science, and societal challenges is viewed as extraordinary.
Chapters
- 00:00DeepSeek Model and Cost Analysis
- 01:29Alphabet's AI Investment and Spending Plans
- 02:26Compute Needs and Infrastructure Strategy
- 04:00AI Productivity Gains and Adoption Challenges
- 04:59AI Regulation: US vs EU Perspectives
- 06:32Balancing AI Risks and Opportunities
- 07:14Ethical AI and Future Risks
- 08:12Google's AI Policy and Product Commitments
Full Transcript — Download SRT & Markdown
Speaker A
Does it support, do you believe them essentially when they say that they trained a model that has similar capabilities to Gemini and ChatGPT at a fraction of the cost, $6 million and all the chips? Do you believe that? Look at
Speaker A
the first thing I say is it's a very impressive model, a very impressive piece of work. I think the team is probably the best team that I've seen come out of China. So that said, I think a lot of the claims are exaggerated and a little bit misleading.
Speaker A
So first of all, when you report how much it cost to do a training run, they seem to report just their final training run, which is honestly a fraction of the cost it normally takes to explore and train and do all the
Speaker A
tests beforehand before you do your final run. So we think that's been sort of underestimated.
Speaker A
And then the other thing is they seem to have relied on some Western models to distill from or to basically fine tune against the output.
Speaker A
So you have to sort of factor that cost in as well. And then finally, like we, it's an impressive piece of work, but we don't see any silver bullet, new technologies or techniques that we haven't seen before, haven't invented before. They just applied it very well.
Speaker A
So it's impressive, but it isn't some new outlier on the efficiency curve. For example, Gemini is more efficient than DeepSeek in terms of its training to performance or its cost to performance.
Speaker A
We just don't talk about that very much. But it's actually more efficient on the frontier of cost to performance.
Speaker A
James, to you then and the company Alphabet's promise, or at least plan, to spend about 75 billion USD on CapEx, data centers, on chips, on H100 and probably Blackwell as well. And does any of that change the spending
Speaker A
plan for Alphabet? Do you scale back the spend as a result of the efficiencies that you're pushing for Gemini, which I mean, I would also that have been illustrated by what we're— No, no, it doesn't, because we're very
Speaker A
excited about the potential for this technology and the progress that we're seeing. And the opportunities are extraordinary.
Speaker A
Think about what this is going to do for people, the economy, advancing science and so on. Demis and colleagues just did with Fold and think about the things we can tackle to improve massive pressing challenges in society. So the potential that we see is
Speaker A
extraordinary. And I think at some point it's worth talking about what this could mean for economies and for productivity growth.
Speaker A
That's what we're excited about. Okay. So the investment still makes it still make sense.
Speaker A
Demis, we're looking at the US pushing through a staggering 100 billion, initially pushing up to 500 billion exclusively. It seems data centers, OpenAI does not. Does that give OpenAI and Sam Altman an unfair advantage?
Speaker A
Well, look, we'll see what that actually transpires to when it gets built. But just to echo what James said, compute is still very much a critical part of what the infrastructure needed for AI, not only for exploring new ideas.
Speaker A
If you're innovating at the frontier, you have to experiment at scale. Otherwise, the results of the experiment don't necessarily hold at the final training run scale. And also we need it for that.
Speaker A
We're seeing incredible demand for models like Gemini 2.0 models, and you need a lot of compute to solve that, which of course is what we've always been after. And then the final thing is with the advent of more thinking models, what's sometimes called thinking models or
Speaker A
inference type models, actually, the more spend compute time you spend at the point of processing inference, you get more powerful and better answers.
Speaker A
So actually, for all of those reasons, you need probably more compute than ever. Okay, sounds like NVIDIA is still in a safe position. Well, and also we're very happy with our use, of course, as well. So I Google and we actually take
Speaker A
advantage of AI infrastructure stock. So you don't want to be overly reliant on anybody. You have your own infrastructure stock going in. And also keep in mind that our current generation of TPUs, the sixth generation, is four times more efficient
Speaker A
than the last two generations. So we continue to push that frontier as well. Okay.
Speaker A
In terms of the productivity gains, James, when do we start for all of you?
Speaker A
As for investors, a key question remains our ROI, return on investment from AI. When do we start to see broad-based productivity gains as a result of this technology? I think it depends on all of us.
Speaker A
Some. We see enormous potential. Lots of estimates from McKinsey, Goldman Sachs, and all this investment in these huge productivity gains is going to depend on a few things.
Speaker A
First, adoption. We know from previous rounds of technology that until you get the large sectors and economies and companies in those sectors adopting it, using these technologies, it doesn't happen.
Speaker A
Second, we know that's going to be pretty important to focus on productivity-enhancing use cases in companies and organizations.
Speaker A
It's not enough for people just to play with these technologies. Otherwise, we're going to have another sort of paradox.
Speaker A
It's also going to be important to try to empower workers with these tools to be able to use them productively. So until those things happen, you know, we may not see the gains. You have experience in policy as well.
Speaker A
Formally at the Obama administration serving with the involvement. Obama administration was Trump right to cut and slash AI regulation in the U.S.? Well, we've always said regulation is important. It's important to regulate well.
Speaker A
So regulation is going to be very important in our minds. Regulations should do a couple of things.
Speaker A
First, it should address all of the risks that we see, but should also enable the things that we want. And so one of the key, key things we hope happens with regulation is that it's focused.
Speaker A
In other words, focus on applications, focus on gaps in the existing regulations. Focus also on the particular use cases and sectors. And we also need some coherence.
Speaker A
We're here in Paris, we don't need a patchwork of regulations around the world. So the more we can get coherence, that's going to be important, a long way from that. So I'll say one final thing, though.
Speaker A
It's got to be balanced. I think balance again, between addressing the concerns and risks, but also enabling the innovation, the productivity that we all hope will happen.
Speaker A
Armstrong, just to make it, the EU AI Act is coming into force this year through next year, and then the US is basically scrapping AI regulation, at least at the federal level.
Speaker A
Does how timely does that make it for DeepMind? Well, look, I just echo what James said is we got to get the balance right between embracing boldly the opportunities.
Speaker A
Things like drug discovery, I think, you know, helping with climate and energy, all the amazing things, advancing science, which is the motivation for why we started off this journey on increased productivity.
Speaker A
I think we're going to see all of those things and badly need them, but also addressing the risks that come with it. How is your assessment of the risks changed? Worst case scenario.
Speaker A
Best case scenario? Well, there's two main things that I worry about and one is how do we make sure we enable the good actors to use these incredible technologies for the benefit of the world, but also restrict access to would-be harmful, you know, bad actors doing harm, repurposing those
Speaker A
same technologies for harm. So that's a huge conundrum for the industry to try and solve. The second thing is as our systems become more powerful, not today's ones, but in a year or two or three years' time, as we start getting agent-based systems that are able to accomplish
Speaker A
tasks on their own that we set them, you know, addressing the kind of what's sometimes called AGI risk, risk inherent in these systems, making sure that control of all that is understandable,
Speaker A
In my opinion, the key part of that is international cooperation. We've got to do that as a whole, as an international community with academics, civil society and industry and government all together.
Speaker A
And that's what I think summits like this. James, very briefly, before we let you go, you've changed the AI principles.
Speaker A
Can we assume now that Alphabet is willing to work on AI weapons systems? We've been very clear about a few things.
Speaker A
We're very focused on application of this technology with benefits far, far, far outweigh the risks. We've also been very clear that we're going to stay consistent with international law and fundamental human rights. Look, we want to focus where we have
Speaker A
unique, distinct expertise, the incredible scientific work our colleagues are doing, advancing, prototyping, use case.
Speaker A
That's what we're going to do. Okay. We're also going to continue to make sure that our product policies as well as our terms of use, continue. That's what we're focused on.
Speaker A
James, thank you very much indeed. Senior Vice President Google and of course, Demis Hassabis, the CEO of Google DeepMind.
Topics:DeepMindGoogle AIGeminiDeepSeekAI regulationAI productivityAI investmentParis AI SummitAI infrastructureAGI risk
