Hey, and good morning, everybody. I'm super excited to be here with all of you. Thank you for joining us on this pretty busy week in London. You're here at the 10th annual Research and Applied AI Summit. So, it's been 10 years we're doing this event.
So, as those of you who have come before, this is all about creating a community of entrepreneurs, researchers, operators, people in government, policy — trying to create a room with really interesting minds where you can have interesting conversations and talk about the current state of the art in AI, what's happening around the world, and what might come next.
And we do a lot of work to try to curate a great audience in this room. And it covers across these different gamuts. So, we have big companies, public companies, government labs, and else.
This is really like a team effort. I mean, largely led by my sister Joyce, who exceptionally flew from New York City where she moved three months ago. So, I'm like insanely indebted to her having done this for me. And so, thank you.
And yeah, my day job is investing in startups. I have many of my portfolio in the room, many friends that have helped over the years. So, thank you again for that, and hopefully we can form some new partnerships going forward as well.
So, I wanted to take a couple minutes to just set the stage for what we're going to talk about today. I used this slide last year to sort of explain how it felt like on a weekly basis we were making kind of incremental progress. You know, new model launches, new benchmark performance, new demos, etc. And then if you zoomed out five or ten years, you would kind of look at what we did and deem it to be magic. I actually looked at this slide this year, and I'm like, I think we're getting magic kind of on a weekly basis now. So I just left it here just for comparison's sake.
And so where did all this stuff kind of begin? I think if we go back 2015–2017, there were already early signs that deep learning would scale. So on the one hand we had examples of neural networks training a lot faster if you added more GPUs. This is from 1 to 128 GPUs. And then initial work from Baidu that showed that word error rates went down if you had larger training data.
And now these concepts are very well understood as scaling laws and they appear sort of everywhere. So first in natural language processing and then in biology with protein structure prediction and protein engineering and then even now in reinforcement learning.
And so where might all of this end up going if we see recursive superintelligence kind of emerging where now we have scaling laws across everything and maybe we're just limited by the amount of money we want to spend on experiments.
The interesting thing just a couple years ago was we doubted whether any of the unit economics in AI would actually even work. Now we start to see that we get better and better intelligence for less and less cost. And now the narrative in public markets and private markets is: are these companies, these big labs, going to be the last companies that ever will exist?
And just a year and a half ago we entirely as a market kind of went through this DeepSeek deep freak moment where there was a paper that came out that in an appendix said the experiments cost $5 million and the entirety of Wall Street read that and not the rest of the footnotes and argued that actually frontier AI could be produced for $5 million bucks. And so why were all these companies developing massive GPU clusters? Why were the R&D costs of labs so astronomically large if this was the sort of situation in the industry?
And this sort of accelerated the moniker of Jevons paradox, which is the idea of you get something for cheap, you use it more. The more you use it, the more you want it, the cheaper it gets and it kind of iterates over and over.
And so when we now fast forward another year and a half, we're basically on like full tilt infrastructure build out. Now the sums of money kind of going into GPU clusters is probably on the order of a trillion dollars. Last year it was about 450 billion. And every state in the world, every country in the world is developing their own sovereignty AI strategy.
And so, you know, we went from this kind of world where a couple of years ago we could barely get videos that would sort of barely get a model that could look at a video and describe an action of what's happening in there. You know, we would create many of these like task demonstrations and use these kind of convolutional models with like recurrence in there to try to predict actions and it was all kind of brittle and didn't really work. You know, now we have these like generative world models that we'll talk about today where every pixel is generated in real time. You can take action in these environments, they're entirely novel and they kind of give us entirely new experiences we couldn't imagine before.
So it's kind of crazy just 10 years ago we basically had some crappy version of Microsoft Paint and now you kind of can't tell the difference between reality and AI. Even just a year ago or two years ago, AI would struggle with basic arithmetic, basic maths. How many R's are in strawberry? A classic meme. And now we're, you know, basically getting mathematical challenges, grand challenges solved with AI. Interestingly, by putting two discrete areas of mathematics together.
And so again, back to this idea of recursive superintelligence, what happens when more and more of AI is developed by AI itself? This is from a blog post from Anthropic just a couple of days ago that really kind of shed light on just how much of Claude is written by Claude. Now, it's about eight times more commits are written by the AI than by the developer. And so, now we have super intelligent coders and indeed also super intelligent hackers.
Some great information that was analyzed from the AI Security Institute in the UK just showing how much of long-ranged kind of cybersecurity challenges could be solved with Mythos and GPT-5.5.
And so, then super intelligent hackers, super intelligent AI, AI can build AI — is robotics and frontier embodied AI the next sort of major scaling axis?
We've clearly gone from a world where China was deemed to be just a rip-off, not a real competitor. Three years ago in our State of AI report there was barely anything you could write actually about Chinese AI. And now, of course, it's at the frontier. It's kind of everywhere. Increasingly actually open-source AI looks like it's synonymous with Chinese AI.
We'll talk about a bit of this today, but we've been through a world where defense and national security was a faux pas in the technology world. No one wanted to touch it. Obviously, we were living in a world that was far less dangerous than the one it is today. And now, as I alluded to earlier, like AI is really involved in the geopolitics of every country in the world, whether it's for security, defense, industrialization, and everything else.
And so, you know, we've gone from this world of maybe AI will kill everybody to please use our apps. And yeah, more and more data center build-outs to do all of this. I think, you know, even 10 years ago we wouldn't have imagined we'd be in this position of investing a trillion dollars. But maybe there are some spanners in the works.
It's not a uniformly popular policy in the US and in Europe to go build data centers next to homes that you've invested money to repair and now you hear this constant buzz and higher electricity bills. And so about 14 states have seen legislation proposed to ban data center build-outs and a couple of them have issued moratoriums.
And so with all of that, I wanted to then take a step back and look through a couple of the achievements that many of the speakers have had over the last few years. On the one hand we'll have Vivek here today who spoke a couple of years ago. At that point, you know, Google DeepMind was working on Med-PaLM, Med-PaLM 2, sort of these models to work in clinical workflows. And now we've accelerated towards really amazing AI co-scientists. And then we'll hear more about that today.
Just two years ago we had Alex Kendall here from Wayve, the UK self-driving company that, you know, 10 years ago started in a little garage in Cambridge and now has accelerated to be a real contender with Waymo.
Also similar timing, we had Synthesia, again a London-based company born about 10 years ago, starting with video avatars and now as the world can increasingly generate pixels that are lifelike and interactive almost for free, this company has been accelerating to become that default.
And then in voice, ElevenLabs, again another hometown favorite, has just accelerated more and more, now passing about 500 million in revenue, soon 500 million. And then on the drug discovery side, there couldn't be one of the more exciting domains to work in and Isomorphic Labs, again hometown favorite, has accelerated to be one of the premier labs in the space.
And so overall the message I think I want to communicate here is like RISE is really about community, it's about trying to create opportunities for you across the spectrum, whether you're a student, whether you're working in a big company, or you're in a startup or looking to start one. And so, we have these community events and meetups around the world. We have open source AI fellowships that we give, and we have startup capital through Air Street Capital.
This is our flagship event. If you like it, there's a bunch of others that we run on a meetup format in the US and in Europe throughout the year. All of the revenue that we generate through this is funded to the nonprofit that we've had for a couple of years, where it has a goal to accelerate education and access and work in open source AI that might not be in the purview of private and public companies.
And then on the startup capital side, which is the day job, I really just love kind of meeting early-stage founders working in industries like biotech, defense, enterprise software, material science, etc. I've worked with many of you in the room, and I would love to work with more of you.
And I do, kind of weekly, ad hoc and sometimes annual basis, we publish essays on Air Street Press, which is really trying to showcase some of the achievements of the community, try to draw some playbooks to the extent that they're useful, and also comment on policy and geopolitics. And then on an annual basis, of course, produce the State of AI Report.
So, yeah, this event, we've had some amazing partners over the years. I think we have a really good track record of having our partners and they just stick with us for a long time. So, this is the way we like to work, and with that, you know, I'd like to thank Aaron Archer, who I've known for I think 12 years at this point. He was the first partner on RISE, and he's been an exceptional counsel to a lot of our companies and friends in the space. So, really encourage you to go chat with him if you have any, you know, advice you're looking for with regards to forming or growing your technology company.
And then a partner from last year that we're really excited about, Lambda, which is one of the best companies for building large-scale GPU clusters. I encourage you to go talk and chat to them. They're one of the favorites actually across the portfolio and have been a fantastic partner from little company to really big one.
And so with that, that kind of wraps up my intro. I hope you are excited as much as I am for today and we'll have a set of talks and then coffee breaks and lunch throughout the day. Most of the talks are about 20–25 minutes and then we'll have some questions afterwards for five minutes. So as you're going through the talk, use your neural notepad at the end of the program, like jot some notes and be prepared for questions as the talks finish. So with that, thank you very much for your attention.
