Guillaume Verdon
speaker
342 appearances
1 recordings
1 series
first heard Dec 2023
last heard Dec 2023
Guillaume Verdon’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
Yes, it is full stack. And so we're folks that have built differentiable programming into the quantum computing ecosystem with TensorFlow Quantum. One of my co-founders of TensorFlow Quantum is the CTO, Trevor McCourt. We have some of the best quantum computer architects, those that have designed IBM's and AWS's systems.
They've left quantum computing to help us build what we call actually a thermodynamic computer.
Right. I mean, that was a challenge to build, to invent, to build, and then to get to run on the real devices. Can you actually speak to what it is? Yeah. So TensorFlow Quantum was an attempt at, well, I mean, I guess we succeeded at combining deep learning or differentiable classical programming with quantum computing and turn quantum computing into...
or have types of programs that are differentiable in quantum computing. And Andrej Karpathy calls differentiable programming software 2.0. It's like gradient descent is a better programmer than you. The idea was that in the early days of quantum computing, you can only run short quantum programs. Which quantum programs should you run? Well, just let gradient descent find those programs instead.
We built the first infrastructure uh, to not only run differentiable quantum programs, but combine them as part of broader deep learning, uh, graphs, uh, incorporating deep neural networks, you know, the ones you know and love with what are called quantum neural networks. Um, and, uh, Ultimately, it was a very cross-disciplinary effort.
We had to invent all sorts of ways to differentiate, to back-propagate through the hybrid graph. But ultimately, it taught me that the way to program matter and to program physics is by differentiating through control parameters. If you have parameters that affect the physics of the system, and you can evaluate some loss function, you can optimize...
the system to accomplish a task, whatever that task may be. And that's a very sort of universal meta framework for how to program physics-based computers.
I think working across disciplinary boundaries is always a challenge, and you have to be extremely patient in teaching one another. I learned a lot of software engineering through the process. My colleagues learned a lot of quantum physics, and some learned machine learning through the process of building this system.
If you get some smart people that are passionate and trust each other in a room and you have a small team and you teach each other your specialties, suddenly you're kind of forming this sort of model soup of expertise and something special comes out of that, right? It's like combining genes, but for... your knowledge bases and sometimes special products come out of that.
And so I think like, even though it's very high friction initially to work in an interdisciplinary team, I think the product at the end of the day is worth it. And so learned a lot trying to bridge the gap there. And I mean, it's still a challenge to this day.
You know, we hire folks that have an AI background, folks that have a pure physics background, and somehow we have to make them talk to one another, right?
Yeah, it's really hard to pinpoint that je ne sais quoi, right?
Yeah, I'm actually French-Canadian. Oh, you are legitimately French-Canadian. I thought you were just doing that for the cred. No, no, I'm truly French-Canadian from Montreal. But yeah, essentially we look for people with very high fluid intelligence that aren't over-specialized because they're going to have to get out of their comfort zone.
They're going to have to incorporate concepts that they've never seen before and very quickly get comfortable with them, right? Or learn to work in a team. And so that's sort of what we look for when we hire. We can't hire people that are just like, you know, optimizing this subsystem for the past three or four years.
We need like really general sort of broader, uh, intelligence and specialty, uh, and people that are, that are open-minded really. Cause if you're pioneering a new approach from scratch, there, there is no textbook. There's no reference. It's just us and, and people that are hungry to learn. So we have to teach each other. We have to learn the literature.
We have to share knowledge bases, collaborate, uh, in order to push the boundary of knowledge further together. And so people that are used to just getting prescribed what to do at this stage, when you're at the pioneering stage, that's not necessarily who you want to hire.
I think our goal is to both run human-like AI or anthropomorphic AI.
I know it's triggering for you. We think that the future is actually physics-based AI combined with anthropomorphic. So you can imagine I have a sort of world modeling engine through physics-based AI. Physics-based AI is better at representing the world at all scales because it can be quantum mechanical, thermodynamic, deterministic, etc.
hybrid representations of the world, just like our world at different scales has different regimes of physics. If you inspire yourself from that in the ways you learn representations of nature, you can have much more accurate representations of nature. So you can have very accurate world models at all scales, right? And so you have the world modeling engine
And then you have the sort of anthropomorphic AI that is human-like. So you can have the science, the playground to test your ideas, and you can have a synthetic scientist. And to us, that joint system of a physics-based AI and an anthropomorphic AI is the closest thing to a fully general artificially intelligent system.
Showing 261–280 of 342 · page 14 of 18
← Previous
Next →