Google's AI Legend Just Left, OpenAI's Math Breakthrough & The Future of Robotics | Lightwork

0:00Wait, no, this is even crazier. Breaking news. Jeff Dean is leaving Google and starting his own company with Sanjay Guamette. We're entering a world where you can buy breakthroughs, right? One brain that you can plug into any robotics body. Is this what leads us to the dark factory? It feels like discovery is abundant now.

0:29we're back with another big week coming to you live from our san francisco office open ai pulls off some of the biggest math breakthroughs yet it solved 10 open problems including one untouched for nearly 30 years google deep mind releases gemini robotics 2 One AI brain that can run different robot bodies, walk, grip, and coordinate machines. And BasePower raised a billion dollars at a$13 billion valuation. And we discussed what they're building and the American-made home battery they launched. Plus two new deals from the Lightspeed universe. Let's get into it. So it sounds like there are big things happening in the math world.

1:14Yeah, OpenAI's next models solved 10 impossible to solve until now math problems. How did it do that? What should people take away from this big release from OpenAI? Yeah, so they basically teased their new model. It's unreleased. It's called Astra. And each of these produced new results for 10 long open problems in mathematics. And so these are not math problems where there is a textbook question and answer. These have been unproven. For one of these, actually, one of them had been open since 1999. So 30 years of people making no progress on this main question. And so these 10 problems were big.

1:57They actually had covered a wide slice of math from high dimensional geometry, coding theory, group theory, operator algebras, quantum complexity, cryptography, all this stuff that, you know, I do not have that math level. but we can talk about the breadth that it covers because I think it speaks to the power and the generalizable reasoning that these models have. But what was more important, not only that it proved it, but that it showed its work in three ways. So with this release, they published a 249 page write-up where they actually showed the actual math. They also had a walkthrough where the model explained its work.

2:39And then it actually ran every proof through a program called Lean, which is basically this automatic grader. And Lean is able to check those proofs line by line and only gives it a pass if every step folds up. And they publish all this publicly so anyone can check the work. But this is just machine checked, right? Like for solving these in the past, somebody would be getting a Nobel Prize, right? Like who gets the Nobel Prize in this instance? That's a really good question. I don't know. We are entering this sort of new paradigm where I think AI is starting to produce original knowledge. And it's not just summarizing existing patterns or pattern matching.

3:22I think that's historically been one of the critiques of AI is that these are kind of regurgitation machines. And, you know, I think a couple months ago, this is like late in 2025, there was this incident where OpenAI, an exec, had said that they had solved a couple of air dishes. They said they'd solved it, but they really hadn't. But this time they were a little smarter in the release, right? Like, I think from a PR perspective, this is like, here, we're putting this out here. We put together this big paper. We're not like, we're going to allow this to get human verified, maybe. But last time it was basically that it didn't actually produce new knowledge.

4:00It basically dug up existing answers. And that was where there were problems. Actually, this time, though, these were genuinely open problems that didn't have an answer and they were actually real dead ends. And so I do think this time it's different. And I think to your question around, you know, who gets credit for this or how this changes the space of math, I think a lot of very theory oriented spaces like math and science are kind of facing this reckoning where, you know, breakthroughs can be paid for even. I think someone brought up a really interesting point that this breakthrough was actually absurdly cheap.

4:35OpenAI said that. It was only like$2 ,000 in tokens to make this happen in the next model, which is Astra. Very exciting. But I don't know what that means for, you know, biology and all these other areas where discoveries used to be so scarce. But I know we're going to be talking a lot about abundance, but it feels like discovery is abundant now. Yes. And we're entering a world where you can buy breakthroughs, right? I think the fact that for$2 ,000, we had arguably 10 Nobel-level breakthroughs, that's only the floor. And you can imagine if we actually threw in real money or companies or universities throwing very real compute and money into these problems, the amount of breakthroughs we can have.

5:23That being said, I think there's an interesting tension between verifying and understanding. So there is a really interesting point. I forgot where exactly I had read it. But someone was laying out the idea that we have these results that are now true, but there might be a world where no humans actually understand it. There's a big difference between knowing the proof works, but not actually understanding how you'd reuse that insight or build on top of it. You just know that it's correct. Reasoning is now cheap, but the understanding and that depth is... The answer is now cheap. The answer is cheap.

6:01But the reasoning is not. And I think there's this question of like, is there a world where the answers and the knowledge output outruns human reasoning and our understanding of where we can, you know, we can prove things that's really exciting and we can prove what's true. But does that move faster than how we as a species can figure out why? It's very interesting because I think for a world like mathematics where it used to be such a hero's journey to go and solve these problems to see AI solving it for$2 ,000, it does mean that probably their knowledge is shifting to exactly where you're saying, right?

6:42It's going to be a little bit more in this judgment and explaining how we got there. Yeah, I think there is, you know, with everything, there is going to be both a realm of verifying and realm of understanding. And I see maybe that shift. That being said, I am nowhere close to being a mathematician and could not understand what it's like to deal with AI tools in this space. But I do think it's a really interesting moment. I think it is one of these spaces where, you know, in knowledge work, so many things are unverifiable. We have these benchmarks that are sort of replacements for how good AI is at solving our work or accomplishing things.

7:22But it isn't as binary as in math, where there are these just like true or false answers. And so it's really interesting to see how AI has massively disrupted a field like that. I think it'll continue to evolve. And I'm very curious to see what kind of breakthroughs come out of this. I'm curious to see reactions from actual mathematicians. But for now, as a very clear outsider, it's just very fascinating to watch from from afar. Claire, Google announced Gemini Robotics 2, which is very exciting. Is this their Android moment for robotics? robotics big breakthrough moment a lot of people are saying this is sort of like the chat gpt moment but for robotics and the way i would analogize this is in the realm of llms and these large language models for a very long time we had these classifiers that were very specific to specific workflows or specific patterns that we were trying to identify but then there was this moment where we were able to build this very generalizable model that could fit into many different workflows.

8:26And I think that's what they're trying to accomplish here with Gemini Robotics 2, which is their second major generation in its robotics AI models. Actually, with this, there are three models. And what it's able to accomplish is build this sort of one brain that you can plug into any robotics body. That's how I would describe the big mission of Gemini Robotics 2. Yeah, I think the cool thing, if you watch the video, was that one robot, well, you know, the standing robot, could speak to the robot that perhaps didn't have legs, and they can communicate and kind of hand off tasks to each other. And I thought that felt like the really special thing, because they highlighted a lot of things that we've talked about on Lightwork, right?

9:08They talked about the dexterity of the hand, which we've seen with 1X, And, you know, we've certainly seen it with other robotics companies, but it did feel like they're creating this layer. And the other thing is they're also opening this up to developers, right? Yes. I don't know which exact models, and I'll go into the three models in this family and how they all work. I don't know which exact ones are open to developers. But to your point, I think the big difference here is that historically, a lot of these models in robotics have been pre-programmed and specific to one body. So they're very good at repeating one very narrow motion, like picking this one box up, putting it in this one shelf.

9:55And, you know, if it's repeatable and it's in a clean environment, it can do that many, many times. But what we haven't nailed is something that is able to learn and adapt to new environments. And one thing in particular is that a skill trained for one robot in this physical form almost never translates to another physical form. So the moment you change the amount of grippers or the moment you change the height or the moment you change whether it is wheel based or bipedal, then you totally have to reconfigure the model and change the mechanics of the brain. Right. And so this is actually trying to just create a brain that can plug into any body and plug that intelligence across all forms.

10:35within the models there's actually three things here one there's gemini robotics 2 and that's what people call a vision language action model in the robotics case a lot of people reference to these as vlas and that's basically the model that is taking in what the robot is able to see in front of it and then take in language instruction and convert that into actual motor commands so i would say this model, Gemini Robotics 2, is like the doing model. As part of that, and an important distinction is that they have a separate model called Gemini Robotics ER2. And this is an embodied reasoning model, meaning that it is a vision language model that acts more like the brain.

11:20So it is not in charge of turning what's happening in the outside world and turning all these physical inputs into commands. Rather, it is just the brain that's processing, here's the order of steps I need to work through. Here's what this item is. It doesn't move any motors directly and basically looks at a room, figures out the sequence of steps, talks to, you know, the human and might say, should I do this this way? Should I pick it up? How long would you like this task to be completed? And then it delegates the movements to the VLA, the vision language action model. You can actually see how both of these work together.

11:56You kind of need both. You need both the brain and the motion part. And this is what it solves for. And then there's a third one, and that one's just Gemini Robotics on device too. And that's basically all you need to know is a baby version of the larger Gemini Robotics 2 model that is run locally on the robot itself. So you actually don't need internet connection and you reduce some of that latency from the cloud. So that's how you get robotics that don't need to constantly be plugged into internet. That's how you get robotics working in more rural or more difficult environments. That's sort of the breakthrough there.

12:32How does this differ from maybe skilled AI in our portfolio? Do you want to first share what skilled AI is? Skilled AI is a robotics company in the Lightspeed portfolio. They focus on creating the brain side of robots. The brain works, whether it's a biped robot, whether it's a robot with wheels, it can basically take on any form and start to use that robot. So I think that it's different, but there are some similarities here. Yeah, I'm really glad you brought it up because, and now that everybody knows what Skilled.ai is, they should look it up because it's a really cool company. But they are playing on a similar idea of the portable brain.

13:15And so this portable brain idea isn't unique to DeepMind. I think a lot of people are going after this. Guild, I would say, is specifically also focused on this idea of some building for the omnibodied robotics ecosystem. Yeah, it's that portability. Yes, yeah. So that you have this one model that works regardless of the shape without being told any specifics of that body in advance. And I think their slogan is actually any robot, any task, one brain. So I think also what's interesting is that it actually is not only applicable to humanoids. It works for robots that come in the form of just two arms, factory arms, quadrupeds even.

13:57So I think really interesting work there. And the adaptability, right? Like when you cut off one of the robot legs, it can learn to walk without it, right? Oh, I actually didn't know that. They have so many cool features that are baked into that model. But these are actually designed very much for the robot, it seems like they're working on today, right? Yeah, that's interesting. I guess I'll have to go look at the adaptability because that's really cool. Because I think one of the big things of robotics today is the fact that we are seeing a lot of these cool demos, but they often happen in these very different ways.

14:30in these vacuums where it's like, oh, that would be really cool. And then it's like, what happens when I bring it into my home or the office or all these complicated places where humans get in the way? Yeah. And I think at least with Gemini Robotics too, the way that they're going after these and capability wise, where they're seeing advances is whole body control. So this idea that you no longer have to segment apart what's happening in the top half of the body and the Lower half the body now you actually have what they call whole body control from feet to fingertips, meaning that coordination, balance, locomotion is all synced together.

15:07And that's actually really hard because historically those have been separated because balancing is a very different motion versus picking something up. But now you can actually have something walk in real time, balance. And based off of how the arm shifts, that will inform how it needs to correct on the leg part. So that's interesting, having it all sit in the same neural network that is taking in all that information. On the dexterity front, to your point around X1, this one is pretty close. I know a couple weeks ago we discussed how X1 has 25 degrees of freedom on its robotic hand. This one actually has 22.

15:43So it's able to do things like tying knots and sealing a Ziploc bag. Interestingly, it can unscrew a light bulb reliably 92 % of the time. However, screwing one back in lands at only around 36%. So you can see how it's good at certain motions, but fails once you bring in a second step or something pretty similar, but in reverse. There's all these kind of edge cases and failure cases. It's so interesting. It's hard to fathom how some of these tasks that are so easy for us are so hard for robots. One of the things that I watched it do was try and take out the trash. You just think that that's like such a easy human task.

16:21But for a robot, it's a very complicated process to understand the dexterity and the weight and all of these things that, you know, we've been talking about with these incredible groundbreaking robotics companies. Another interesting part is that sometimes you actually need multiple robots to accomplish one task. That feels so easy for us as humans. Another capability they introduce with this is called multi-robot collaboration. And what that means is not that you have one AI brain that controls a swarm of bodies. That is, we don't have that yet. Is this what leads us to the dark factory? We don't have that yet.

16:58What this is, is that there's a robot robots of different types can now coordinate to finish a job faster. So you can have two robots, for example, tidy a room. One robot is, for example, taller and it's in charge of opening the fridge and pulling out ingredients and condiments and setting them on a counter. And then you might have a smaller fixed arm robot that doesn't look like a humanoid. It's, you know, stationary on the counter and it's more precise, but it can't move around. And that's focused on stacking, you know, assembling, slicing maybe. And you can see how this multi-robot collaboration is important.

17:33Again, it's not one brain, but what it is, is that you have robots that can sort of like hand off tasks. The way we think about robotics today is, you know, we like to imprint our view of like human work and it needing to come in the form of the way that humans complete tasks. And to your point, a lot of this is harder than we think. And maybe how we accomplish it is by breaking it down into all of these kind of smaller, more systemic form factors. Yeah. Seems like robots will do it differently. Okay. So we have a developing story. Breaking news. Yes. Jeff Dean is leaving Google and starting his own company with Sanjay Guamette.

18:13Big news. Do you want to share a little bit more about Jeff Dean? Because he is a really important character in all things AI and deep mind. I mean, Jeff Dean is one of the godfathers of AI and this new company discovery loop, like the, the entire team has worked on Google search, Google ads, Google news translate. They've worked on alpha star and alpha fold and Gemini and all of these different incredible achievements. And now they're starting to work on these loops, research and engineering tasks that are difficult to do, is the focus of this new company. Yeah, it's huge. You know, I think it's funny because I think it was a few weeks ago or last week, Jeff was actually a speaker at YC Startup School to 6 ,000 people.

19:09And he talked a lot about what he was interested in. And one of the things he said was his big interest in this idea of the automated version of the scientific method. And he verbatim said, you propose an experiment, you implement what you need to run the experiment, you evaluate the experiment, and then you get the results from that. So ballsy to tease your new startup on stage at a YC event before announcing. I know. Little did anyone know that he was foreshadowing it all. And he wants to implement all these automated loops in many domains across science, chip design, biology, even making AI models smarter, which is kind of what we've discussed a lot around this idea of recursive intelligence, iterative intelligence, the idea that models are now self-improving, which is, you know, an entirely new paradigm shift.

20:00and even ties a little bit back to what we were discussing around the math breakthrough with Astra. Yeah, it feels like now all of this, these discoveries are now in the realm of what LLMs can solve. We've seen it with Astra. And so it'll be interesting to see what they solve. And I feel like they have this like big mission on their website, like basically to solve all sorts of problems from solar to clean drinking water. It's kind of everything. Yeah. When I saw the breaking news, I was actually I I have to be honest, I had a little freak out because the first headline I saw was that Google DeepMind CEO Demis Hassabis is stepping down.

20:43And I, you know, for me, Demis is such an iconic figure that I thought he was leaving DeepMind. I was like, there's no way. Yeah. But it turns out this is just more of a restructuring now that Jeff Dean who was formerly chief scientist and a bunch of other Google AI executives are leaving to start their own company. Now Demis is going to shift from CEO to be the unit's chairman and also chief scientist for Google. So it's a big deal. I mean, on X, it seemed like the CEO of Google wrote a very nice message saying that they are going to both be an investor and the cloud provider for the company, which is lovely and exciting to be backing their own.

21:24Looks like it'll also be a PBC, a public benefit corporation. So interesting to see. I think this is similar to a lot of labs that we've seen where we're still early, early in seeing how those breakthroughs in research actually diffuse into actual applications. Big news from a Lightspeed portfolio company. Base Power raised a$1 billion Series D at $13 billion valuation. Claire, talk to us about Base Power. What makes this company so special and maybe different than the way people are thinking about powering their homes today? One of the fastest ramps in the history of energy hardware and any startup building in the energy space.

22:08For context, in under three years, this company has raised more than$2.5 billion in total, and they've tripled their valuation in 10 months since the last round. So we're excited to be partners with them. They are building in a category that we know moves so slowly because it involves atoms and factories and regulators around energy. But at its core, Base Power is not just a battery power company that is competing with, you know, a Tesla Powerwall. It does sell batteries, but the bigger vision here is actually the network that it has, the network of batteries that it's built and it actually performing more as an electricity company that happens to use these home batteries as its infrastructure.

22:53So a lot of people actually call them a gen tailor. They are a combination of a power generator and a retail electricity provider. How it works is that base power will give you as a consumer a really big battery for your home, similar to what a Tesla Powerwall might provide. Sure. But like a Tesla Powerwall and some of these other things are quite expensive. They're like, that's a five figure purchase. This is not like a small purchase. Yeah, they can be anywhere from$15 ,000 to$20 ,000. Base power is a lot cheaper. Yes. Here you don't buy it. And they help you install it. They own the Powerwall.

23:34And you just pay this small setup fee. I believe it is as low as anywhere from$95 to$695, depending on where you live. And then you pay a low monthly subscription fee and you agree to buy your electricity from them. And this has become a really big need, especially in the markets that they're serving, right? In like Austin, Dallas, you think about like they have these like snowpocalypses where you lose power and it's it is actually so necessary to have a power wall. And yet they're so expensive. So this actually makes a lot of sense from a consumer standpoint, right? Like thinking that, you know, instead of buying this really expensive thing, I can partner with BasePower, pay this little bit of money and add myself to this much larger grid.

24:19So what it means for BasePower is that BasePower has all of these suburban homes with all this excess power on these power walls that it can then resell to others that require power. Basically, this can actually almost serve as like its own data center. If you think about like all of these different providers thinking about building a data center, there's all this excess power in the ecosystem. I think the whole business runs on this idea that you can, to your point, connect thousands of these little home batteries together and create the sort of what they call a VPP, which is a virtual power plant.

24:57So instead of having to have this one giant power plant, which we know that when a state needs more backup power, that requires massive battery insulation. You basically put a warehouse full of batteries out in the desert or a countryside. And those are huge. They hold a ton of energy, but they take a lot of years to build out. You have to buy land or permits, a lot of environmental regulation to connect it to the grid. So there's just a lot of slow bureaucracy. Instead, to what you brought up, this kind of takes advantage of that decentralization and you have thousands of these homes and nodes that connect together.

25:36They networked together. They add up to a utility. And that's the exciting part about this. Yes. And then what base power is able to do is when you are not using these, they actually make money off of this connected electric grid. So for reference, electricity prices usually swing very wildly during the day. Power is often very cheap at night. Base will charge up all their batteries cheaply then. And then when the grid is slammed, which is, you know, say a very hot summer afternoon in Texas, there are a lot of those. When everyone is blasting their AC, that's actually when prices spike a ton in the grid.

26:18So they sell that stored power that they've been collecting from all these kind of decentralized homes to the grid at a huge markup. So this is basically power, in short, at buying low and selling high, and you're doing it kind of at scale thousands of times a day across thousands of homes. Very exciting. And truly, it's like the most interesting time in the energy space right now. I think, you know, if folks haven't seen our great breakdown with Guru on what's happening in the data center stack, there's just a lot happening when it comes to American electricity demand. And it is actually rising for the first time in a generation, largely driven by electric vehicles, refurbate manufacturing, and obviously data center demand.

27:05And so to what I brought up earlier, this idea that instead of traditionally building new gas plants or these new transmission lines, we need capacity now. And this is a really interesting, unique solve to it. Yeah. And there's so much more untapped potential in all of these homes across America.

27:28K2 Space raised a$500 million Series D at a$6.8 billion valuation. K2 builds the biggest, highest-power satellites flying. Its first giant satellite just marked 120 days on orbit, and the company holds more than a billion dollars in signed commercial and government contracts. Their Torrance factory is built to scale towards 100 satellites a year. Just two years ago, the industry consensus said that satellites should keep getting smaller, but K2 is betting the other way. Whatnot raised a$547 million Series G at a$20 billion valuation in a round co-led by Lightspeed. It's a live shopping marketplace where sellers stream and buyers bid in real time.

28:20Today, you have businesses of all sizes, from small businesses to retailers and global brands, that are all making live commerce a core part of their business. Whatnot says that over the past year, the number of sellers reaching more than a million dollars in lifetime sales has now more than doubled. And this new capital will go towards AI-powered tools for both buyers and sellers.

28:51Thank you.