#259 Anjney Midha: a16z’s Strategy to Turn AI Startups into Unicorns

0:00We're entering a regime where the model is taking action versus just producing words. we're starting to get into an agentic system. The more autonomy you give that model, the ability to call those tools on its own and self-learn, the more agentic I guess it is. But we are squarely already in the middle of agents. I would say a lot of code workflows is probably where I'm most excited about agents actually working. The last few years have been the unabashed growth of transformer pre-training scaling. And that meant for a long time, people thought that meant that general models would win. And I think actually now we're transitioning to an era where general models with specific last miles are winning.

0:33Building multi-agent software is hard. Agent-to-agent and agent-to-tool communication is still the wild west. How do you achieve accuracy and consistency in non-deterministic agentic apps? That's where agency, A-G-N-T-C-Y, comes in. The agency is an open source collective building the internet of agents. And what's the internet of agents? It's a collaboration layer where AI agents can communicate, discover each other, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows.

1:26Build with other engineers who care about high-quality multi-agent software. Visit agency.org and add your support. That's A-G-N-T-C-Y dot O-R-G. Visit them today to support high-quality multi-agent software. So yeah, Anjani, can you talk about your background before a 16 z and uh and and about the anthropic angel investment that's that's pretty fascinating but obviously you must have done quite a bit before that to be an angel investor uh i i got my start a long time before that in machine learning as a graduate student at stanford i was in the bioinformatics department there and this was when there was an earlier wave of machine learning around 2011 and 12, where everybody in Silicon Valley thought they needed a machine learning strategy.

2:25But I was mostly on my way to a career in academia. I was really fascinated by using some of the new techniques that were becoming possible in deep learning in healthcare. And being at Stanford at the time in the med school meant that I got access to tons and tons of really fantastic, really valuable data, sort of clinical data, healthcare data, patient data. And I thought I was going to go be an academic when I got a chance to spend a summer at a firm called Kleiner Perkins, basically helping their portfolio companies get started with their machine learning pipelines at a time when everybody in Silicon Valley thought they needed it.

3:09So one of the partners there said, look, we need somebody to be our first machine learning engineer. And that's kind of how I first learned about venture capital. I ended up spending four and a half years there at Kleiner as an investor. They ended up asking me to come over from the engineering side to investing, both at the early stage and the growth stages in what now we would call developer infrastructure businesses. But at the time, it was basically companies that were figuring out how to build new products and services on top of GPUs that had only recently become available in the cloud, on places like AWS and so on.

3:47And then around 2017 is when I decided to leave and start my own company called Ubiquity 6. And our goal was to try to use some of these modern deep learning techniques, primarily in computer vision, to allow high precision 3D mapping for applications like indoor robotics and self-driving cars. At the time, there was this explosion in computer vision AI techniques that made it very possible to do high-precision 3D reconstruction on commodity hardware like smartphones. It turns out that there's a little app that took over the world called Pokemon Go that you might remember around that time. And that ended up being one of our primary sources of customer demand was game developers trying to build applications, augmented reality applications in real world spaces.

4:43And we ended up serving millions of users who were walking around the world mapping real world locations for us. And one of the biggest bottlenecks that we had to solve for them was massively multiplayer networking, which meant when you had hundreds of people in one physical location, how do you keep everybody's physical coordinates updated in real time? It's kind of a hard networking problem to update everybody's positions in real time. And so we built a fair amount of interesting technology to allow that. And it turns out when the pandemic happened a couple of years later, and everybody was online, not walking around in the real world, but doing a bunch of multiplayer experiences online, that infrastructure became really valuable because we had built essentially a form of networking called serverless networking, which meant it was really easy for thousands and thousands of people to get together in the same sort of virtual experience very quickly without it costing the developer a ton of capital.

5:42And so I ended up selling that business to Discord at the end of 2020, which had also exploded in the pandemic from being primarily a chat application for gamers to being used by all kinds of groups. And there, my job was to both integrate our backend infrastructure into the company and launch the company's developer platform business. They wanted to allow developers to build third-party applications on top of that infra we'd built. And this was around the time when I got the call from friends who were running research at OpenAI. Tom Brown, who was one of the leads on GPT-3 and had been a longtime personal friend, gave me a call and said, Anj, we're thinking about leaving and starting this new company.

6:26We'd love your help on figuring out how to fundraise and do all the foundry things you've done as a CEO and founder who's sold his company. And so for the first six months of 2021, Dario, Tom, and I just did many, many working sessions on business plan, capital fundraising, how to assemble sort of the commercial side of Anthropic. topic. And then, you know, a few months later is when I was I came on as their angel investor and help them then go raise the first hundred million dollar seed round. And then later, we we got fairly involved in acquiring lots of compute for the company. It was a it was sort of net.

7:09Today, it's well understood that foundation model businesses need hundreds of millions of dollars in compute to get going. But at the time it was a fairly novel thing. And around the end of that year, 2021, is when I started realizing that the scaling laws and language models that we were seeing at Anthropic were going to hold for other modalities like image and video. And I had the chance to, in building the developer platform business at Discord, one of the things you always have to do is figure out what are the early killer apps that you think will drive value right on the platform. And I had the chance to team up with an old friend who was working on a little text to image app called Midjourney.

7:49And we created an app store front page at Discord where we put Midjourney on the front page and helped them launch one of the world's first text to image models. And the Midjourney team did an extraordinary job taking that business from basically zero to many tens of millions of dollars in revenue very quickly. And that kind of gave me the inkling that we were in the middle of a new platform ship fundamentally where entirely new startups could be built in new modalities, whether that was video or audio or code. And I left then Discord in early 2023 to start investing in companies full time and helping researchers and scientists take their research out of the lab and turn them into new foundation model infrastructure businesses.

8:37And since then, you know, I've had the chance to work with some amazing folks like the founders of Mistral and who are working on open source language models. Robin Rombach and team who are the creators of Stable Diffusion and then founder a company called Black Forest Labs that I'm on the board of. And that that that's why I that's what I spend most of my days doing is working with researchers and scientists, turning their research into into businesses. That's why I joined Andreessen Horowitz shortly after I left Discord to do that full time. And that's what I spend most of my days doing now.

9:11Yeah. Wow. Going back to the Pokemon Go, you weren't involved in building that, were you? No, no, no. I was not at Niantic, but they were one of the developers who wanted to use our technology. And early on, actually, John Henke gave me a call and said, this is exactly what we've been looking for. How quickly can you get the service up and running? So I really owe it to him for, you know, gaming was not at all on our roadmap. Indoor robotics, indoor navigation was kind of our primary focus. And I credit Niantic entirely for opening our eyes to the world of location-based gaming. Yeah. And with then from Anthropic, you stayed focused on foundation models.

9:57I mean, at least for a time. And just maybe to jump ahead a little bit, what I'm interested in hearing is how you see, I mean, we're all familiar at this point with foundation models and they raise some issues that I'd like to hear your thoughts on, you know, the open source versus proprietary debate. of and and then you know distillation and small models models for the edge but recently i've been talking to people about you know kind of backwaters in ai research that have been passed by uh during the transformer uh you know phase but when you started you were talking about uh vision right uh you know that was a super supervised learning phase and that's all anybody talked about and everyone was you know drilling down and in these incremental improvements and then transformers came along now everyone's you know working on how to uh you know tweak foundation models to do this or that or uh but there are these other strategies You know, I had on the podcast a bit ago, Sep Hockrider, you know, that's not how the Germans pronounce it, but, you know, the guy who came up with LSTM.

11:36Right. And he's, you know, still pursuing that architecture and doing really interesting things with it and things. There's a lot of promise. reinforcement learning was kind of in the background for a long time and and it's it's being talked about again because it's being used to train foundation models but you talk to the guys up in alberta and and they're doing amazing things and have great optimism about building intelligence using rl alone or at least as a primary strategy. And then I was talking to yesterday this guy, Friston, with the free energy principle and energy-based models. And there's all kinds of stuff going on there.

12:35You know, there's Mamba now, and that's opened up. You know, I had a call with, I'm not going to be a member. I'll cut that out. But anyways, the guys who have productized that in a model called Jamba. Right. So, I mean, are you focused on sort of transformer-based models, or are you scanning the horizon for these other strategies to see which ones might become important in the product space? No, look, I'm constantly looking for new architectures because the idea that a one-size-fits-all architecture just works is deeply counterintuitive, right? Yeah. And this is why I think it's called a bitter lesson, which is that if you get your training as a machine learning, in sort of classical machine learning, what you learn is that there should be specialized architectures.

13:42If you value efficiency, that there should be a more efficient architecture for every different use case or task. That task-specific architectures historically have been more efficient when it comes to production, right? Yeah. And that's just turned out to be unintuitively, frustratingly untrue, right? Like the transformer has just marched ahead. And I think what's going on is that the bitter lesson is just turning out to be true because Moore's law continues being one of the biggest drivers of efficiency of computational workflows. And what we're learning is that, you know, to be honest, these transformer models are so efficient at extracting sort of latent representations of knowledge that in a sense, the architecture, you don't have to get that fancy on architecture if you can process and filter the data correctly.

14:37Now, the big question I think is, the right one you're asking is, where's the limit, right? of how far do transformers get us and I I have been shocked basically by how good RL is at improving performance um let me put it more more clearly what I I do think that we were the last few years have been the unabashed growth of of transformer pre-training scaling Yeah. And that meant for a long time, people thought that meant that general models would win. And I think actually now we're transitioning to an era where general models with specific

15:28last miles are winning. That last mile can often be a combination of an LSTM, an SSM, a completely different non-transformer based approach. And in some cases, they're diffusion models. right so if you if you take the case of multimodal image generation um i think we're heading to an era of transfusion where you take a transformer which is autoregressive and then you combine it with some of the best parts of diffusion so that you get quality um in image generation and that's where we're at we i do think we're headed to an era of hybrid architectures i'm not a transformer maximalist but i do think the network effects of of infrastructure where a lot of the world's like sort of inference serving pipelines are now designed to be for auto aggressive you know work makes it makes the bar higher and higher for a new architecture to to try to jump over before it gets adopted yeah yeah yeah uh but you're hopeful or or you anticipate that there will be a new architecture I mean no one saw transformers coming um you know this is where depending on who you ask if if you ask a traditional machine learning researcher they'll tell you there are no new architectures they've all been invented before you know and a hawk rider will tell you that so will you know the in the entire there's a whole body of of sort of european researchers who will tell you that the transformers is basically just you know sequence to sequence learning which was invented 40 years ago you know with with boltzmann machines right yeah um and so some i think i think there's two sides of the argument one side one school will tell you no way craig there there there are no new architectures we've all invented them before this is just the new kids on the block inventing stuff we've done 40 years ago um and then there'll be the other school of thought which is which is often non-traditional researchers you know the the anthropic guys actually came from physics backgrounds you know dario has a phd in biophysics um jared kaplan who's a chief scientist and you know was was sam did the formative sort of seminal work around scaling laws was a physicist by training not a computer scientist and they'll tell you of course these are new architectures and of course we'll have you know sort of this is fundamentally a new body of work and we might have you know new new um representation learning mechanisms in the future and i to be honest with you my where i fall in the debate is does it even matter right because the thing that continues is sort of being the most important is that they just, they work and they're driving such incredible advancements in the end user's experience that none of us expected that it doesn't really matter to me very much.

18:10Even though my training is in machine learning, I have learned to let go of the ML purist in me and stop worrying about the architecture and asking, you know, do we have enough training data for the models to keep improving? yeah um so what what kind of things you're looking at now that that are sort of beyond their horizon look the the number one thing that keeps me up at night right at night right now is evaluation right how do you actually tell how good these models are we're we're past well past the era of the low-hanging fruit right where when gpt3 first came out and you know daria and tom sent me a screenshot of of their ablations that showed just by increasing the compute by 100x on on a bunch of tasks like um you know they had they had a chart showing how well a human could detect if a piece of if a news if a news article had been written written by an ai model or not and just by 100xing the compute they basically crossed the turing test on on that path yeah right and there were these low-hanging fruits sort of tasks all around us two years ago four years ago actually at this point where in coding for example just by 100xing the compute gpd3 relative to gpd2 was able to solve programming problems that had never seen before in the training data right there was no even if there was no python code in the training data and there's only javascript it was able to still solve python problems during inference time which is what what they'd call in-context learning.

19:45I think now we're headed to the era where those academic benchmarks like MMLU and so on are no longer really, we have moved from the research era to the deployment era. And for the deployment era, you need benchmarks and evaluation to tell you how good these models are in the real world. And to do that, you need a completely different mindset and methodology than I think the previous era, which is you'd have these sort of static benchmarks. benchmarks, you'd kind of train a model to beat those benchmarks, then you present at a research conference. And that was good for the low hanging fruit era.

20:20Now we're in the era where if you want to know if a new model actually solves a physician's problem in the field or a scientist's problem in the field, well, you need these models to be trained in that real world evaluation loop. And that's what I'm looking for right now. Yeah. Yeah, and benchmarks from a layman's point of view, you can train to a benchmark. And I don't necessarily trust these companies to be doing it correctly. Exactly. So everyone comes out with their chart and their model is like the middle finger pointing up there against everybody else. But who knows how they got around to that.

21:07Well, Goodhart's law is real, right? Which is that when a measure starts becoming a target, it ceases to become a good measure, which is exactly what happened with all these static academic benchmarks like MMLU and so on. So now I've become much more partial to crowdsource the wisdom of the crowd, things like LMSIS, right? Where you have real world users show up and rank two side-by-side responses from LLMs. And that's much, much harder to try to sort of, I think it's much harder for that to be overfitted on. Now, you could debate that you can still overfit on that, sure, but it gets harder and harder to do that at scale.

21:47It gets harder and harder to do that in real world tasks. And therefore, I'm much more a fan of sort of that type of evaluation to solve. Look, the big problem to answer here is AI reliability, right? Yeah. The biggest problem holding back AI models from being useful, I would say, in the most mission critical industries of life, you know, defense, health care, financial services is reliability. And I'm just very convinced that the answer there lies in good, robust measurement. right and and that means evaluation with real world with real users in real world context because otherwise the measurements from sort of academic setting are not reliable benchmarks or measures of performance ultimate yeah uh are you looking at uh at any of these you mentioned uh LSTM are you you know the Huck Ryder am I saying it right has a new company I'm not going to be able to remember uh NXAI or something like that where he's trying to productize his the advances he's made on LSTM architecture I I haven't talked to him about the company I'm not familiar with what their particular plan is but i would say there's a number of companies in that vein who believe that specialized non-transformer architectures are more reliable because they're more interpretable in the real world right yeah as opposed to uh trans the transformers a black box right and so there's only two solutions i think either you have an you you have an architectural change that um makes these black boxes transparent right and the body of work they're called broadly speaking interpretability that that i think is fantastic i spend a lot of time paying attention to what's happening in interpretability there's actually a couple of great papers that came out last week from anthropic that show that we're beginning to to to peer into what what's going on in the transformer um so that's either either you've turned the black box into a clear box or the second approach is you start measuring the output of the models to be sufficiently predictable that they're reliable.

24:15And I'm kind of bullish on both. I don't think you need an architectural shift like LSTMs to solve that. I think if you could make transformers more interpretable, you could get the same benefit. Yeah. You're talking about Anthropik's work on I can't remember now, intelligence graphs or what did they call it? But they've developed a tool to kind of like a functional MRI to be able to look in and see which nodes in a network are being activated at different stages of it. Exactly. We call it circuit tracing, right? Right. And the idea is that you can group the parts of a neural network that are activated when you ask it something into these clusters they call features.

25:05And seeing which clusters of these features fire for a particular topic, you can start to trace why a language model says what it does. And you're exactly right. The analogy would be inventing a microscope for what's going on in the LLM. once you have a microscope then you can move on to the part the next step which is okay can we start actually gene editing if you don't have a microscope to look at the genes then you can't actually start editing but once you have a microscope then you can start talking about how to actually do steering and and and control yeah and it seems to me that they they were able to do that in those experiments they the i'm thinking there was uh one of the examples was they started a poem with a word that rhymes with rabbit.

25:56I can't remember. No, that's exactly right. That's right. Yeah, and they could see that the model was looking at habit and, or no, it was carrot, or habit and carrot, and they could sort of direct the model toward one or the other. Exactly. I think what they identified was that the models plan ahead for where they're going to go. And if you can kind of trace the circuits that the model uses to plan ahead, and you can block out or mask certain endpoints of those traces, then you can steer it towards other endpoints. And so the idea was that they could get it to write a different second line of the poem than it was originally planning to by kind of masking, you know, the original plan.

26:46And I think it's very interesting. I think we're starting to get, it's finally starting to get with interpretability going from research to engineering, right? Because it's obviously a toy example, but this is how we do engineering, right? we you first try to prove it in a petri dish and then if it works then you start scaling it up and so i i you know my my understanding is the next step now is to try to scale it up to prove that they can actually steer it in a useful enough fashion outside of those toy examples it's very exciting but up until until that happens and it may be yours i'm a big fan of evaluation uh in real world scenarios like you know lmsys and chat arena i think that move is is is positive i do think

27:34um one one of the core tensions in reliability in in in llms is control over the weights and which is why we're entering an era where geopolitics is obviously pretty important lots of sovereign nations now want their own ai stack they want their own models because ultimately unless the models are fully interpretable there's no other real way to control them unless you have the weights. I guess I would say there's three big solutions to AI reliability. One is mechanistic interpretability or just making the models more interpretable in general. You could have side-by-side evaluation, things like Arena, or you could have open-source, open-weight models that give full control over the inference to the customer running the inference.

28:20I spend my time thinking about all three of those things right now. yeah well what about agents that's the other hot area that that everyone's talking i would think that you guys would be all over that or is that someone else at the firm oh i think that everybody's got to contend with models going from being next word prediction machines to next action prediction machines right and the beauty of the bitter or the the bitter lesson playing out with with transformers is that it's, it doesn't, it's not that hard to fine tune a model to take action on what, you know, would be called tool calling, right?

28:59So the first kind of generation of agents we're seeing right now is just LLMs trained to call APIs when they need help, right? And so you can tell an LLM, hey, I need, tell me the time today. And if you don't know what the time is, then that's fine. Just call out to a clock. Now, some people will tell you that's not really an agent but i i i don't i'm not that dogmatic about these definitions to me basically if if we're entering a regime where the model is taking action versus just producing words we're starting to get into an agentic system um the more autonomy you give that model the ability to call those tools on its own and self-learn the more agentic i guess it is um but we are squarely already in the middle of agents i would say you know a lot of um code workflows is probably where i'm i'm i'm most excited about agents actually working you know the reality is for things like browsing general purpose agents are very brittle right yeah but in workflows like coding uh where the outcome's actually quite easy to verify and score you have we have unit tests right in software so that's why in software engineering and code generation it's quite easy for agents to actually be told whether be given a scorecard.

30:13And I think if your question is, how do you measure AI reliability with agents? I think it's very simple. The general heuristic is exactly how you would score a human being on that task. You write a bunch of tests and you figure out whether it passes those evaluations or not. And the good thing about programming and software is that a lot of those tests are automated by design. Building multi-agent software is hard. Agent-to-agent and agent-to-tool communication is still the Wild West. How do you achieve accuracy and consistency in non-deterministic agentic apps? That's where agency, A-G-N-T-C-Y, comes in.

30:59The agency is an open source collective building the Internet of Agents. And what's the Internet of Agents? It's a collaboration layer where AI agents can communicate, discover each other, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for interagent communication, and modular components to compose and scale multi-agent workflows. Build with other engineers who care about high-quality multi-agent software. Visit agency.org and add your support. That's A-G-N-T-C-Y dot O-R-G. Visit them today to support high-quality multi-agent software.

31:51Yeah. Two questions for someone in your position. One, how do you keep your finger on all of this? Are you like me? You subscribe to a million newsletters. You have a million meetings. uh and you're capturing the conversations uh with some AI like otter I use otter other people use other things and then you're kind of constantly uh you know summarizing and looking through for ideas or or are you relying on on deal flow and what people are coming to you with I think so the answer is yes, all of the above. But there's just no substitute, I find, for deep, long-form conversations with scientists.

32:50Yeah, that's right. And so this was the moment when three, four years ago, when Dario and Tom first gave me that call after GPT-3 and said, you know, we think we're, we've discovered this thing called scaling laws. And we are, we think there's a new lab to be built that prioritizes interpretability first. And I said, sure. And they said, you know, can you help us out as an investor? And I said, sure. How much do you need to get started? And Dario said, I think we can get by with five. And I said, okay, look, 5 million, not a problem. I should be wired over next week. And he said, well, you're off by a couple of zeros there i'm going to need 500 million and i said okay things are going to be a little different but sure walk me through you know usually i find that when when some of the most the leading scientists uh kind of challenge a base assumption you have it's usually a good and you feel uncomfortable it's it's a good i i find it's it's a good practice to dig into that on discomfort and kind of big ask the five whys on from a first principles basis why are they so convicted in that right and um you know in in in his case dario tom and i spent many weeks unpacking why they felt they needed 500 million at the end of that conversation i realized actually that 500 million was extraordinarily efficient to train to build a lab of a capital race to build a lab around because by you know by then open ai had spent over four billion dollars getting to that point yeah and that's when we really got into the idea of compute multipliers and why having compute multipliers allows you to train uh models for six times more efficient uh compute spends than you know the first generation and and i find those long-form conversations you just can't substitute no amount of reading a paper can explain okay can can compensate for those in-person working sessions and that's what i love doing with all the founders you know i work with I often find it most rewarding to be the first call for a scientist or a researcher about commercializing their research before there even is a company.

34:56And that's what I spend my days usually filling. And I find that often means you're quite early. The problem with spending your days with scientists who are living two steps ahead is that most of the world today is not ready for their realizations. And so when I introduced them to, you know, Dario to 22 other investors up and down Sandhill Road, they got 21 no's. And, you know, the reactions were everything from borderline, this is snake oil, to, you know, it just doesn't make sense. There's no way that just throwing more compute will eventually produce models that will be able to solve all intelligence problems under the sun.

35:32It's so absurd. Now, they turned out to be right. And so I often find it's not easy to communicate my excitement about these realizations to other people in the field. So I actually I try to spend as little of my time talking to other investors as I can, because I find by the time other investors have conviction in something, it's already consensus. Right. Yeah. What I'd like to spend most of my days is in long form conversations with with scientists and researchers who are in the lab. and I find the most exciting moment is when science is going from being or some strain of technology is going from being inevitable to being imminent because that's what can be most useful to them because they're often looking for my help to commercialize their work and scale the impact that they can have by turning their research into a product that could be used by millions of people and there's no papers you can really read on that it's often the hard work of sitting down and discussing face-to-face the technology, the capital required, the compute needs, what the industry needs.

36:43And that messy search for product market fit is what I find most exciting. Yeah. And how many was this? Yeah. And I read a lot of papers and I'm a journalist. uh you know and i i'm not a practitioner i'm not an ml person i mean i didn't go to school for that and a lot of times i'll see something and i'll think wow you know you could you could build a product off of that uh and a lot of that research just gets shelved or buried in in archive and no one ever does anything and maybe you know maybe it's the idea is rediscovered 10 years later and someone says oh yeah well so and so wrote that uh 10 years ago is most of your i mean are you investing i mean that's that's that is your role right yeah that's right i'm a general partner i spend my days investing um but i i think the the discipline of investing weirdly has changed so dramatically in the last few years that it actually looks, at least in the world that I spend my time in, which is AI infrastructure, you know, at the earliest stages of taking AI technology out of the lab, it looks much more like the early days of the venture capital industry than, you know, let's say the 2010s, which were, you know, in the earliest days in the 1970s, when you had firms like Kleiner Perkins and Sequoia getting started, these were founded by former semiconductor engineers, right?

38:19You know, Eugene Kleiner and Tom Perkins, obviously from the Fairchild line, right? Genentech, for example, as a company was founded literally in the basement of Kleiner Perkins by a former Kleiner partner and Herb Boyer, right? A scientist at UCSF. And the way these companies were formed is often you'd have a leading scientific mind who felt like their research was ready to leave the lab, who'd often team up with a former operator who understood that capital could be used in really strategic ways to accelerate the commercialization of that science. And when I got the Kleiner, you know, I was 19 at the time, I was really lucky that Brooke Byers, who was the B in KPCB, took me under his wing and kind of, you know, gave me a crash course into how venture capital used to be practiced.

39:03And that was quite different from the way it was being practiced by some of the newer partners there because we had gotten into this era where i think fueled largely by the rise of cloud and mobile um a lot of the the the technology that businesses that were being built did not really have a heavy scientific component right um and so it i i think yeah so you're right i do spend my days investing, but it often looks kind of like, it looks quite different. Even though these businesses that I end up helping these scientists start are often software businesses, you know, Anthropic, Mistral, Black Forest Labs, Sesame is one that we just took out a stealth that is building a conversational voice companion.