How Open Source Became AI's Backbone | Inferact with a16z

Elena Burger and Matt Bornstein are joined by Simon Mo, co-founder and CEO of Inferact, the open-source inference engine powering many of today's most advanced AI applications. Together, they explore how open-source AI evolved from a research project into critical infrastructure, why inference has become one of the most important layers of the AI stack, and what it takes to bring frontier intelligence to developers around the world. The conversation covers vLLM's origins, the rise of open-weight models, why companies increasingly want control over their AI infrastructure, and how open-source inference enables the next generation of AI applications. They also discuss model licensing, the economics of open-weight AI, Kimi K3, distillation, AI infrastructure, and why Simon believes the gap between open and closed models is rapidly disappearing. Timestamps: 00:00 - Intro 01:00 - What Is vLLM & Why Serving LLMs Is a Fundamentally Different Problem 05:10 - When Open Source Became Critical Infrastructure 08:26 - Where vLLM Sits in the Stack 13:35 - The Open Weights Letter & Why Open AI Development Must Be Protected 16:42 - K3 Economics: Bridging the Gap Between Open & Proprietary 19:57 - Licensing Evolution: From Apache 2 to Commercial Terms 28:51 - Why Open Source Inference Is the Only Way to Scale Agents 32:21 - The Hugging Face Incident & Why Guardrails Break Down 36:57 - Building a Company from an Open Source Project 43:32 - The Distillation Debate: Is It Critical or Incidental? Resources: Follow Simon Mo on X: https://x.com/simon_mo_ Follow Matt Bornstein on X: https://x.com/BornsteinMatt Follow Elena Burger on X: https://x.com/VirtualElena Follow Inferact: https://x.com/inferact Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  The fun thought experiment is if GPUs dropped in price by 99%, then do we get back to a real open source world?

• 0:07Speaker 01

  If moderation is never solved, in the future, people will go to Openweight by default. Because that is where you know for sure you can control your guardrail for trusted use cases.

• 0:18Speaker 02

  And can you talk about where VLLM sits in that stack?

• 0:21Speaker 01

  VLLM is an inference engine. It is kind of like databases and operating system and other critical software to power AGI, Nvidia, AMD, Google. Their newest chip will make sure VLLM can run on them. And in a lot of cases, they use VLLM as a benchmark while bridging almost a 10x gap. For proprietary models, there is a regular mode and fast mode. But for Openweight, every provider can offer potentially even 10 different levels of speed.

• 0:49Speaker 02

  Five years from now, open source AI models, have they closed the gap with frontier models?

• 0:55Speaker 01

  Capability wise, I don't really see a big gap, not even today, because...

• 1:01Erik Torenberg?

  Open source AI has become one of the most important forces shaping the industry. But making frontier models available to the world requires far more than releasing model weights. It requires an entirely new infrastructure layer. In this episode, Elena Berger and Matt Bornstein are joined by Simon Mo, co-founder and CEO of Inferact, to discuss how open source inference became critical infrastructure for AI, why enterprises are increasingly turning to open weight models, and what the next generation of AI systems will require. They also explore model licensing, inference economics, and the future of OpenAI.

• 1:41Speaker 02

  Today, we're here with Simon Mo, co-founder of Inferact and a lead maintainer of VLLM, the open source inference engine now running on half a million GPUs at any moment. We're also joined by Matt Bornstein, an A16Z general partner. Simon, Matt, thank you so much for joining us. I think first, we should start with open source AI and kind of the more recent history of open source AI. So VLLM actually has its origins kind of back in 2022, pre-ChatGPT. And your team set out to make a slow open source demo faster and instead just found this pile of unsolved problems.

• 2:22Speaker 02

  So can you talk about what made serving an LLM so fundamentally different from the ML workloads? Everyone already knew how to run.

• 2:31Speaker 01

  Yep, good to be here. So serving large language model is a fundamentally different problem because serving it requires to run it on accelerators like GPUs or TPUs. And it is a computationally intensive process that will require a lot of engineering and ensuring that for each request, user can see the LLM's response quickly and efficiently. So this typically means we need to handle differences in input distribution, how long each request is, output distribution, which is non-deterministic, and batching and scheduling a lot more in the, as a core of the inference engine.

• 3:13Speaker 02

  Yeah, and so the team, the project has been around for about four years now, but the company is a little bit more recent. So I want to throw this out to either of you. Matt, I know you've known the team and observed the team for a very long time. So at what point did you see this sort of transitioning from being a much beloved open source project to critical infrastructure and then a company?

• 3:39Speaker 00

  Yeah, I mean, I think you have to go back a little bit, right? Open source was the norm for AI models early on, right? I mean, we literally had this company called OpenAI, which, you know, it's become a little bit of a joke. It's not as open as it once was or not nearly as open as it once was. But early on, all the frontier AI work was being open source or at least released into open weights, which is a little bit different than true open source. And people could mostly run these models sort of on hardware or computers that they already had. I'm curious, Simon, like maybe you remember, like, what was like the first model that like you actually needed to go out and get special software and like a special set of computers to be able to run?

• 4:14Speaker 01

  Probably BERT. And before that, it was like ResNet for computation, like images, computer vision, classification. So ResNet already need to run on Nvidia K80, which is kind of one of the first SKU in other places. And but way over, but even at this point, ResNet, you can still kind of run on a commodity, even CPU devices, it's just very slow. But for BERT, where running at is like, wow, you have to run it on GPU to make it anything faster and efficient for anything, translation or any task. So that was like before 2020 even, yeah.

• 4:53Speaker 00

  It's so funny to thinking about this. Yeah, I'm like, all the memories are flooding.

• 4:56Speaker 01

  Yeah.

• 4:56Speaker 00

  Hugging Face had a thousand BERT variants on it.

• 4:59Speaker 01

  Yeah.

• 4:59Speaker 00

  You have to go find like the right BERT variant for your particular task. And yeah, and you're right, you had to, some people, I guess, had their own GPUs and could run it. But yeah, a lot of people had to go provision cloud and stuff like that. That's really funny. And so, yeah, so, so look, I mean, BERT was an early language model that newer models are much bigger, much more sophisticated, take up a lot more memory, a lot more compute. And so VLLM, really from the early days, right, was about running these more powerful models that you couldn't just sort of do it, figure it out on your own.

• 5:27Speaker 02

  Yeah, yeah. And I think to get us to this present day, I think it would be good to talk about kind of at what point it really became critical infrastructure for these even larger open source models. And when did we even start to see these larger open source models kind of come into the field?

• 5:47Speaker 01

  While we really see the criticality of the software stack, it is always about bringing open frontier intelligence. This is, open frontier intelligence is a catchphrase for the K3 model release this time. And it's also about VLLM's mission about inference infrastructure. So if we look at it, when did this start being critical? Is when people are relying on it for their application, for their day-to-day productivity. So maybe 2023, when GitHub Copilot and ChatGPT are something that people just cannot live out with anymore. And that point, Openweight model is already backing and become a cornerstone of how people are living every day.

• 6:32Speaker 02

  Yeah.

• 6:32Speaker 01

  And at that point, we will need the accelerator running open weight model and open weight software, to open source software, to make sure it works well. Yeah, and have all flexibility and control over it.

• 6:45Speaker 00

  Yeah, it's sort of interesting from a startup standpoint. Like you mentioned, sort of GPT-3 or like early ChatGPT, those closed source solutions were starting to become critical to like a small group of people in that time. And open source existed, but it was a little bit of a curiosity or sort of an enthusiast thing. As the frontier has expanded, particularly with closed source models, more and more open source has been dragged in as kind of like critical behind it, like if that makes sense. Like at any given point in time, including now, I think models from OpenAI and Anthropic are kind of more widely used and more critical kind of in general than open source models. But I do think we passed a threshold, and I want to say about a year ago, where a bunch of smaller companies or like new application companies, as they were trying to figure out, how do I really build an AI

• 7:30Speaker 00

  without just being a wrapper on top of OpenAI? The answer to that question turned out to be open source. I mean, this is what Cursor did, this is what sort of Decagon and Harvey are in the process of doing now. And a bunch of other like really, really strong application level startups sort of made the determination, we can't build just on closed source. We need to do our own mid-training, our own post-training, our own sort of inference and deployment tricks. And all of that means it must be built on top of open source. The closed source vendors won't give you the access to do this. So my read is like kind of a year-ish ago. Open source became really central in a way that's not always visible because it's deeply embedded in some of these products. But some of the most innovative products and applications now really depend on this very deeply.

• 8:10Speaker 02

  Yeah, yeah. And can you talk about where VLLM sits in that stack? Where we do have these larger enterprise companies that are choosing to use open source models? Like where does VLLM sit in the stack for them?

• 8:21Speaker 00

  Yeah, I mean, you should, like, just about everybody uses VLLM. You should describe it.

• 8:25Speaker 01

  Just about everybody uses VLLM. VLLM is an inference engine. That means its job is to turn available GPUs into a running endpoint for intelligence. So that means it is kind of like databases and operating system and other critical software to power this economy or power the AGI that everybody really uses today to ensure they can have cost effectiveness, efficiency, reliability, and also always staying on the frontier. Because for VLLM, we support more than 1,000 model architecture up to today. And a lot of those are proprietary, but also a lot of those are open weight, right? And a lot of those model architecture, when they are becoming, transitioning

• 9:10Speaker 01

  from a research prototype to world accessible open weight model architecture, they are live on VLLM immediately. So that's what a process we call day zero model release. And additionally, VLLM also work closely with all the hardware vendors. So that means across like Nvidia, AMD, Google, and Amazon Intel and a lot more. Their newest chip will make sure VLLM can run on them. And in a lot of cases, they use VLLM as a benchmark to make sure it runs well on them. So this kind of fusion of where models run and where it gets to meet the hardware is where the magic happened and this is where VLLM is.

• 9:47Speaker 00

  And you've told me some of the behind the scenes stories, like it's actually not easy. These day zero model releases, it's like a lot of human drama in addition to like technical work. I guess, are there any stories there that you think are okay to share?

• 9:58Speaker 01

  Oh, it's actually a very fun co-design process because from model labs point of view, right, these are brilliant researchers who have built this model. Now their biggest question becomes, how do we get this out of the world and make sure everybody's able to use it and run it well? And we have worked with model labs that are very, just because they just use VLLM already in production or in their research process, they will just done everything for you. Because this is a moment when we go to them, it's like, hi, where is the VLLM team? We would like to support your open source model. We would like to offer in a way this kind of open source, but white gloss service to get your model running well on architecture. And then you return, we get the model labs like, oh, we're good at working already

• 10:41Speaker 01

  because we're running it for the ROI process. Here you go, just review our code and merge a pull request. And on the other end, we really have model labs that just don't know how this can work.

• 10:51Speaker 00

  Because systems is not like their core.

• 10:54Speaker 01

  Yeah, because this thing is not their core and they have been training or maybe they have their internal inference engine that just don't know how we will adapt to the open way ecosystem. And by the way, this is also a very much a multi-party kind of involvement process. Every model is typically involves the the model lab, involves a primary or secondary hardware vendors, involves us, involves Hugging Face, who are the model format and like model hub vendors. And then depending on the appetite, the model lab involves ten or twenty different kinds of release partners. These could be inference clouds, this could be public hyperscalers, whoever is going to run this model and you want them to ensure the model are running successfully, right?

• 11:37Speaker 01

  So even up to today, if you look at the K three model release, it's a whole partnership and a drive to make sure that the model is. Once the model is released, because it's just a few terabytes of files sitting on the internet, that people are actually going to be use it really, really well. Even from the beginning of twenty twenty three, twenty twenty four, if you remember when Mistral dropped their first model, they just dropped a torrent link for PTP. And then everybody's like struggling and trying to get it up and running. And then we're working behind the scene with the Mistral team, trying to get the inference engine support working in V R. And this was the most probably early on exciting weekend that we are able to spend on this.

• 12:19Speaker 01

  And then and after the weekend, when everybody's trying to really analyze what's going on and Monday, Tuesday, well, Mistral and us just announced, here, you can run it on V R successfully here. And everybody will able to easily reuse a lot of the work and start building on top of it.

• 12:36Speaker 00

  That was sort of a fun time where like enthusiasts like me could just like scramble to like download the model and like get it running somewhere.

• 12:44Speaker 01

  Yeah, yeah.

• 12:44Speaker 00

  I'm glad the professionals have taken over because it never worked like very well. It was like a fun moment in time.

• 12:50Speaker 01

  Yeah.

• 12:50Speaker 02

  So so to bring things, you know, forward to the present, I think open source models and also, you know, distillation have have been the news recently. Inferact signed the Nvidia Open Weights and American AI Leadership Letter that that was signed by also A sixteen Z, Meta, Amazon, dozens of other companies. Can you just talk about, you know, your decision to sign that and and sort of what what you were really kind of responding to in the market and and kind of in the news?

• 13:26Speaker 01

  Yeah, so for us, what I really want to stand behind is open way absolutely matters in the ecosystem. The world cannot just be controlled by proprietary APIs and where open way, open development and research of these models are blocked or banned, right? The pledge that Infra agrees enough for is we want to help and foster this ecosystem. Where we are typically in a little bit downstream of this ecosystem, right? Inference engine are not part of the pre-training process nor the R L process, but where where the model actually meets the world.

• 14:02Speaker 01

  And from what we're seeing, people are just really using their imagination and ability to materialize this imagination of open way model. They're able to leverage this open way model so much effectively.

• 14:16Speaker 00

  There's almost two pieces to this, right? There's like the the cost thing where it's like the closed models are too expensive. And then there's sort of the control thing where I I want to sort of be in control of my infrastructure and and and in control of the model, right? If I need to extend it or or put on my own guardrails or anything like. I'm just curious, have you heard from from customers? Like, are are both those things important to them or or like, are they kind of willing to pay as long as they have the control or or maybe there are different use cases?

• 14:45Speaker 01

  I think like it fluctuates over time, so control matters a lot over the last few years and then costs just start to matter over the last few months. So cost really matters, starting from people trying to migrate off, right? Their expensive coding plan and like every skyrocketing token maxing spend. But control has always been in the backbone of this is they want to even in a way to control the cost, right? But also it's about controlling the system performance against what they're paying for. So there, for example, for a voice agent company, they want to control their own model so that it can make sure the model actually respond by their required time.

• 15:25Speaker 01

  So the customer, when they're on the phone, they can ensure the agent is responding according to a S L A. And this sometimes is only you can do with your controlled intelligence, because you know the whole hardware you're running and the whole system you're monitoring. Versus signing up for relying on your critical infrastructure with a proprietary API where they might go down anytime or have violation of the contract anytime.

• 15:50Speaker 02

  Yeah, Simon, you also, to to go back to the cost point, you actually make the point in an essay you recently wrote about the release of Kimi K three. That actually the economics is besides the point and it's actually, you know, in in the case of these, you know, just really, really great open weight models that are on the frontier, that were designed by really brilliant researchers, like like. These models are in some cases just as expensive as, you know, the closed source models. So in those cases, kind of what is the point of running them and kind of what do we learn architecturally in the in the course of running them?

• 16:33Speaker 01

  Yeah, so first on cost, it's not necessarily they are as expensive as a proprietary model, but rather first, the cost. Cost discourse has been discussed over and over again with even G O M five point two a few months back. So open way model are sometimes definitely a lot cheaper. And but for this model, there's a big sort of step change where we're bridging almost a ten X gap, but strike somewhere in the middle. Where Kimi K three is not as expensive as Claude or G P T Soul, but it is a lot more expensive than G O M five point two. Why is that? And I do believe this is the point of where pricing intelligence with the market correctly and understanding where it is and.

• 17:18Speaker 01

  But then the majority part of the discourse should be focusing on, wow, this model is bringing a Opus four point eight level model to our own infrastructure that I can use, I can run, I can fine tune. I can be able to understand exactly how many tokens do I need, understand the exact performance profile. The reason here, for example, is for proprietary model, there is a regular mode and fast mode, and that's only the two switch here. But for open weight, when you're running it, every provider can offer potentially even ten different levels of speed, going from like the slowest mode, which can be a lot cheaper

• 17:56Speaker 01

  to four hundred tokens per second, almost up to five hundred in many cases for some workload. And this is typically two X or three X faster than the fast mode out there today. So this kind of level of control, even in terms of performance, and then let alone control over how customer interacting with the model, control over data retention. Keeping in mind, Fable doesn't have zero data retention policy and at least a lot of the data need to be staying there. And let it control security and compliance a lot more. Yeah, this is why I'm particularly excited about K three, not just from the cost perspective, but a lot more on bringing this level of intelligence to something people can own.

• 18:41Speaker 02

  In terms of, you know, calibrating things like speed, calibrating other things, just sort of on the back end, what needs to happen? And and kind of what are you seeing your users do and like, like, who who is being really clever about this?

• 18:56Speaker 01

  So we do see users are able to get the maximum benefit out of this model when they enable fast mode. Like what I'm talking about here, of course, is V R on its own fast mode, getting up to four hundred and five hundred tokens per second. Because it is really a big step change from like, especially when developer are interacting with the model, they can see, oh, I can really just get my task done faster here. And the model are now no longer stuck in thinking, rather it is just executing, executing and interacting with the environment. So for premium developer blocking focused task, we're seeing is very benefiting.

• 19:34Speaker 01

  But also K three are just be able to have the ability for people to modify it and fine tune on top of it, allow them to make it better for their own workload. And this is definitely happening today as well.

• 19:47Speaker 00

  Can you just explain what the licensing term is for the most recent open source models compared to the past? And and like, why you think they're doing that?

• 19:56Speaker 01

  Oh, yeah, so historically, the the open way model are just like Apache to like our software, which is like, take it, modify it, do it, whatever you want with it. Here is a gift to the world. And then recently, the model lab are trying to understand a way to economically fund their own model development after all, model training and research are. And the data are very, very expensive. So we have been starting to see terms, even to the Llama days, for for when Meta was releasing Llama. They do have a term of if your daily active user or like annual recurring revenue will see some threshold, please enter into a commercial agreement with Meta specifically, right?

• 20:36Speaker 00

  I do remember that the numbers were like specifically chosen at that time that you could go find it was like two companies in the world that like fit the definition that they had excluded from their license.

• 20:45Speaker 01

  Yeah, exactly. But like, people have taken a hint from that, especially now the labs are trying to figure out a way to economically fund it, especially when they're open source the model. Everybody can just take it, run it themselves, whereas nobody will use your API anymore in many cases, while their API currently still taking up shape, right? And now we're seeing a very healthy ecosystem development, starting from even Mini Max recently, when they're releasing the M two point seven model. They have a term specifically focusing on a usage. And Kimi initially also has like, if you have derivative derivative works, like this is kind of big news back then with fireworks and cursor about how they build on top of Kimi model.

• 21:28Speaker 00

  Yeah, and and it's, if I if I could just expand on that a little, like, I I don't think it's greed, at least what I've seen from open source model labs, right?

• 21:38Speaker 01

  Yeah.

• 21:38Speaker 00

  Open source models, really what we're talking about are open weights, right? And and it's just not software, right? Like, an A I model is not software at the end of the day.

• 21:49Speaker 00

  And so, open source software used to be supported by people donating their time, or big companies kind of authorizing their employees to donate their time. So it was sort of like a bulk in-kind, you know, donation of people's time. That really doesn't work at AI, right? Like, I can't just like go home at night and like train a frontier open source model with friends for fun. Like, we need millions or billions of dollars of computing resources in order to do it. So I think it does support your point that, like, obviously, there need to be economic incentives and there need to be funding mechanisms in place. Frankly, I think even more so with Chinese models than with domestically-produced models, right?

• 22:31Speaker 00

  If there's no source of economic, if there's no source of funding for Moonshot to continue to train models, like, we know where the funding will come from instead, and it's not like something that we, right? You know, it's government and things that like are actually worse for us, I think. So like, I think you raise sort of an interesting point that this is an important economic structure. And like, I think this means we'll see more of this in the future. Would you agree with that?

• 22:55Speaker 01

  Yeah, it's really about sustainability in the end. It's about how do you make sure that all this initial CapEx, almost to train the model, fail again and again and train the model again? Like, how do you really pay it back? And how do you make sure that there's enough confidence and funding from everybody involved to go to do the next one, right? And I recently heard someone or recently heard someone making analogy to this, to the pharmaceutical industry. It's almost like, how do you make sure that the R and D process of new drugs are properly funded and is proper sustainable method to making sure that people are willing to take big risk, big bet, to go to do research for new drugs and then later,

• 23:37Speaker 01

  because they know there's a economic incentive in the end when the new drug reads to the market. A portion of those, of course, like besides the just distribution channels, right? A portion of those revenue will flow back to continue to fund the next R and D effort. And this is where we're kind of seeing, similar to the model development now, yeah.

• 23:57Speaker 00

  That's a really interesting analogy, because it's like, once a drug, a molecule is released, you have the strongest possible control, which is nobody else can manufacture it at all, right? It's like the most closed possible source, right? Like, it's like a secret. But, yeah, like in the case of models, especially open source models, you know, once it's out there, anybody can take it, use it, extend it, et cetera. So having, yeah, so having some economics attached to it probably does make sense.

• 24:25Speaker 01

  Yeah.

• 24:26Speaker 02

  Well, actually, that raises a question for me too, which, Matt, you were alluding to this earlier about how different open source models are from the dynamics of open source software maintenance. When it comes to open source AI, what actually needs to be maintained? Is it the infrastructure around it? Do the models themselves need maintenance at all? Just kind of what are those dynamics? Because I think even the developer behavior around it is pretty different.

• 24:56Speaker 00

  Yeah, I mean, Simon sort of said this already, but maybe I'll just expand a bit, which is, you know, you see the results of a big training run, you know, where training now means it's pre-training and then, you know, RL kind of, I know, mid-training or sort of post-training on these things, or, you know, pre-training SFT RL, right? Like, it's a very complicated pipeline. We see only the result of this at the end. And the numbers are big. You're like, oh, you know, this was a hundred million dollar training run. But what you often forget is like, there may have been five failed training run, large-scale failed training runs before you even get to it.

• 25:31Speaker 01

  Yeah, the blood, sweat and tears.

• 25:33Speaker 00

  Yeah, exactly. One of my favorite artifacts, maybe we could even track down the link, is one of the early Llama models. They published the whole, like, conversation log between the between the people who were babysitting the training clusters, like while the models were training. And it's so funny, it's just like, oh no, everything's gone wrong, like chaos, like panic. And then the next comment is like, okay, we solved it, everything's okay. Clusters have losses going down. So there's a lot, a lot, a lot that goes in behind the scenes before these models get released, you know? Once they're out there, I guess it's a little bit more in your zone, you know, to kind of make sure it's like operationalized.

• 26:09Speaker 01

  Oh, yeah, but this is also a very interesting point. Once it's out there, is a whole community effort trying to optimize this model? Because the model is trained on a given type of hardware and given type of architecture. And but when it's out in the wild, everybody has different cluster topology and use cases. And it's about how do you turn like a use case of one now to a use case of almost infinity? Now you have people trying to adapt it to the edge devices and people trying to run it at largest scale ever. Adapt it, making sure it runs for voice agent, but also for coding agent, which are entirely different kind of use cases. So this is a whole community effort trying to further optimize, specialize, and making sure the running of it is reliable and continue to be able to optimize against it.

• 26:53Speaker 01

  So that's a whole village later throughout the open source to make sure it's improved.

• 26:58Speaker 00

  And that's cool, because that really is like open source software. I mean, and this is what we do, but you know, like, this is like anybody can contribute and like make these better. The fun thought experiment is if GPUs, you know, dropped in price by 99%, right? Like, if GPU-based compute actually became, you know, kind of cheap and widely available, like, then do we get back to a kind of a real open source world where, you know, one person sitting in their basement or a hundred people working their free time can like come up with something new, try many of these sort of model training paths that, you know, that like are in the queue somewhere at one of the big companies? Then, you know, kind of see, you know, really expand and advance the field collectively.

• 27:39Speaker 02

  Yeah, yeah. Well, I mean, this relates to what you were talking about at the beginning. It's like, you know, at the beginning, the amount of compute you needed to be at the quote-unquote frontier was negligible, and now it's just like, it's enormous. And how do you, how do you get that, you know, back to consumer parity again?

• 27:56Speaker 00

  I've said this on the podcast before, but I'll keep saying it. AlexNet, first, you know, kind of like neural network to run on GPUs that we care about, ran on two GPUs. And that's not, that's not like there are no missing decimal points or commas in there, literally two. Now that would get you literally nowhere.

• 28:12Speaker 02

  Yeah, yeah. So I guess this relates to another question that we've had, which is inference has gotten harder over the past 18 months because of this combination of scale and diversity and the kinds of models that we have. And of course, like agents that are doing increasingly long-running tasks. So what makes open source in this world and in this scenario not just like nice and nice thing to have, but absolutely necessary?

• 28:42Speaker 01

  Yeah, so scale comes from a few points. It comes from whether or not you can run this gigantic model on a data center rack. And but also it comes from, can you optimize it to the extreme point, to the speed of light, so that you are getting the most value out of it? But also, you need a whole community and collaboration and partners of effort to validate it and making sure that there is no, that last remaining bug that appear like 0.0001% of the time. So this is where scaling up the deployment and then making sure more people are running it as like largest footprint possible will ensure everybody's experience of running this model is better.

• 29:22Speaker 01

  And this is where kind of, this is why open source inference is the current leading way right now instead of closed source inference engine. And frankly, right, all the, a lot of the open, a lot of the open, sorry, a lot of the inference cloud and API as a service today do use and leverage open source inference engine under the hood. And the reason to do that is just there's just so much battle, tested and learned recipes and things that they can build on top of. And this is where open source is the current leading way to run models.

• 29:59Speaker 02

  Something that I find kind of funny in like looking back at not just the history of VLLM and Inferact, but also a company like OpenRouter or even Ollama, all of these different teams kind of got started around 2022 and 2023. Some of them even before ChatGPT, in your case, built around open models. And what kind of special thing do you think was happening at that time? Just like in the world of research and in the world of AI, where we see these companies now and we look at them and you think of VLLM as mission critical or OpenRouter as mission critical? And what do you, what do you think was going on at that time?

• 30:44Speaker 02

  Where even before we had, you know, a big consumer use case, we had, you know, teams like yours that were, were building out these things and kind of, how did you guys know, I guess is the question?

• 30:57Speaker 01

  I think there's two parts to this. Our team always have open source kind of angle. We're from UC Berkeley, a long tradition of open source software and our system research principles. But also just curiosity overall, everybody was so curious about how are these AI models going to benefit the humanity and how can we use it better? And this is where open source is where you meet the most mission-aligned people together. Right, OpenRouter, Ollama, as you mentioned, they're all great partners and friends in the ecosystem because we're there to understand how everybody else can leverage AI model better and making sure they have a good use of it.

• 31:37Speaker 02

  Speaking of, you know, leveraging open we, open source models really well. So I think another thing that was in the news really recently was Hugging Face using a Chinese open source model to help contain a cyber attack carried out by a rogue un-Sandboxed OpenAI model that was being tested. So like, what can we take away from something like that? And can you just first, I mean, walk us through your impressions and understanding of what happened? And then kind of what you think the takeaways there are?

• 32:12Speaker 01

  Yeah, like, this kind of goes back to our previous point about control. So for the Hugging Face incident, they break it down pretty well on their website and blogs. So really thanks for their transparency there. In the end, it's about all the closed proprietary model API. Their guardrails are a little bit arbitrary, but also very difficult to enforce. That means they have so much false positive in the guardrails that will have legitimate use cases just being blocked. And this is like a evergreen problem, even in the social media days. Like, how do you design content filters correctly and like be able to do the moderation correctly? So if moderation is never solved, which is gonna be very, very hard,

• 32:56Speaker 01

  then there's always a place where you have a model where you know and trust that you are publishing to and be able to use from. So Hugging Face, they have to use OpenWay model. But in the future, also see for the trusted use case, people will go to OpenWay by default. Because that is where you know for sure that the guardrail is lessened or you can control your guardrail for trusted use cases. Even for us, like this is actually applied to us today, where a lot of the Anthropic model do are banning frontier AI research. And then when we're studying GPU kernels, even as a invalid memory access error, we are triggering the red line.

• 33:41Speaker 01

  And so a lot of our developer within InfraX and for V L M are like retreating from using Fable five because you have a two hour job and you trigger the red line, which is false positive. And then you have to lose all of your work. And so a lot of our developer are using like Kimi K three today even just making sure. Because it's similar quality and it has a guardrail that makes sense to us. There goes while using it.

• 34:11Speaker 00

  I think your social media analogy is a really apt one.

• 34:13Speaker 01

  Yeah.

• 34:14Speaker 00

  Because like in both, in both cases, what's kind of happened is you've taken like distributed human activity and kind of centralized it in one place. Like, like before social media, people talked to each other.

• 34:24Speaker 01

  Yeah.

• 34:24Speaker 00

  And published articles and like messaged each other on message boards. But then you centralize it all under one kind of like profit-seeking enterprise. And all of a sudden, the incentives will change like a lot, right? And we have a specific carve out for social media, which is you're not responsible for what people say on your platform. And that allowed the moderation problem to be tractable, I think.

• 34:47Speaker 01

  Yeah.

• 34:47Speaker 00

  Right? It's like, okay, we're gonna do our best to eliminate obviously illegal things. You know, like things that are just gonna degrade experience on the site, you know, you know, in really bad ways. But like, you know, if it's on the edge, you know, it's something that's maybe out of fashion to say, but not illegal.

• 35:03Speaker 01

  Yeah.

• 35:04Speaker 00

  Like, we're, we just can't police everything. We can't be the police of like all human communications. I think something similar is sort of happening in AI, right? Where a lot of work, not just talking, but work is kind of being consolidated in one place. If I need to write code or create a spreadsheet or get advice, you know, health advice or like anything. It's all happening in like one or two sites.

• 35:26Speaker 01

  Yeah.

• 35:26Speaker 00

  And they don't have that carve out, right? Like, they don't have that exemption of like, hey, we're not responsible for what actually happens on the site. And like some of them, especially, you know, Anthropic, like is kind of going further than even what would be sort of like legally required and they're sort of taking ethical stances on these things, which may be right or maybe not, right? That's sort of their decisions.

• 35:45Speaker 01

  Yeah.

• 35:46Speaker 00

  How do you actually do this, right? Like, the problem's actually bigger than just communication because it's like actual work in action. And you don't have this exemption and you have this sort of layer of ethical stuff on top. So I, yeah, I'm just totally.

• 35:57Speaker 01

  Much harder.

• 35:58Speaker 00

  It's, I'm just totally with you. It's like, and it seems like some of these, you know, companies have gone like err on the side of caution, which is probably a reasonable thing to do, but very, very frustrating as a user, right? Like I, you know, I saw online, some translation attempts are blocked because they think the particular combination of like language and content is somehow like triggering like politically sensitive topics and, you know, things like this.

• 36:22Speaker 01

  Yeah.

• 36:24Speaker 02

  Simon, we're nearing the close of the conversation. And I just wanted to take a step back a bit and ask just about InfraX and, you know, running the company. And I know that Ian Stoich of Databricks is an advisor and a co-founder of InfraX. And I'm just curious what you've learned from him in terms of taking an open source project like V L L M and building a company.

• 36:48Speaker 01

  Yeah, Ian, as a co-founder, he has always been thinking about open source and where, how do you support open source better? And then now with experience from Databricks and AnyScale at, and even Arena, which is a public platform. And now at InfraX, his focus really stemmed from how do we build such a, in a world where this software is being used by so, so, so many people on such a critical project? How do you make sure they're getting the best quality and where is the value come from, right? If the open source movement and open source inference engine is a starting point,

• 37:30Speaker 01

  then what are the gaps that we can fill as InfraX? And then what are the gaps that we should absolutely put in? Ian has always been open source first. So for whatever we need to build, we would like to build them in open source. And then for whatever value we can deliver, we want to really close the last mile and making sure that customer and partners are getting the best out of it.

• 37:53Speaker 02

  Just, I guess, to close this out, five years from now, do you think open way, open source AI models, have they closed the gap with frontier models completely? Are frontier models always one step ahead? Kind of how, how do you see that shaking out?

• 38:09Speaker 00

  Five years.

• 38:10Speaker 01

  Five years is a lot.

• 38:11Speaker 02

  All right, all right, one year, one year. Five years, who knows?

• 38:17Speaker 00

  We're either, you know, we're all gonna be like just floating around in our Wally pods on our spaceship.

• 38:22Speaker 02

  Exactly.

• 38:23Speaker 01

  Progress, right?

• 38:24Speaker 00

  Yeah.

• 38:26Speaker 01

  For, for me, really, at this point, there's kind of a point we haven't talked too much about is what really differentiate opens open way model from closed way model, right? In the end, there's not much differentiation. It's more about the distribution strategy and go-to-market strategy. And the capability wise, I don't really see a big gap, not even today. Because for how these model are coming to being, they're really starting from the first principle, right? You have a compute cluster, you have training data, and you have brilliant researchers that group together and really to build this amazing artifact that is this pre-trained model

• 39:10Speaker 01

  and then later R O D, post-trained model, and that the world can use. But if you look at the ingredients, right? The, one of the most important part is just the data. It's about who gets what data? And then what are the environment you are building to let the model improve on itself and make better, right? One of the very useful benchmarks that we have on Arena for, for K three has been front-end coding, right? That means for Moonshot, they have built some of the best environment for front-end coding, right? They have published amazing demo on the ability for this model to

• 39:49Speaker 01

  code and then see what the rendered is and then kind of continue looping and this iterative process. Now, this is about their environment to improve the model. It's not about just source data. It's not about where they get the data from, rather is who can build the best environment? And who can make the most sort of optimization and algorithmic choices to leverage autos learning from this environment? So the next year is all gonna be about that. It's about how open way model labs are differentiating and really getting the model to meet the real world and have this kind of what people are popular today, like recursive self-improvement almost to really improve the model overall.

• 40:33Speaker 01

  And so really project out in a year, there's not gonna be any difference.

• 40:38Speaker 00

  Yeah. And you've used this term brilliant researchers a few times.

• 40:44Speaker 00

  There are brilliant researchers everywhere in the world, clearly. What, why do you think, you know, in the U S, all the smart researchers are working on closed models and in China, all the smart researchers are working on open models?

• 40:57Speaker 01

  I mean, from my point of view, they are attracted to interesting problems, not necessarily on the open or closed stance. But rather, but however, open way model does give people a really, really good boost on the impact of such models. So that is like a plus. And I, I think all the brilliant researchers are attracted to how to improve the model overall, right? Like, actually, one interesting point about this, maybe fairly technical for this Kimi K three model is they removed a rotary positional embedding. So rope has always been there for a lot of the transformers model. And guess who removed it?

• 41:39Speaker 01

  Is the inventor of rope, who are like, yeah, like the, Jenny, he wrote the first paper introducing rope as a concept. And then he now also wrote the explanation of why you don't need it in, as part of the technical report in this K three model. So like when we read it, it's like really come full circle. You have all these brilliant, humble researchers that are able to really study how this work and really study the secret of training and pre-training and share it across the whole world and recognizing and iterating on their past, right? So really a miracle, I would say, for this model to come alive.

• 42:19Speaker 00

  It's so, it's so interesting, like, you know, AI is in this funny zone where empirically it works incredibly well. But then you go ask the theorists and they have like no idea. Right? And so like, you have these kind of iterative things where when you go read, you know, primer on transformers, you read about positional embeddings and why, or positional encodings and why it's so important? Because otherwise you can't sort of like understand meaning, you know?

• 42:41Speaker 01

  Yeah.

• 42:41Speaker 00

  And then it turns out, once you understand one level deeper, because we've been doing this for a few years and you have all these smart people like, oh, actually you don't need this.

• 42:47Speaker 01

  Exactly, exactly, yeah.

• 42:49Speaker 00

  You know, simpler actually is better.

• 42:51Speaker 01

  Yeah.

• 42:52Speaker 00

  We, we didn't, we didn't talk about distillation much so far in this conversation, but I think it's very relevant to this. Like, I, I have just one question, which is like, I'm not going to ask like, is distillation happening? I think this is kind of speculation on the part of everybody, you know, in the world. But like, you work a lot with these Chinese labs. Do you think? Distillation is a critical component of what they do? Or are they just doing good work? And distillation, if it's done, is an incidental part of it?

• 43:23Speaker 01

  I would lean to the latter part. Specifically, as I mentioned previously, environment matters so much today. So these are RO environments. These cannot be distilled. You don't have other people's environment to really distill a copy from. It's about constructing it, also understanding the learning process. You cannot distill how the model learns within environment. A lot of these are just not doable today. There are things potentially you can do with rewriting the data sets, making better pre-training data. But again, you can do it with any models. Any models that are going to follow instructions are going to be useful in terms of utility there. So I really don't think currently what we're seeing,

• 44:08Speaker 01

  this is a big cornerstone of what's powering the progress today. In the end, what's powering the progress is still just really smart people with very interesting algorithms, data, environment, and they will produce, of course, compute, they will produce the models.

• 44:24Speaker 00

  I think it has really interesting policy implications. I tend to agree with you, by the way, that we have smart people everywhere working on a bunch of smart things, and it's not about distilling data from any one place. It has really interesting policy implications, right? Because it's almost tempting if you're in the White House to say, oh, sure, we'll just turn off distillation, all our problems will be solved. But I think it's more the case that they're just smart people doing interesting things. And so it's like, how do we adapt to that?

• 44:54Speaker 01

  And creative innovations. Like, one part in my essay would mention that open source and open way really helps innovation because it set out this race track where everybody can learn from each other and see where each person, like, each player is in this race track. And then you are able to improve and stay on the shoulder of each other to improve yourself. So that is where everybody can move forward faster.

• 45:20Speaker 00

  Yeah, and one thing we're looking for a lot from an investment standpoint is people doing more open source model training all over the world, not just in China, not just in the United States, but all over. Right, because you get that magic of collaboration when everybody's doing it and achieve global harmony in all of the business.

• 45:36Speaker 02

  Well, I think that's a good note to end on. Matt, Simon, thank you so much for joining us. Thanks for tuning in.

• 45:44Speaker 01

  Thanks so much. Cool. Thanks, Jim.

• 45:48Erik Torenberg?

  Thanks for listening to this episode of the A16Z Podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at A16Z.substack.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast.

• 46:32Erik Torenberg?

  For more details, including a link to our investments, please see A16Z.com/disclosures.