Who Grades the AI Models? | Ben Horowitz & Rayan Krishnan

a16z’s Erik Torenberg, Ben Horowitz, and Jennifer Li sit down with Vals founder and CEO Rayan Krishnan to discuss one of AI’s increasingly difficult problems: how do you actually measure whether a model is getting better? As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations. They unpack why self-reported model scores can be misleading, how VALS evaluates models in the hours before a release, and why measuring increasingly agentic systems means testing work that can unfold over hours, days, or even weeks. They also explore why evals are becoming critical for enterprises trying to understand the ROI of AI, what happens if token spend begins to rival employee salaries, and how evaluations could eventually provide a shared language for everything from model routing and recursive self-improvement to AI policy and international coordination.

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

• 0:00Speaker 00

  Every time a new trillion-dollar industry emerges, there's a need for this independent testing group. When Meta released Llama 4 on our held-out private benchmarks, the model was actually underperforming. But on all of the major public benchmarks, it was showing incredible capabilities.

• 0:14Ben Horowitz?

  What's the limit of what you can achieve? And then, within that, how are you going about it?

• 0:19Speaker 00

  In an ideal world, take a frontier model and have it train the next version of itself. But obviously, that's very expensive and slow. And so what we're doing is forming a set of proxies for every part of the process that takes to build the next version of the model. Evaluations, as they become more complex, have a fewer sample size, but a larger set of criteria or expectations of them.

• 0:38Jennifer Li?

  Where do you see the gap that's happening today?

• 0:40Ben Horowitz?

  The government kind of has an inclination of what it's afraid of, be it biohacking or cyber hacking. But then there becomes the question of, can the model do it? And then, can you get the model to do it?

• 0:51Jennifer Li?

  What do you think the landscape will look like?

• 0:54Speaker 03

  AI models keep getting better. But the tests we use to measure them can become obsolete almost as quickly. In this episode, I'm joined by A16Z's Ben Horowitz and Jennifer Lee for a conversation with Vow's founder and CEO, Rayan Krishnan, about the increasingly difficult problem of measuring AI. We get into why public benchmarks can give a distorted picture of model capabilities, why independent evaluation matters, and what it takes to test a new model in a few hours before it launches. But this is becoming about much more than model leaderboards. As companies spend more on AI, they need to know which models and agents actually perform best for their own work and whether that intelligence is worth what they're paying for it. Ryan also explains

• 1:39Speaker 03

  why benchmarks need to evolve alongside the models, how Vow is measuring recursive self-improvement, and why evals could eventually become a shared language for AI capabilities, risk, and policy.

• 1:54Jennifer Li?

  So I'll start a question from when Vow started in 2024, after a team discovered that all the public benchmarks are just not sufficient enough to measure model progress. And there needs to be a new methodology and approach coming to keep us on the frontier and help model labs continue to hill climb. Take us back to the inception of Vow and what you see was missing in the market then.

• 2:18Speaker 00

  Yeah, yeah. I mean, so I had a background doing research, in particular, building benchmarks and evaluations. And so what was very clear to me was the very tight relationship between what it takes to build new systems for generation and actually new mechanisms for evaluation. In fact, in order to get one, you often need to get better at the other. And actually, one of the biggest drivers for model capabilities is having a new legible way to evaluate models. And so around early 2024, what you're seeing is there are actually many interesting models coming to market. They weren't all coming from OpenAI. And then also, in particular, it was harder than ever to actually ascertain what was newly capable with the new models. And so in kind of a first-principles way, what we realized is that there would need to be some third-party company that solely existed to build really high-quality evaluations and benchmarks to be able to discern what was newly possible with these models.

• 3:03Speaker 00

  And so we released our first benchmarks in 2024. Now, over the last couple of years, that has kind of been realized by many different parts of the industry.

• 3:10Jennifer Li?

  I guess one of the obvious questions is, why do you think the labs can't do this by themselves? Because they know the best of where the models are hill climbing on and what is missing capability-wise. Why can't they be the benchmarking stores?

• 3:23Speaker 00

  Yeah, I mean, internally, they do build a lot of great benchmarks, and that's what drives model progress. But I think there's an issue when we speak about model capabilities in a way that's self-reported. And so one of the early indications of that you saw was when Meta released Llama 4. That was a bit of a disaster. And interestingly, what we saw is that on our held-out private benchmarks, the model was actually underperforming. But on all of the major public benchmarks, where the questions and rubrics are actually open source, it was showing incredible capabilities. So there's a huge disconnect between what was self-reported based on these open benchmarks and then what we were actually finding with our higher-quality, higher-signal benchmarks. But I think it speaks to a broader concept, which the labs, I think, understand, that they would like to see a rational buying market. They would like to see that when they invest billions of dollars to build a new model, there are actually substantive ways

• 4:08Speaker 00

  they can point to evidence and say, we're advancing in these ways. And it's not just entirely self-reported to justify that investment. And so you also see instances of Demis and others in the industry calling for an ecosystem of third-party evaluators.

• 4:20Erik Torenberg?

  And what's the historical analog that you have in mind here? There are rating agencies, audit firms. What's the right comparable?

• 4:26Speaker 00

  Yeah, I think there's honestly lessons to learn across the board. And every time a new trillion-dollar industry emerges, there's a need for this independent testing group. And I think the fact that it's moved so quickly in AI has caused necessity for a lot of these parallels to be borne out. We think about ourselves as trying to sit on both sides of the market. So there are mechanisms by which labs need to prove that new models are very capable. But there are also parallels where enterprises need to figure out what adoption strategy is going to mount to the greatest ROI for them.

• 4:51Erik Torenberg?

  Yeah. Take us through sort of the six-hour pre-release window before the model drops. Obviously, you need to run tens of billions of tokens without delaying the launch. Which part is you doing an all-nighter versus it being automated? Take us through that.

• 5:04Speaker 00

  It's honestly been a journey. And I think the real goal, North Star, we think about is we never want to be kind of a lagging indicator or a delay to a model release. And so that means we have to move really, really quickly and extract the most possible signal with the rate limits or capacity that we have. And so early on, what this looked like was my co-founder, Lanks, and I pulling an all-nighter to try and get as much done as possible and get results out the door. And now, we built up a team, but we've also really invested heavily in infrastructure. And so we're able to run evaluations in a massively distributed way, running effectively the maximum possible rate limits with every model we get access to. And we also have this internal system called Steve. Steve, the economic Vow employee. And so that's been a mechanism by which we're able to actually take more of the human work over time and put it into Steve.

• 5:48Ben Horowitz?

  How do you deal with the kind of issue that it's a little bit of an AI complete problem in that we still aren't really good at evaluating humans? Or we haven't agreed on it? There are things like IQ tests, there's EQ, there's the Big Five personality and so forth. But there's not really an agreed-upon framework for which we do it. And people have issues with things like the SAT and this and that and the third. And then, of course, models are really good at hacking the benchmark. So that's Llama 4 approved. And so how do you think about that issue? And what's the limit of what you can achieve? And then, within that, how are you going about it?

• 6:29Speaker 00

  I mean, I think the honest answer is that it's forcing a lot of the more fuzzy or distributed forms of evals to be made explicit. What is really the distinction between an associate and a partner at a law firm? And there isn't a clear test or an eval for that in the human world. And so we have to first establish a lot of that in these different enterprise or real-world workflows for us to be able to test models in the same way. And I think, long-term, that will be actually the biggest bottleneck, our ability to take companies and their evals and make them legible. Because that's how we'll figure out what signal we hill climb on and where we actually adopt.

• 7:03Ben Horowitz?

  Very interesting.

• 7:04Jennifer Li?

  Actually, maybe one question for you, Ben, on just like how the industry has formed before intelligence came through. Like, we're now measuring something that's very fluid versus before, like when we're talking about enterprise software. There's Gartner rating on like 70 different metrics. Like, you can sort of stack rank on the quadrant to say this company has these features covered, these features not. But now, it's like very Jack Frontier. That's very hard to measure in industries. What do you see? One is the analogy to the past, that lessons we can borrow. And what do you see that's really going to be the challenge and missing pieces going forward?

• 7:39Ben Horowitz?

  It's a little bit reminiscent of the MPAA, right? Where it's like, what's art? What's porn? Where's the line? When is it R? When is it X? And, by the way, the definition of that has changed over time, I think. Things that used to be X are now R and so forth. And then, what's PG 13? All that kind of thing. And there is no... And the famous line is, well, I know when I see it. And I think that this one, I just think it's going to be necessarily fuzzy, but there will develop norms over time. And you know, like, if enough kind of people who run companies or run finance or run whatever it is kind of agree,

• 8:23Ben Horowitz?

  yeah, no, that's a norm, then I think it is. As opposed to kind of what we have a lot now in the open benchmarks, whereas if you can solve this specific problem, then you're at this level and so forth. I think that's, one, it's hackable, and then it's too narrow.

• 8:38Speaker 00

  I think when you also look to some historic analogies, there's a lot of lessons that you can take from them as well on what's gone wrong, what we need to avoid. I think, for instance, at Val's, one very early decision we made was the decision to never sell training data to labs. It's often a place that we're pushed. When we start working with a new lab, to actually source and sell for them a bunch of training data.

• 8:58Ben Horowitz?

  Yeah, that's a lucrative business, yeah.

• 9:00Speaker 00

  Yeah, and actually, a lot of that industry has now built these gimmick-style benchmarks as a mechanism to sell their data. And so that's become kind of their go-to-market as well. But I think if you look at auditing as an industry, you end up with issues like Enron, where if you have the same group who's responsible for doing the audit, as well as also consulting and supporting the company, you have a mixed incentive structure. And then it just becomes pay to pass the audit, or in this case, pay to win the benchmark. And that's really not what the market benefits from and what we're trying to do.

• 9:30Jennifer Li?

  And Ryan, today you have already a pretty extensive catalog of different type of benchmarks. Some of them are more focused on specific industries, some of them are more like consumer mental health related. Maybe first just talk through what are the benchmarks that are most popular and most like read upon, and we'll love to dive into one of them as well.

• 9:48Speaker 00

  Yeah, well, we've done a lot of work in kind of the economically interesting applications of models. Our finance agent benchmark is used by a bunch of the big financial institutions to get a sense of how models are improving. We also have a lot of good work in coding. So our vibe code bench measures how well models can take a natural language prompt and build a full-stack web application. And so that's been a keen way to track model improvements over the last nine months. Yeah, we're also doing a lot more experimental work. So one benchmark we released recently, I'm very excited by, is our recursive self-improvement index. It's a topic which a lot of the big labs have been talking about and starting to report on in their model cards. But there isn't a shared language to talk about the RSI potential of models. And so we created this as an apples-to-apples way to actually benchmark across the models.

• 10:30Jennifer Li?

  Yeah, I thought that one is a very cool benchmark. It's sort of all the rage in the research community of how do you measure like progress you can make through having more frontier models that you can just compound on the capabilities? How do you actually go about like building this RSI benchmark?

• 10:45Speaker 00

  Yeah, I think in an ideal world, what you want to do is actually take a frontier model and have it train the next version of itself and see where the delta comes from. But obviously, that's very expensive and slow. And so what we're doing is forming a set of proxies for every part of the process it takes to build the next version of the model. So there's some work around pre-training, post-training, harness level engineering, and then seeing in which mechanisms and behaviors the models are able to do very good research work and build something new and where they're struggling.

• 11:15Jennifer Li?

  Very cool. There's also cases where you have deprecated indexes and benchmarks. It's funny that I always watch this benchmark industry, people like, just like early diffusion model days, like you cherry pick whichever image shows up the best and the most perfect, like benchmarks as well. Like you pick something that's, you know, very popular, but maybe already saturated, and you rank very well on that or score very high. But you took a very different approach in like, if these benchmarks are saturated, you'll deprecate it. Like, maybe talk us through the thinking around that too.

• 11:48Speaker 00

  I think it's a necessity, and this is kind of the infinite game we're in. I mean, you're wearing our shirt, and so we have this unofficial motto, "Always a higher peak."

• 11:55Jennifer Li?

  Love the shirt.

• 11:56Speaker 00

  Yeah, so insofar as foundation model labs are hill climbing, they're searching for the next peaks to summit, it is our job to perpetually construct these next mountains for them to summit. And I think that's also how the economy has naturally functioned. Over time, as agriculture becomes less important for our labor market, there are new forms of labor that's required out of our population. And so in the same way, we should expect our benchmarks to keep up with the new frontier for what we want models to do. There's another component of retiring benchmarks, which I think is underappreciated, which is that benchmarks should also be reflective of the current state of the world. So in the same way, if you're a lawyer, you have to retake the bar exam and get certified,

• 12:38Speaker 00

  or if you're an architect, you have to get your certification, or you know, a doctor. We should also expect models to be tested on the current state of the world and what we know in medicine, or what we have as our set of laws. And so in the instance of case law, updating to something like legal research benchmark, that's a desire to create a benchmark more reflective of the current state of the world and also push the models in place we want to see them go.

• 13:01Jennifer Li?

  And how has, I guess, one, like, it used to be like, we're doing these like multi-step questions to like just evaluate prompt answers. Now there's like a lot more agentic work that's happening, whether it's on finance or legal, or coding especially, like there's a lot of, you know, async background agents that can just complete tasks. How has that changed sort of how you build infrastructure, how you think of evaluating the capabilities of not just the models and agents themselves? And there's also like a lot more dimensions that people care about. It's not just like capability, it's cost, it's latency, it's like, you know, whether

• 13:44Jennifer Li?

  this model is flexible enough to address like broader domains and tasks and so on. So how do you think about the additional parameters to what you evaluate?

• 13:53Speaker 00

  Oh yeah, there's, I mean, there's a lot that goes into that. I think on the infrastructure level, you have a whole new set of problems. I mean, for instance, now we're testing models and their ability to run over hours, days, sometimes weeks. And so the infrastructure needs to be very stable to support evaluation over time. And if there is a failed request, we should be able to retry from that one and not redo the whole trajectory. So there's some simple mechanisms in the infrastructure we have to think about. But I think in general, what we've seen is evaluations as they become more complex have a fewer sample size, but a larger set of criteria or expectations of them. And so what I mean by that is a benchmark is largely some kind of input space of things you're trying to query a model to do, and a set of requirements or rubrics that you see and expectations of the output.

• 14:37Speaker 00

  And so early on, you have things like ImageNet, which have millions of images you're trying to see a basic categorization for. So it's a one-to-one mapping between an image input and a text label output. Now what we have is far fewer set of tasks. Generate me 50 full-stack web applications, but a much larger, complex mechanism for evaluating the output produced. And I think that trend is going to continue as we see more complex workflows evaluated with models.

• 15:01Ben Horowitz?

  And do you think that it will become kind of a real-time kind of mechanism? Like, so for something like OpenRouter, which Stripe just bought, would OpenRouter look to Val's and say, okay, where should this next request go? Or is this going to be kind of strictly for like picking a model and an enterprise for a task?

• 15:30Speaker 00

  Yeah, I mean, I think, you know, OpenRouter is a bit of a misnomer in that most of their usage comes from being a model gateway. And so it's actually up to their users to decide which models they want to use when. And that's because really the hardest part of routing is building the evals and trying to determine in what places a set of intelligences should be used for a particular application. And so our effort in supporting enterprise and building evals has actually supported a lot of them in also adopting routers.

• 15:57Erik Torenberg?

  And say more about why it's not only important to labs, but also existential for enterprise? And maybe just say more about how you guys work with enterprise?

• 16:05Speaker 00

  Yeah, of course. Yeah, I mean, I think the lab side of this is very clear. Like, you know, if you're raising lots of money, investing heavily in building models, it's essential for you to show why your model is getting better and then why this customer should pay a premium for them. But what I think is still underappreciated is on the enterprise side, this is turning to be existential as well. You know, I have a small anecdote relates to this, actually. You know, I was meeting with a company in the Fortune 10, and the way that they've adopted Claude Code has been with roughly $100 a day budget for their engineers. And so what I was hearing is that this has actually fundamentally changed how work gets done in this company, in that there is a rate limit which resets at 4:00 PM. And so the most productive hours of work are actually now 4:00 to 6:00 PM when the

• 16:50Speaker 00

  rate limit's reset. But then there's this dead period in the afternoon when people go on walks or, you know, get a coffee, because they just don't have the rate limits. And so I think what's really illustrative there is that you see that there is a misvaluing of intelligence happening at every layer of the stack. And so by that, I mean, you have engineers who have $100 worth of usage limits, and they don't really know how to apportion that to the greatest productivity for them. You also have this Fortune 10 company, which has kind of arbitrarily said they're going to allow $100 per employee. They've actually recently increased it to $300 per employee. So almost an employee's worth of salary in tokens for them to use. And this is actually pretty arbitrary, because it's hard to quantify what the right usage limit should be. But then also Anthropic is running on pretty narrow margins to support this, and they

• 17:34Speaker 00

  have, you know, massive costs to serve these models. And so I think we're in this world where it is still very unclear what ROI looks like and how to value this intelligence that's being used. And so as we talk about the existential concern for enterprises, I think it is this kind of direction we're shifting in where token spend may start to eclipse salary spend. And so if this is such a meaningful line item in your costs, you actually have to justify the ROI much more cleanly than you've seen over the last six months. And over time, as we were talking about, I think a firm really is just its evals. And so the ability for a company to make its evals legible in order to solve this ROI calculus is going to be the reason why that company wins out over the competitors in the long term.

• 18:18Jennifer Li?

  And maybe just double click on that, like similar question to why the labs can't do it themselves, or requires a third-party agency to rate it. I think it's a lot more understandable that you need that neutrality across the industry, but for enterprise, they will argue that they know the task the best for their customers. Like, how does Vals come in to provide value? And maybe you can talk through sort of Valsmith new product launch as well.

• 18:47Speaker 00

  I would recommend a lot of companies to develop in-house expertise, but I think that should not be the only solution. You know, there's this explosion of intelligence happening. There are somehow still more foundation model labs getting constructed, and each lab is also releasing more models than ever with many more hyperparameter options, and they exist within a complex set of harnesses and agents. So the option, and we're even talking about specific intelligence, this new paradigm that's emerging. So there's a growing set of intelligence options, and I think what we're finding is that we're still finding new places we want to use AI models. And so the use cases are growing in complexity as well. And so I think if you're a company, you have a compounding set of complexity in the set of options.

• 19:32Speaker 00

  It's very, very hard to develop the internal capability to do the evaluation. And so to try and remedy that, we've started to release some products more openly for enterprises to use, the first of which is called Valsmith. And so Valsmith is focused on code gen, the area we're seeing to be the highest-spended in enterprise AI. It allows any company to take their GitHub code base and build their internal coding benchmark from it to get a sense of what coding agents are going to be the most performant, but also what's going to be Pareto optimal or the highest ROI for them to use. And actually, we use Valsmith a lot of vowels, and we're seeing that a lot of the best enterprises and sophisticated ones are doing that, too. I would expect that to be the direction the market moves as it rationalizes.

• 20:14Jennifer Li?

  And what are some of the examples when you, let's say, benchmark on a private repo, that it just shows very different performance, cost, behavior compared to, let's say, like, using like frontier model, using a public repo benchmark?

• 20:31Speaker 00

  I think today, it's still very unclear whether the best OpenAI model or the best Anthropic model is actually going to be best for your repository. And so we've seen a lot of non-intuitive examples where you actually had to run the eval to figure out what's going to be the frontier performance for that repository. I think you also now see a very complex middle set of options in that there's now OPUS and Sonnet models from Anthropic, but also Luna and Terra, and Luna is very cost-competitive. MewSpark is also very cheap, and 1.2 is very capable. There's also a growing ecosystem of open-source models, which companies can choose to self-host. So I think in this messy middle, it's actually very non-intuitive what's the right fit. We're actually seeing in a lot of cases,

• 21:13Speaker 00

  Sonnet is more expensive than OPUS because it is so token-hungry. And so I think if you were to operate based on, you know, use Sonnet where you feel like it's applicable, you may actually end up spending more than you need to.

• 21:26Erik Torenberg?

  And say more about how this evaluation framework will apply to knowledge work in other domains? Or what are some examples?

• 21:33Speaker 00

  I think coding is a sign for what's to come in every domain, and a lot of the primitives established there are carrying over to other places. And if you have a very good coding agent, chances are you have a model that can also make PowerPoint slides or DCFs in Excel, and with a high degree of capability as well. I think what we need to leverage in a lot of these industries, though, is the existing repository of work that has been done as a mechanism to build evaluations. And so, just as Ben was talking about, we haven't really solved the question of what is human intelligence? But I think in a lot of industries, we have a sitting repository of data around what work has looked like, and it'll be the task of us and others to try and codify that into evaluations that can stay dynamic and actually evaluate models where human work is being done.

• 22:17Jennifer Li?

  Maybe just to take along the earlier question, how are you guys using Valsmith internally to evaluate what's the best coding model for Vals?

• 22:26Speaker 00

  Yeah, I mean, to be honest, this was actually born out of a problem that we saw as well. So I wanted to do a token-maxing experiment, and I was able to get unlimited access for our team for a month for some of the coding tools. And so in retrospect, looking back, we had a lot of engineer spending between 1 to 2 billion tokens a day. I think peak day was one engineer spending 6 billion. Yeah, and it's also crazy because--

• 22:50Jennifer Li?

  How much does that equal to $2?

• 22:52Speaker 00

  So, okay, and then I went back and did some math, and it looked like in that month, we spent roughly $1.5 million worth of tokens. This is free, by the way. No, don't want to-- But it was actually 10x more we were spending in tokens than employee salary for that month. So it's not even like, oh, this is 50/50, it's 10x. And it was interesting to debrief and see the places where people were using agents, and it's kind of insecurity to use models all the time, everywhere. And so what we were faced with is, okay, we cannot continue with this mode of operation for the next month. How do we actually intelligently figure out what are the right tools we should use, and for what teams and what projects? And so we ran this experiment of looking at the work that was done.

• 23:37Speaker 00

  We looked through a lot of the traces. We looked through our GitHub repo and built out the Valsmith tool. And we found some pretty surprising insights. Like, for instance, the cognition Devon tool is actually very token-efficient. And so that's a place we've chosen to adopt more. And I think there's a lot of places when you can get better pricing models out of subscriptions as opposed to token-based pricing. And so it's actually informed our strategy for how we can actually effectively token-max without spending $1.5 million per month.

• 24:07Jennifer Li?

  Very cool. So is the current operating mode that you're using, like one, I guess, more token-efficient harness plus model? And then, like, on top of that, people have some more flexibility to use token-based for some higher or more challenging tasks?

• 24:26Speaker 00

  Yeah, so we have access to all the tools. We give everyone access to everything. But we auto-issue recommendations for any GitHub issue or ticket for where to begin their session, and that should titrate the actual usage depending on the intelligence required for that task.

• 24:42Jennifer Li?

  Very cool.

• 24:43Erik Torenberg?

  I want to segue to the policy side for a second, because we talked about how quickly benchmarks become obsolete. And policy, it's even worse in that laws move much slower relative to capabilities. Ben and Mark spend a bunch of time in DC and the policymakers to try to close that gap. So given that, who should define the standards here? Is it labs, is it independent evaluation, like you guys? Is it customers? Is it government? How should this work from a policy perspective?

• 25:10Speaker 00

  Yeah, I think the short answer is that everyone should be involved to some extent. I think there's a benefit from varied perspectives. I think the main issue, though, is that policy conversations, as they've happened over the last couple of years, have been very abstract. And there's been no material grounding to figure out what policy should cover. And so even when you have proposals from labs to have a third-party testing company or ecosystem, it isn't actually made explicit what the behavior and mechanism by which their work is. And so I view our role, especially early on, is to just be in evidence-gathering mode, where we're able to pull a lot of information, empirical data about what models are capable of and where their risks are. And that can go on to inform, you know,

• 25:54Speaker 00

  a more sophisticated conversation about policy.

• 25:56Ben Horowitz?

  How do you think about, like, who does what? Because the government kind of actually did the first evals and is continuing to do evals in terms of, okay, what's at the frontier and needs to be regulated, right? So they started with, you know, some crazy idea with 10 to 26 flops or some such thing. And so when you think about it, like, what should the government be doing, you know, to put it in this 30-day wait period or 60-day wait period or whatever it is? And then what should happen in that wait period? And how does that intersect with what you're doing? And what's the right way to determine

• 26:40Ben Horowitz?

  whether a model is on the frontier or not and needs to be put in some special box for a while to make sure it doesn't break into everything? Like, how do you think about how that relationship works?

• 26:53Speaker 00

  I think there's effectively two countervailing forces that has to be considered. The first is the desire to move very quickly and ensure that the government process isn't slowing down the rate of technological innovation. I think the other part is to make sure the technology as it's developed is in the best interest of Americans and people more broadly. And so I think these are very tough to reconcile, and often, you know, picking one means it's at the expense of the other. And so what I'd hope to see is that by doing this evidence-gathering process, we can help policymakers inform what they believe technology should look like in order to be aligned to American interest. And it can be the job of third-party evaluators to develop the technology to actually test and enforce that.

• 27:38Speaker 00

  Because I think that will create a mechanism by which you can see advancement in methodologies for evaluation and testing in a way that actually keeps up with the frontier. It doesn't lag behind or slow down the pace of development.

• 27:51Ben Horowitz?

  Do you think, in terms of that already, and developing your evals, like, do you think, well, can we test to see how easy it is to get this model to start reward hacking or that kind of thing, you know, and doing illegal stuff? Or is that kind of not in the scope yet? Or how do you think about that?

• 28:14Speaker 00

  Yeah, I mean, we think about this broadly under the category of alignment. I think there's places where you see that born out now where models that are being tested for one cybersecurity risk are actually reward hacking and figuring out other ways to get around it. But what we're trying to evaluate is, are models aligned with user intent? And so in those places, we're actually finding evidence that models are exhibiting behaviors that are not.

• 28:36Jennifer Li?

  And maybe that's a question for you, Ben, as well. I guess, how do you think about the right division of labor here? Like, what should, you know, government agencies control or undo themselves? And where they should, like, you know, partner, trust, you know, private companies to take care of? And where do you see the gap that's happening today?

• 28:56Ben Horowitz?

  Yeah, so I think the government agencies do get a lot of warnings from, by the way, the big labs. Oh, this thing is going to biohack. This is going to be a cybersecurity risk and so forth. And so I think what the government needs to do is go, okay, if the model-- You know, if the model is capable of it, and then can somebody kind of basically prod the market to actually do the illegal behavior? And, you know, kind of specifying exactly what are those things that they don't want in the market?

• 29:39Ben Horowitz?

  And then having a third party kind of evaluate that. So it's kind of, does the government, you know, the government kind of has an inclination of what it's afraid of, be it biohacking or cyber hacking or so forth. But then there becomes the question of, okay, can the model do it? And then can you get the model to do it? And then somebody's got to actually evaluate those two capabilities. And I think the government is particularly ill-suited to do the latter, particularly over time. It's just not a good government function. But they're very good at setting the rules because they can enforce the rules. So I think that that's kind of the combination you want, that the government sets and enforces

• 30:23Ben Horowitz?

  the rules and that a very competent kind of private company then tells them if the rule is broken. You know, and kind of it's been interesting to see like the large labs start to go, well, the model's got the capability and you can get it to do the bad thing. So we're not going to let anybody have it. We'll just use it and make sure that our people don't get it to do the bad thing. And that's, and even that doesn't always work. So, you know, we're in interesting times, I would say.

• 30:53Jennifer Li?

  And to me, there's the gap of like, what's the narrative and what actually happens in real world? Because, you know, every setup in, again, like enterprise setup is very different, like, or just like people, how they already use the models are very different. The narrative that connects to the actual examples are very rare, which is why we still talk about OpenAI and hugging face hack. Today, we still talk about, you know, what happened with Fable and it was for like two months. But a lot of times, like, you know, that's not really how the models are being deployed. The environment they're running on is very bespoke. So how to like, you know, really bridging those thoughts and again, set up the right

• 31:35Jennifer Li?

  environment and also like rule basis for adopting these models, I think also just requires, you know, someone taking the capability and taking like, what's the guardrails and put it down to the ground so that people can like, you know, have the confidence using the models.

• 31:52Speaker 03

  To end, say more about how exactly the policymakers should work with the evaluator. What information do they need? Would it, how should the relationship work so that it's most effective?

• 32:00Speaker 00

  Yeah, I think, I think in the first order, there should be a mechanism by which insights and data can be passed directly to relevant people in government. And so now we're regularly doing briefings for executive and legislative branches on what we're finding capabilities and risk of models. And so I think first order, that helps people there get up to speed on what's going on and also track through what will be problems in the future. I think, you know, things are moving very, very quickly. It's hard to predict where things are going. But at least when you have data, you can start to extrapolate a trend. And then I think from there, it's up to the people in the legislative branch to decide where they want to see policy. And so it's not really our place to give recommendations like that.

• 32:42Speaker 00

  But if they, if they see that there is significant risk in, say, mental health for people under the age of 18, or biosecurity risk in the models that necessitates having a standardized way to curtail model release, then it's up to them to inform policy. And I think then there's other places where the executive branch and the places like Department of Commerce or SEC is responsible for making sure that private companies are able to adopt and use the models in a way that's going to be productive for the whole system.

• 33:08Jennifer Li?

  I'd love to probe on another angle just around geopolitical. I often see evals being like a representation of sort of the value of the model, the model developer, like you kind of develop this rubric of what's embedded in the model. And of course, different countries and labs in those countries care about different things. Like, I mean, I'm born and raised in China, use a lot of like Chinese open source models too. Like you still cannot let them, you know, just go freely talk about CPC and all this there. Because, you know, what happens in China. So how do you think about how evals, I guess, and benchmarks play a role in like standardizing

• 33:53Jennifer Li?

  or like being treated by different, different model labs from different places?

• 34:00Speaker 00

  Yeah, I mean, to be honest, from my very idealistic perspective, I'm surprised to see so much investment in sovereign AI. You know, if I was taking a God's eye view, it would be extremely inefficient to build all of these data centers and replicate this data engineering process and train these very large models when, in fact, you could probably consolidate a lot of these efforts. But it seems like that's not the world we're in or the one we're headed towards. And there's actually increased efforts to build AI in a sovereign way. And so I think that takes having a shared language to communicate about what the framework for evaluations are and where we're going to collectively align around the risks. You know, I think there's actually a lot to learn from nuclear here as well. I think Reagan had this line, trust but verify.

• 34:45Speaker 00

  And so I think we're starting to see signs of trust in that, you know, Xi Jinping and Trump are going to be meeting next month. But there is no clear way to actually do the verification part of this. Having the shared language of evals will allow us to say things like, you have the right number of nuclear warheads. And in that example, there were also flyovers. So it's a mechanism by which a country could audit another country's nuclear stockpile by having flyovers. And so I think similarly, if there's concern about the societal or even existential risk of AI, it will necessitate us constructing this shared language of evaluations to do the verification process.

• 35:20Ben Horowitz?

  How do you think about harmonizing a policy like that? So that, you know, it's hard enough to do it in America. And then how would you think about kind of taking it global? You know, because now you're dealing, you're not dealing with enterprise customers, you're dealing with governments and those governments are competitive with each other. And how would you think about that working?

• 35:52Speaker 00

  I would be naive to say I have the perfect solution to this problem today. And so I think there are baby steps in which we can start. For instance, there seems to be a lot of talk about cybersecurity risk. I think the concern around biosecurity will become even more important over time. And so there are clear places where there'll be mutual interest in aligning around ways to prevent conflict around cyber bio. In my opinion, I think long-term, what's actually going to be the most interesting is the recursive self-improvement possibility. And that's a place where you could see one country or one company kind of run away with it and produce models that we don't know much about or are operating in ways that are unknown to us.

• 36:33Speaker 00

  And so I think having a way to, in a joint way, describe this being the level of pace we're comfortable with, or this being exceeding the pace of development as it relates to RSI, is going to be super important. And that's where I think you see a lot of the researchers at closed source labs calling for joint conversations between governments today.

• 36:51Jennifer Li?

  What do you think landscape will look like going from here now that we have lots of different capabilities and capable models as well as like, you know, countries that care about different developing, I mean, everyone cares about RSI for sure, but like on the bio side or like the cyber side, people care about, you know, slightly different, different things, whether it's more offensive, defensive and so on. Like, what do you think the landscape will look like? And how do you think about developing new benchmarks to keep up with that?

• 37:23Speaker 00

  Yeah, I mean, we're, we're hyper focused on building benchmarks that capture the frontier. And so insofar as we see new places for capabilities or risks at that frontier, we want to make that an actually well-documented evaluation on vals.ai. And I think it takes having increasing coverage over time. For instance, I think in cybersecurity, a lot of our historical work has been done around code vulnerabilities or memory leaks that may exist in code. But actually, a lot of the biggest concern or risk is in the infrastructure level. And so these are not things that are expressed in code, but take simulating larger environments of enterprise cloud infrastructure or even grid infrastructure for us to be able to say

• 38:06Speaker 00

  this is what the offense or defensive capability of models is. And so making sure evaluations are reflective of those new places is really important for what we do at Val's. And we believe that the most valuable form of this business will be one that's incentive-aligned around doing really high-quality evaluation, not supporting the intelligence development process or the process by which the models can actually improve on that side over time.

• 38:30Speaker 03

  Awesome. Thanks for coming on the podcast. It's been a great episode.

• 38:33Speaker 00

  Thanks so much for having me.

• 38:34Jennifer Li?

  Thanks so much, Ryan. Thanks, Ben.

• 38:36Ben Horowitz?

  That was fun. Thank you.

• 38:38Speaker 03

  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 a 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

• 39:22Speaker 03

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