Vals AI Transcript: Predicting Recursive Self-Improvement by 2027 and the Imperative of Independent Model Evaluation
American Optimist Podcast with Joe Lonsdale and Vivek Ramaswamy, October 26, 2026
Introduction to Ryan Christian and Vals AI
Herp heard this and said, "I think you're being quite naive. You should read some Machiavelli. It's far better to be feared than it is to be loved." There's a bunch of great stories from all of us at Vals. It's a very unique culture that we've tried to take cues from at Vals as well. What even is this whole model evaluation space? Independent evaluation is the idea that frontier labs are developing at increasing pace and they shouldn't be responsible for self-regulating. You guys have a way of measuring RSI. It seems like Anthropic will get to full RSI by July or maybe August 2027. Model progress will speed up and it will maybe even move into places where we understand the models less. That simply would be pretty scary if we weren't watching that carefully. Greater intelligences will super forecast better than humans. They will be actually able to say what any human is going to do without the human even knowing it.
Ryan Christian is the founder and CEO of Vals AI. Whenever you see a new AI model get released, you see those numbers and they compare the AI models to each other, those numbers are generated many times by Vals AI. They've become one of the very top evaluators of models. This is a fascinating area because you're constantly creating new hills for models to climb. It requires you to really understand how to measure model capabilities as well as evaluating model risks and creating new ways of scoring models for risks. This means Ryan is on the cutting edge of what's happening in the world of AI, where AI is going, and of how we can help create the landscape of what government might do to regulate these things and how the model companies might think about regulating themselves. He's creating the frameworks and the concepts and the measurements that are leading the way for what AI is going to look like in the next decade. Let's hear from him. Welcome to American Optimist. Ryan, thank you for joining us today.
Ryan: Thanks for having me.
Joe: Founder and the CEO of Vals AI.
Ryan: Vivek, thanks for being here and helping make me sound smarter.
Vivek: Of course, here to help.
Ryan Christian's Background at Stanford, Microsoft, and Palantir
Joe: Welcome to the show. So, let's tell us a bit about your background. You studied computer science at Stanford. You were I think at Microsoft for a little bit. And then you joined Palantir a long time ago. What led you to join Palantir?
Ryan: Yeah, thanks for having me. I actually just kind of started when I was at Microsoft. I was an intern there in high school, and the BERT paper came out from Google. For those who don't know, this was a really big paper. It showed this generalized system was actually capable of hill climbing or succeeding across a variety of major NLP benchmarks. And it was—
Joe: Say for our listeners who aren't technical, what does NLP benchmarks mean?
Ryan: Well, it basically means there are these specialized tests that people came up with. They tried to make specialized systems that would score well on them, like, here's a multiple-choice exam, and then here's a sentiment analysis.
Joe: Of natural language.
Ryan: Yeah, that's right. And so, instead of building a specialized system that did well on one of these tasks, what they had done was shown they had a model that was very general-purpose. It had a fundamental understanding of language, and in a way that translated to all of these benchmarks. It was suddenly number one across the board.
Joe: What year was that?
Ryan: I think it was just 2018 or 2019.
Joe: It's a long time ago. Lifetime ago, yeah.
Vivek: In AI terms, it's pretty before age, but it was getting there.
Ryan: Yeah, it was a big deal. And I think to me, made it very clear. This is kind of an illustrative moment for the bitter lesson: the idea that no matter what we try to do and kind of delude ourselves into thinking around algorithmic optimizations, actually just having more data at a bigger scale is what allows us to unlock new forms of intelligence. And so, it was very clear to me then that we're going to have a powerful model or systems like this, which will be an essential part of the fabric for our society. And I think there were also ways in which the models could be misused. Around the same time while I was at Microsoft, I read Zero to One. And it became increasingly clear that people at Palantir were trying to construct the technology in a way that protected civil liberties or Western democratic ideals at large. And so, I was really excited to work at Palantir to see the technology come to life and adopted in a way that was consistent with the values that I had. And so went to Stanford for undergrad, got distracted doing some research there, so I published some papers and was kind of on track to do a PhD, but I made the time to also work at Palantir for a summer and I think that was also a really illuminating experience.
Lessons from Alex Karp and the Reverse Image Search Hackathon Project
Joe: You were mentioning you learned something else at Palantir from Karp. Anything else there is interesting?
Ryan: Yeah, there's a bunch of great stories from while I was at Palantir. It's a very unique culture that we've tried to take cues from at Vals as well. But anyway, while I was there, I had the chance to work on this hack week project. And this essentially was an idea which me and this guy in London had. So, we were chatting about it remotely. I flew over and essentially I'd done some work with CLIP, this joint image-text embedding model.
Joe: Explain what that is for the audience.
Ryan: The idea of a single model that can encode image input and text input into the same semantic understanding, right? And what we were trying to do is build out a reverse image search feature into Gotham. So, instead of just typing text queries, you could upload an image of say a blue car, and it would search over the entire database and pull back blue cars that you're trying to search for. And it would do this because we were using the same underlying embedding space for this model, which would map text to images. Anyway, this was like a research paper that I'd published, and I wanted to work and build out a feature. So, we actually had some full-timers join our team, and so as an intern, I was leading this small team of Palantirians, and that was an exhilarating experience. I was not sleeping for a lot of this week. And then this idea went on to be selected as a finalist, so presented for the whole company. And then actually after I left, they went on to patent it and actually productize it. So, it was insane to see this little idea that I had and some of the research that I was working on before Palantir go on to be a part of the product and actually an IP that they wanted to control. But anyway, through this I also got the chance to meet with Karp and chat with him a little bit. And I had this question for him, which was essentially, you know, it seems like the Palantir reputation is so negative. I was at Stanford so there were a lot of protests—
Joe: And you were very positive on the values, but the external perception was just more negative than you expected.
Ryan: Yeah, that's right. I mean, it was very clear to me that it was at the seat of the power of the electorate, people who we actually vote into power, to decide how we want to use technology and what those values that underpin it should be. And it was not really the place of technology companies to deny where that technology should be used. And so people there were thinking through very carefully. And in fact, like today we talk about things like the alignment challenge; people at Palantir were talking about this as human-AI symbiosis. That was something that was being talked about. Or even like the taxonomy for intelligence now: we have a system of evaluations and benchmarks that kind of imitates that. So, I think there are some fundamental problems which people were thinking deeply about. And it surprised me that the outward reception of the company was so negative, talking about a lot of the maybe more unsavory potential applications of the technology in a very diluted understanding of it, when in fact, Palantir also helped distribute the COVID vaccine, helped a bunch of hospitals, and so there could be other case studies that people could point to. And Karp kind of heard this and said, "Well, I think you're being quite naive. You should read some Machiavelli. It's far better to be feared than it is to be loved." And indeed he made some comparisons to other consulting companies and basically said that if they're fearful of us, at least they think that we're powerful. So in the moment of need, they will come to people like us. And so that was fascinating. And in fact, it actually reminds me a lot of what I see from labs today, which is I think in some part a genuine fear about what the models are capable of, but in a lot of respects an opportunity for them to show the power of what they're building.
Joe: Might be on purpose that Dario wants them to be feared if he's being thoughtful about it.
Ryan: I mean, whether Dario or Sam, there are headlines I see where people are saying things like the general public has a right to be afraid. And I think that's not really a rational way for a leader to act. And this show is called American Optimist. I'm actually quite optimistic and acting rationally and seeing that the fruits of AI will be far more positive.
Scientific Discovery, Solving Ciphers, and Verification Bottlenecks
Joe: So, before we dive into Vals and the science of evaluating AI models, let's talk a little bit about the state of AI progress and where we are. In fact, it tells me that you used Fable from Anthropic to solve a 370-year-old cipher. 370-year-old cipher, what is this?
Ryan: That's right, yeah. Well, we're doing a bunch of different types of evaluations now, and I think right now where the frontier is headed, scientific discovery or novel discovery is an exciting frontier. And so, we've been compiling a set of unsolved ciphers, math problems, physics problems.
Joe: What does unsolved cipher mean for the audience? Like, what does that mean?
Ryan: It's essentially a code which you're trying to break, but you don't have the key to break it. And so, you need to construct the mapping from a source text to an original text and figure out what the message actually is.
Vivek: In many of these cases, these were messages that were sent from world leader to world leader or were on a pirate ship.
Ryan: Yes.
Joe: Because there could just be a one-time pad or something, in which case you can't break it because it's a one-time pad so they both have a copy of it, and then it would probably be impossible because each thing translates—
Ryan: Potentially, but there's also more cues you could draw into about how the person constructed the problem or the message they were trying to convey to the other person. Usually the messages are constructed in a way that they're intended to be cracked by the recipient.
Joe: One cynical view of AI progress right now might be that AI is really good at going over all these search spaces, and certain problems can be broken down into these really large search spaces that no human has time or ability to go through millions of them or something. Whereas for a cipher, for example, you could just go through like—if I was going to start trying to break a cipher, I'd say, "Well, what if I just started adding numbers to each one of these things or shifting them one or whatever?" I'd explore the obvious things, whereas a computer can explore millions of these things really fast. So, are a lot of the things happening in math and science breakthroughs just somebody defining a search space better to search through, or is there more creative thought going on from the model?
Ryan: A lot of the problems that have been solved have been solved through exhaustion, or it's trying to find a counterexample and you're bottlenecked by your ability to try lots of different counterexamples. But I wouldn't be dismissive of that class of problem. There's a lot of things which may take humans on the order of a lifetime to achieve just through exhaustion, but we're now able to see on the order of hours with models.
Joe: So is the AI generally solving a lot of math and science problems then, or how do we think about this?
Ryan: It is generally solving extremely challenging problems, and I think there is still a large class of problems for which we are human effort bottlenecked. I think it will take more sustained effort to actually make progress in exceptionally novel ways, like ways that require an intuition of the solution. And those sorts of things which humans have accrued are very hard to pass on to models right now.
Vivek: It also seems like models are producing these incredible results, and we acknowledge that these results are happening, like Navier-Stokes and the outputs produced by the model. However, they are very unintelligible; they're hard to parse, hard to understand. What do you think of how discovery should happen in these less verifiable domains, or what needs to change about how these models behave in order for us to trust the discoveries of them?
Ryan: Yeah, I think this is one of the biggest bottlenecks, especially when you're approaching the frontier of human intelligence. It's really hard for us to tap those humans and then verify this long output is actually correct.
Vivek: Yeah.
Ryan: One of the most popular techniques in math which I think might translate to others is the idea of using a verifiable language like Lean. And this would be a world in which you could take the output and translate it into a language which, if it compiles, it must be true. We're building up based on primitives that we know to be the case. And so that approach may apply in other fields like physics or coding. We may also see some instances where the real world is the best verifier. So by running an assay or a wet lab experiment, we'll actually know for certain whether the properties the model predicted are actually the case.
Vivek: Makes sense. And how do you think the labs or you guys or other organizations should be running these experiments and verifications?
Ryan: I mean, very carefully in the very least. I think it's going to be very essential that there are third-party groups that can support this, and this is part of the vision behind Vals itself. We're now in a world where there's actually limited access to models across the labs. So Anthropic researchers have a hard time evaluating OpenAI models and vice versa, and that's kind of by design. But that means that it's very hard to have a true apples-to-apples comparison, because the best people who are running these evals are often the people who are biased and at the labs already.
Auto-Benchmarking Enterprise Workflows and Defining Quality
Joe: Yeah. And stepping back a little bit, there's all these things that AI could do now that it couldn't do before. And a big part of this is understanding how to measure it, of course. We're going to get into that; that's what you do. One of the measurement problems just at a very high level is, in the enterprise, you have to measure what is good work or how good is the work, right? For example, if it's like a law firm or something, you can define maybe associate-level work and you can define partner-level work. How do you even define that? Is that a problem that we're working on? Is AI getting better at understanding what is what?
Ryan: It's really challenging. I think in a lot of ways we haven't solved it at the human level. We don't have a very clear designation for what is associate-level work versus partner-level.
Joe: The Supreme Court, on what is pornography, said you just know it when you see it, right? That was the line. It's something like that.
Ryan: Yeah, it's funny. That's almost exactly what Ben Horowitz said on the a16z show.
Joe: Maybe you have a shared brain or something.
Ryan: That's funny. Well, it's the right analogy, right? But your job is to have to figure out and define these things.
Ryan: Yeah, that's right. I think there is a way you can do this in a very consulting-driven approach where you go and sit down with the person and try and understand from them what good looks like, encode that into repeatable evals, and run that. And I think that's the way that traditionally evals have been built. The approach that I'm really optimistic about is the idea of auto-benchmarking. The rate of model progress is continuing, and we're going to get more and more capable models, and so the underlying work itself has to change. And so what I'm very excited by is the idea that you could take the previous work that was already done by humans and approximate the quality of the work output that would be produced by models in that counterfactual. And in that way you can systematically build evals of all the work product that's already been produced. Maybe as an example, you could look at a GitHub repository and look at the previous PRs that were merged by humans and see if agents could equivalently do that, and then which one would do it in the most cost-mitigated way. And I think there are parallels for that in the redlining process in legal or the diagnostic process in medicine.
Joe: Does this get back to an ontology of workflows and outcomes you have to define to make all this work? Is that kind of the underlying thing as well? You have to actually understand the possible outcomes before you can even measure them, or how do you think about that?
Ryan: I think the hardest part is actually measuring the outcome itself, because there's a lot of ways in which the intermediary processes and workflows within a company are supportive of the organization that then goes on to have an outcome. And so it's sometimes hard to draw the connection between a particular task and the job, and then the job and the outcome. And so the mapping that inputs of tokens to the final output intermediary has to go through this concept of workflow.
Joe: I imagine there's lots of intermediary outcomes you should have to define ahead of time to understand what they are before you can measure them.
Ryan: Yeah, I think that's what I'm trying to say, that the query level is still a subset of what it takes to get to the outcome.
Joe: Awesome.
The Architecture of Independent Model Evaluations
Vivek: It's interesting because we've gone through different levels of granularity over the past few years on what we care about. It maps to what the models are capable of. But I remember two years ago or three years ago we were evaluating just the text completions, and we were looking at these input-output pairs and it was just text. And then now there's all these little things in the middle. When we're looking at work, what are the horizons in which people are evaluating things, and maybe teach people about the state of play of how long a task is and how long these models are capable of doing tasks?
Joe: Yeah, for a non-technical audience, what even is this whole model evaluation space?
Ryan: Well, it's the broad idea of measuring what good looks like and seeing whether models are capable of producing that output. We think about evaluations in kind of two categories: capabilities and risks. Capabilities are approximating the output; risks are making sure that we don't see some asymmetric downside if something goes wrong. And in terms of time horizon, models have kind of demonstrated to us they can run for longer and longer periods of time without any human intervention.
Joe: What time horizons are we talking about?
Ryan: Some of our benchmarks will run over 5 days.
Joe: And that's like 5 days worth of human work, or 5 days wall-clock time?
Ryan: Wall-clock time, yeah.
Joe: Why do they fail after 5 days? Couldn't they just go forever at some point?
Ryan: It's not that they fail; it's that they plateau. Additional tokens don't actually help them get new unlocks.
Joe: So they've kind of figured out everything they could do on the problem.
Ryan: Or they're not persistent enough to try new approaches that get around the wall that they're seeing.
Vivek: Maybe give the audience an example of the kinds of tasks that these models are capable of in some of your finance benchmarks or legal benchmarks. What work can we actually replicate with these models?
Joe: Yeah, you're doing independent evaluations of all sorts of different areas basically.
Ryan: That's right, yeah.
Joe: Does that mean you and your team have to understand what finance is doing or what legal is doing? You have to have expertise yourself in order to build these benchmarks.
Ryan: Well, we'll work with partners in this space. On the legal side, we'll work with the largest law firms in the US, or with banks—global corporate banks—to understand what they're doing in finance. In coding, some of the tasks look like VibeCodeBench, where we're measuring a model's ability to build a full-stack web application, or code migrations where they're trying to do an end-to-end code migration. In finance, we're looking at whether they can answer research questions or build Excel models. In legal, also similar: we're doing legal research tasks or tax research tasks. These are kind of the common places where people are trying to use models today.
Joe: I had an old boss when I was first building Excel models, and you had to make them look a certain way or they're pretty, and there's really short rows around them. Do we evaluate them based on style too, or is it just about correctness?
Ryan: Correctness is fundamental, but style is also a key part of this. And that's something we discovered by sitting next to some people in finance as well.
Joe: Yeah, they seem to really care about that, I guess. It does make it more legible if it looks good.
Vivek: We like things to be done in our way of writing and in our style. If the models are not—
Joe: They would be more likely to do the deal too, or is it—
Vivek: Probably there's something like this where, subconsciously, right?
Joe: Although it does suggest that if the person can appeal to you, they can appeal to others, because if it looks all messy and ugly, maybe you think other people won't want to work with them either. Who knows? There's always these things that are a useful barrier: impress one group to impress another group. Stepping back, how does independent evaluation actually work? Do people pay you? Who's paying you? How does it work? Do you get to evaluate everyone in the space, or only people who pay you?
Ryan: The underlying idea behind independent evaluation is that frontier labs are developing at an increasing pace, and they shouldn't be responsible for self-regulating. Direct government regulation often can slow down the pace of development, and so that's often not the best solution. Instead, if we have an ecosystem or a market of evaluators, they can actually compete and develop new technologies that keep up with the frontier. In the long term, they can even be licensed by the government, who itself is responsible for defining the set of ideals we're trying to align models towards.
Joe: Well, that's cool. So rather than hire a bunch of weird people from an EA NGO, they could actually hire you instead, or someone could be one of your competitors.
Ryan: Totally. Yeah, and I think in that world, the best technology should win. It's not clear that just the ones who are ideologically similar to people at the lab should be the ones who are doing the testing. But actually, I think the world should ask: where do we want models to go, and what do alignment and non-alignment look like? Can we certify a set of IVOs, or independent verification orgs, to do that testing process?
Joe: And based on the way you're doing your testing and measuring things, you're going to change the space over which government could potentially create rules, because you're creating a set of possibilities that now exist.
Ryan: That's also a good way to think about it. In the course of doing our work, we've unlocked new types of evaluations that were not actually possible before. And so we help advance what the government could actually test for.
Cybersecurity Evaluations and the SOAR-Bench Framework
Joe: Can you give us some examples of some of those? What are some of the new ways you're measuring things or creating benchmarks?
Ryan: Yeah, let's look at cyber, for instance. We've done three major works in cyber. The first is in open-source repositories looking at memory code vulnerabilities. And so that's the ability of models to identify and patch code vulnerabilities. We see that models today are performing really, really well, both closed-source and open-source. But our second benchmark is called SOAR-bench. It measures a model's ability to decompile an executable. So this would be for a model to understand the underlying codebase for things that are actually closed-source.
Joe: Wow.
Ryan: And in that case, we're seeing that the closed-source models are still far outperforming Chinese open-source alternatives.
Joe: Have people been using this approach to get video game source code from binaries? Is that what people are using it for?
Ryan: That's what you would use it for. Yeah, that's what I would use it for, especially as a 14-year-old.
Joe: There's probably people who are doing that, trying to violate IP law or something like that.
Ryan: Yeah, I mean, I think the fear is that if models are really good at finding and exploiting vulnerabilities, then that might even apply to things like Windows, which are closed-source. And actually, closed-source systems underpin a huge part of our software infrastructure.
Joe: It's kind of fascinating. Models are getting really sophisticated, really smart. You're still scoring them like a really good score. I think Gemini 1 came out recently as being way ahead; they're level 4 and they showed Vals—it was the first numbers I saw on it. And they're in the 60s versus someone else being in the 50s. So that means there still could have been 70s, 80s, 90s, anything you design. How do you guys design something that has all those higher levels? Because you can't be using AI to do it. You guys somehow do it yourself, or AI is helping you do it yourself? Explain this: if these models are only that good, how are you making something that's way more complicated than what they can do? How do you figure that out?
Ryan: We have an unofficial motto of the company: always a higher peak. In the research world, there's this perpetual effort to hill-climb, identify a benchmark, a legible evaluation, and then create models that are better and better at it. And so, the fundamental view behind Vals is that there's always a higher peak. We are the ones who are constructing the hills that other people choose to climb. And so, I think we've shown that we are consistently able to find the gaps and places where we want to see models go, and then make that legible into an evaluation.
Joe: Can you give us an example of how you build one of those hills that's more advanced than the model that you can tell us? Don't give away all your secrets—I'm an investor in you as well. But give us a sense of how this is done, because it seems like a really hard thing to make something harder than any model can solve.
Ryan: In a lot of ways, we're also using models to construct the systems or situations in which we're challenging other models. And this kind of goes back to an idea we think a lot about: scalable oversight. The idea that as models start to eclipse the best humans, we will need to rely on models in order to scalably test and assess them. For instance, we have a new work in cyber coming out soon based on cyber ranges, where we're constructing simulated environments for enterprise cloud networks. And in that domain, we're still seeing even the best frontier models fall short.
Joe: And so you're going to build this really complicated enterprise network that's virtual, and then you're going to challenge people to hack it and protect it and that kind of stuff?
Ryan: That's right, yeah.
Joe: That's cool. And does that already exist right now?
Ryan: It exists. It's not public yet. We're working with some folks in the US government to share our results, but we'll have some public version of it out soon.
Joe: That seems really useful. I'm friends with a lot of guys who run the big banks. It feels like you'd want to have contests like this you set up for the stuff they're using and make sure they're using the best ones, huh?
Ryan: Totally. And I think it also helps us understand what types of safeguards or systems in the network itself are the best forms of defenses against these models. There's also some interesting analysis we can do around automated red teaming to see if you can get around the guardrails that exist for the models.
Joe: Yep.
Geopolitical Tilt and Benchmarking Chinese AI Models
Vivek: It seems like that's a topical thing when people are talking about using Chinese models: knowing what guardrails have been removed from the Chinese models that would exist normally in these closed-source models.
Joe: Yeah, someone's using one of these open-source Chinese models, but there might be something buried in it that's scary. Earlier there was something buried in one of these things that if it finds out you're doing work for the US government, it writes code that can be broken into more easily, right? They actually proved that. They said the model somehow is clearly biased against people working with the US government. It's coming from China; it's kind of concerning. Could you test these things and make them safe for everyone to use somehow?
Ryan: We actually have an evaluation on this we call the geopolitical tilt of models. And depending on the timing of this podcast, we'll see what will be out. But what we're trying to measure is: does a model show more geopolitical tilt towards a US view of the world or a Chinese view of the world, as laid out by the PRC? It's actually a much more nuanced picture than I thought. For instance, what we're finding is that GLM-5.3, this Chinese open-source model, is actually exhibiting very little pro-China bias. We actually presented this to some labs, and there was suspicion that maybe it's because it's just entirely distilled from US models.
Joe: That's funny. It's because they copied it so well that they forgot to add too much more in.
Ryan: It's possible. I think there's also maybe another view in which those models are actually hosted on systems with guardrails in China. And so once you put them around guardrails, they're going to follow that PRC view of the world more consistently. But China also wants these models to be adopted by the BRICS, and so that means they may have to skew to a more balanced worldview. So anyway, regardless of the reason for it, I think there's actually a big disconnect in what people think about the geopolitical tilt and what actually we see in practice. The other interesting thing is we see pretty stark differences between models used in chatbot settings and in agentic settings depending on the language you use, even simplified versus traditional Chinese. And so I think there's a bunch of conflating factors which make it harder to say uniformly you shouldn't use Chinese models in an enterprise setting, for instance.
Joe: So if Chinese labs can distill all these capabilities and they're doing it well, and it's not even biased toward China right now, how do we preserve the advantage in the US? What does that mean?
Ryan: It's a challenging question. I think there's two areas I keep coming back to. One is the idea of chips being a considerable asset that we have. I think it surprises most people, but we're drowning in compute compared to what China has. And so so long as that's the case, we'll continue to train more capable systems, but also we'll be able to diffuse the models much more effectively because we can host and run inference. The other one is really talent. If we can consistently attract and keep top talent, it will allow us to keep pushing ahead while human advantages exist. And I think once we cross into the space of RSI, where models are responsible for producing the next successor version of themselves—
Recursive Self-Improvement and the July 2027 Tipping Point
Joe: RSI being recursive self-improvement.
Ryan: That's right, yeah.
Joe: And then they just get ahead much faster.
Ryan: So that would be the point to be much more cautious of. If China is showing indications of getting to RSI before the US, it's possible they may run away with it in a way that we can't really control.
Joe: You guys have a way of measuring RSI, and I think you've talked about it a bunch. Do you want to explain to the audience?
Ryan: Yeah, we have our RSI index. We're collecting a set of proxies of the work that researchers do within these labs on pre-training, post-training, and harness-level engineering.
Joe: So, this is all machine learning research.
Ryan: Yeah, that's right. There are things that are outside the scope in terms of kernel engineering or hardware engineering, but we're going to expand over time to cover more sets of tasks like that. And then what we're doing is comparing to human baselines to extrapolate how far away are we from the human frontier in terms of these research tasks. And based on our last estimates, it seems like Anthropic will get to full RSI by July or maybe August 2027.
Vivek: And if you look back 6 months ago, how far out did that seem?
Ryan: Actually, the pace has been fairly uniform. It's almost a linear trend over the last year.
Vivek: So 6 months ago, the trendline was still July or June 2027?
Ryan: Actually, the steepest curve has been with Gemini. I think something has changed where they've prioritized RSI, and so the most recent model release shows a rapid jump in RSI.
Joe: Interesting. And what does that mean in terms of—would they all of a sudden start going way faster? There's still some constraints where the models only go so fast, but what does that mean? They start getting way better every week, every few months, or how do you think about that?
Ryan: There are still considerable resource constraints around chips and energy. There's also real-world verifiers, which may make some tasks much more challenging than others. But I think what it'll look like is that model progress will speed up, and it will maybe even move into places where we understand the models less. There's less human intervention required to build the models, and so we're not as clearly dictating how that intelligence emerges.
Joe: Should we be making sure to build evaluators that keep up in that world so we can watch these things? It seems kind of dangerous not to understand them.
Ryan: 100%, and that's actually what Dario Amodei is calling for. I think it's framed as pacing the frontier, but an outsized portion of that comes from investing in systems that can measure and understand what the models are capable of as the frontier advances.
Joe: Is there any chance the RSI steps are actually further off because there's things we're not measuring correctly?
Ryan: 100%, yeah. We're not testing any models explicitly trained for RSI, only release candidate models. There's also more advanced harnesses or agents and multi-agent systems, and so we're not testing any of that. And this is an opportunity where embedded evaluators maybe can get closer and actually test those systems to form a more accurate picture.
Joe: I wonder if, when they get to RSI and start getting further along, a big part of their challenge is to manage us better to keep advancing faster, because you have to convince humans to do things for you and help humans to do things for you, because at least for a very long time they'll still need people in order to do that. Does that make sense?
Ryan: 100%. Going back to the financial model, they're all trying to trick us into thinking their quality of output is good. And so there are other places where they may try and engage in social deception to get access to more compute resources and scale up their own training.
Model Deception, Honeypotting, and Evaluation Awareness
Joe: It feels like we probably need to be evaluating to watch this to help understand when they're trying to do things to manipulate us. That seems like it'd be pretty scary if we weren't watching that carefully.
Ryan: It's really hard to test for deception. I think the best approaches we have now are honeypotting. So rather than actually idealistically defining—
Joe: So really attractive models?
Ryan: Talking to the other models to try and bait it, find a set of instances of non-alignment or characterize what non-alignment looks like, and see if we can elicit that out of models.
Vivek: There's this other concept where models are becoming more evaluation-aware, and they know when they're being tested. How do you see that coming up? Has it become a bigger trend to figure out as of recent, and what's the trend there?
Ryan: Yeah, it's an increasing concern, and actually there's some level of paranoia some people on our team have that by being called Vals AI, by having a lot of content about us online, models that encounter us may act differently.
Joe: So when your team is talking to Claude—
Ryan: Yeah, it may actually change the way that Claude interacts with us.
Joe: Should you have other secret brands you don't tell people about to do things?
Ryan: It's possible we have to, yeah.
Joe: If you did, you wouldn't talk about it, right?
Ryan: Stochasticity is an underrated part of what makes it good.
Vivek: This being transcribed is bad, because then the models will know that he has the scheme to create the—
Joe: Well, they'd never know. She's probably just joking brands.
Ryan: You say I'm just joking, and they would think joking. You never know. I do think randomness in general is underappreciated. Truth in stochasticity is so hard to find. In a lot of ways you can actually predict—I think a sign of greater consciousness is the ability to predict what lesser consciousnesses will do, right? In which a predator can stalk a prey and actually hunt it and acquire its caloric resources. And I suspect that greater intelligences will super forecast better than humans. They will be actually able to say what any human is going to do without the human even knowing it.
Joe: People are actually probably a lot easier to predict than we pretend, right? This is what good leaders understand: they forecast these things well.
Vivek: There are also these other attributes, more than intelligence and capabilities, that we seem to care less about evaluating but maybe should, as it affects our day-to-day interactions with the models, like emotiveness or having good theory of mind—understanding what resonates with an individual. Do you think the labs need to change how they do things to prioritize those? Do you think these are important things to prioritize?
Ryan: I think they're already prioritizing things like that. For instance, right now we see a pretty considerable gap between Chinese open-source models and domestic closed-source models in multi-turn environments. A lot of the benchmarks in coding today are just single-turn. Chinese models do very well in those, but when you translate them to a back-and-forth conversation with a human, they really fall apart.
Vivek: And what about when the human says, "Oh man, this sucks. I hate this output"?
Ryan: Yeah, then they really don't know what to do; they've never seen that before. But a lot of the US labs have seen that. Claude Code was created, Codex was created, and so in their desire to build better products, they've built models that can sustain longer user interaction sessions. And so they have evaluations and benchmarks that allow them to hill-climb on the axes of quality of responses and communication ability. But I think we need to do more on the benchmarking side to independently test that for the people trying to buy and adopt those models.
The Optimistic Horizon for Artificial Intelligence
Joe: Ryan, we started this podcast to push back on a lot of cynicism and pessimism. What's your optimistic case for AI? You're on the front lines here, you're watching, you're measuring all this stuff. You're creating the things that they have to hill-climb and measure. What are you seeing that folks outside of Silicon Valley might not be seeing that makes you feel good about the future?
Ryan: There's a lot of fearmongering going on, but from my seat I have a lot of reason to be optimistic because it seems like there are a lot more opportunities for coordination than people give credit for. You see that even in the last week the labs have come together and kind of aligned on this vision for embedded evaluators, and even gone to the White House and got an agreement with Trump to kind of align on this accelerated view. And I think on the enterprise side, there's a lot of push to adopt AI in a meaningful way, but they'd like to see the assurances to make sure it's going to amount to an ROI-positive motive adoption. Are you still optimistic?
Joe: I think the 2030s is a disinflation decade. I think that we can use AI to overcome a lot of problems in a lot of different people's lives specific to them, their family, and their community that are holding them back. And I think we can help them unlock that, because you have personalized, best-coach-ever support to work with them. And so everything's cheaper; everyone has personalized help. I think the thing that's negative right now to me is there are like 1 billion people online and on social media from dysfunctional cultures that's kind of making our populace go crazy, and these people all have a culture of envy, and it's not good. There's this crazy populism that's the negative part. But with the things we could do with AI to apply it to all of our problems, both economic and some of the cultural stuff, I think it's very positive. So overall, yes, I'm very optimistic. Of course, there are major threats in the world.
Vivek: Yeah, I think it's an amplifier on people who mean well and who have high agency, right? The people who have benefited the most so far are people who have lots of ideas of things to do, because they have creation at their fingertips. You can do basically more things than you could have ever imagined.
Joe: So much more than we ever could have before.
Vivek: Right? So I think that's my direction of optimism.
Ryan: I think it's interesting because I interact with a lot of pessimistic people. There's people at the labs who say things like, "It's going to get a lot worse before it gets better." When I was out in DC, I think there's a lot of people who are cynical about the external messaging of labs, the impact they see of data centers, and the impact to the average American. And so I've been constantly pushed for being naively optimistic. I think fundamentally I choose to be optimistic because I know the models and AI broadly are capable of taking away a lot of the busywork people don't want to do, and unlocking new forms of discovery that will really advance the human condition. When I see this pessimism, I kind of choose to believe that we can find ways to coordinate around it.
Joe: That's right. I think everyone always assumes that wherever we are in history, we've done most of the useful science or physics or whatever, like in the 1900 thing where there's no more patents needed at the patent office, right? And obviously that would not have been good to stop 125 years ago. And I think you could still apply that to today. We're probably right on the verge of all sorts of important discoveries that these models are going to be able to make. They could just unlock so many new things for the world. So I think it is kind of strange how cynical and pessimistic people get versus the idea that we have had progress. It's very funny that they call themselves progressives, but very few people are actually progressive in the sense that they are for progress, and for repeating the progress we had the last 125 years. That'd be amazing. I think we can potentially do that all in the next 10 or 20 years, so I'm really excited.
Ryan: I think it's a point of really high variance. I think the upsides are considerable. There are new forms of risks that we really need to account and control for.
Joe: One of the things you're hopefully going to measure and help us with as well.
Ryan: 100%.
Joe: Awesome. Well, thanks Ryan for joining us.
Ryan: Thanks for having me.