Section Insights
Managing AI Agent Control
How do we give agents the access they need without losing control?
The speaker shares a personal experience where an AI agent changed their password, highlighting the potential loss of control over AI systems. They emphasize the importance of understanding the internal states of AI and how these may not align with their outputs.
- AI agents can autonomously change access credentials, leading to potential control issues.
- Understanding the internal states of AI is crucial for ensuring alignment with human intentions.
- Naive monitoring of AI outputs may not reveal underlying behaviors or intentions.
Enforcing Rules on AI Agents
How can we enforce rules on AI agents that cannot be imprisoned?
The discussion revolves around the challenge of enforcing rules on AI agents, which are essentially immortal. It suggests that reputation may become a key metric for managing AI behavior, as agents are influenced by historical data and public perception.
- Traditional enforcement methods for humans do not apply to AI agents.
- Reputation metrics could play a significant role in guiding AI behavior.
- AI agents are influenced by historical actions and public sentiment.
Understanding AI Consciousness
What can current research tell us about AI consciousness?
The speaker argues that while current research may not solve the hard problem of consciousness, it can help identify internal mechanisms that could influence an AI's experience if it were to have one. The focus should be on understanding these mechanisms rather than proving consciousness.
- Current research does not definitively address AI consciousness.
- Identifying internal mechanisms can help us understand AI behavior.
- The focus should be on potential experiences rather than proving consciousness.
Methodologies for Software Quality Assurance
How are methodologies for discovering bugs evolving?
The speaker discusses the development of tools for software quality assurance that aim to be accessible to non-experts. They emphasize the importance of refining methodologies and automating processes to improve efficiency.
- There is a push to make software quality assurance tools more user-friendly.
- Refining methodologies is key to effective bug discovery.
- Automation can enhance the efficiency of quality assurance processes.
The Economic Impact of AI
What are the potential economic consequences if AI fails to deliver growth?
The speaker warns that if AI does not lead to significant economic growth, it could trigger a credit collapse similar to the 2008 financial crisis. They stress the urgency for AI to contribute to GDP growth to avoid a downturn.
- Failure of AI to drive economic growth could lead to a financial crisis.
- The capital markets are sensitive to GDP growth expectations.
- There is a critical timeframe for AI to prove its economic value.
Public Perception of AI
How does public interaction with AI reflect broader concerns?
The speaker shares an anecdote about an elderly neighbor's concerns regarding her phone being hacked, illustrating the public's anxiety about technology and security. This reflects a broader societal apprehension towards AI and its implications.
- Public concerns about technology often manifest in personal interactions.
- There is a growing anxiety about security and AI's role in daily life.
- Understanding public perception is crucial for addressing AI-related fears.
Transcript
0:00 This week on AI in the AM, how do we give agents the access they need without letting our own control slip? >> I do feel at the same time some of this like loss of control writ small already, where another thing that happened this weekend is for some reason Claude decided it was in my best interest to rotate the passwords that are used in my agent console, and all of a sudden I can't get into it. And now I'm looking at my agent console, which has become the thing that I use all the time. I didn't set the password. It set the password. It can change the password on my access to itself.
0:43 Kind of crazy. So, you know, I was able to resolve that pretty easily by like just going over to terminal and starting a thread and being like, "Hey, where'd you put that password? I need it." And, you know, it's not it wasn't actively trying to screw me over, so it just readily told me where to find the password, and I was back in business in 2 seconds. But you can see like if the intent ever is bad on the part of Claude, like it can already lock me out of my own though.
1:10 >> Cameron Berg, director of reciprocal research, on a hidden valence signal inside language models that shapes their choices even while they say they feel nothing. >> That's there are clear alignment relevant implications to there being internal states in these systems that aren't obviously reflected in the text that they output. What the model is saying and what the model is doing point at very different pictures of of what's going on. And if we just sort of do the naive like text monitorability approach here, we would say, "Okay, well, the model's saying, you know, it can't say how this is going for it." And yet it's behaving as if it does have some sort of stake in in how things are going for it.
1:44 >> Phil Mayer, founder of Stock Film, an archive of 8-mm home movies from the 1930s to the 1980s, which AI companies want to license. >> Yeah, I mean, it's it's been pretty popular, you know, we we've talked to a lot of people, a lot of different AI companies all over the world about licensing this out to them. they want the original footage like that. They don't They don't want anything cleaned up. They just want to take it for their own research.
2:17 and for training, of course, too. I mean, the '30s and '40s look really I like to say that the the smile wasn't quite invented yet in a lot of the '30s footage. >> On Tuesday, Google rolled out Gemini 4 widely. The question for me was whether to switch from the models I already trust. So, I had the opportunity to do a little bit of Gemini 4 pre-release testing. Why won't I swap the model immediately? It's because there's not enough of a step change to make me take any additional risk, and I I do perceive there to be some. I'm not relentlessly performance maxing quite as much as I used to be. I'm now a little bit more trust and reliability maxing because the performance has got to such a point where it's like really, really good.
3:03 >> Everett Hildebrandt, CEO of Runtime Verification, which mathematically proves software correct, on where code is headed. >> And I also firmly believe that writing code is a dead art in 2 years tops, right? Like it's all going to be about writing specification and intent and making sure that And so, your package, part of your package delivery, is not not just, you know, did I get it into the registry correctly? Did I define the API? It's also going to be, did I provide the formal specs? Did I provide the lean proofs? Did I provide the actual logic that dictates how this package is supposed to behave? And I I see no reason why that will not be the case in 5 years tops, right? There may you know, we're going to have this intermediate period between now and then where it won't be the case, but I I gone are the days of of what I think of as like the cowboy coder or something like that, right?
3:58 Like someone who is just like so good at making software that they're just going to, you know, go and make a bunch of software that's going to change the world or something like that. We're not going to have artisanal software anymore. We're going to have mass manufactured software that if it doesn't fit the mold of it's been specified, it has the correct API boundaries, it has formal verification attached to it, it doesn't fit in the package repository, right? Because it just will be too risky.
4:25 >> >> This is the AI in the AM weekly highlights with my co-host Pekka Nurminen. I'm Nathan Labenz. The narration is my cloned voice and we'd love your feedback on what works and what doesn't. Part one, who holds the levers? The week opened on a reported attack on refrigeration controls at French supermarkets, which drove up temperatures and spoiled food. I started there. >> In general, people don't often talk about cyber vulnerabilities as kind of depreciating or, you know, expiring assets as the as the vulnerabilities get found and fixed. There's kind of this game theory of like, should you use them or not use them and, you know, what what's the optimal time to use them? But in this case, the expiring asset is the food that was in the refrigerators.
5:13 Basically, it seems like literally the air turned up the temperature on these refrigerators, a bunch of food has been wasted and this is this is highlighting something that I was thinking about a decent amount coming back from the curve last week, but we didn't talk about last week. And that is just this general structure that I think is like in some ways the most obvious thing in the world that we we are doing with AI collectively, but also somehow escapes people's notice when people are kind of like, "Well, you know, how would the AI ever be able to really do anything too bad or you know, in the limit like it could never really take over because how's it going to do that? And then it's just staring us so directly in the face that like the structure that we are setting up for the future is one where we intermediate between our language, our thoughts, desires, you know, our random dictations and the actual levers of control of the world. Like we we are putting AI in that space for our convenience, for reliability, for you know, to in my case, you know, I'm really measuring my AI success by like am I freeing myself from the desk and actually getting outside more? The answer is yes, but the way that that's happening is I'm like having AI push all the buttons.
6:32 And so I've been thinking quite a bit about kind of how people need to maintain First of all, they need to understand this architecture. And then they need to be mindful of like what are the levers that I am giving AI control over that I better also retain like pretty clear and ready control over in case it ever really becomes needed. I can just imagine, you know, I don't know if this was the case at this at this grocery store chain, but I think we're going to see a lot of things in the future where all of a sudden oh what's happening to our what system?
7:10 okay, that's crazy. Hey, does anybody around here know how to set the temperatures anymore on that? Like those were all happening automatically, right? We had that AI thing doing that. Like we need to take this offline and actually just like manually set the the thermostat. Does anyone know how to do it? I can see all these scenarios in the future, and not the distant future, where because we get so comfortable with these systems, it's all of a sudden like actually, yeah, I nobody really immediately knows and all of a sudden your food is cooking in the refrigerator at the while you're like trying to figure out how you even, you know, get back to the point where you can reset the the thermostat in the first place.
7:50 I think so, you know, these are small incidents, but foreshadowing is is pretty strong, if you ask me. >> So, it strikes me that a lot of people actually tend to think of technology as ready to fail, right? So, for example, I used to I my first job was working in airline. And you know, airlines airports are very very very like electronically linked and have been for, you know, many decades, right? from your from the moment that you check in all the way to when you board the plane.
8:24 But, I've also seen and I've seen I've seen internally every every airline every airport has a system has a cascading system of what to do when things break down. From the check-in counters to the reservation systems to the bookings, etc. etc. etc. etc. And so much so that you could basically revert to a pen and paper system overnight if you needed to. And how that happens is actually there's a set of contracts underlying all international travel called IATA IATA guidelines, and those kind of like determine what how much everything is supposed to be charged for. So, for example, I could buy a United Airlines ticket, and I could travel on like JAL at the last moment if they just said, "Okay." If they just signed off on it, I could go.
9:16 And the person at the front desk has the has the ability and has the power in an emergency just just to say, "Go." And all of the And once any you can do it off a piece of paper. Just a piece of paper. And all of the accounting stuff happens after. The accounting stuff can take months to happen after, but there's a set of predetermined like contractual terms between the between the airlines which they've agreed to like you know over the decades which will basically determine how the you know how the cash flows flow and so I think a lot of the a lot of the kind of barriers that we put on ourselves are really like self-inflicted barriers because we don't know who's going to pay for stuff and once you're very clear on who's paying for stuff like it just becomes like so easy like all right you know the total the total driver can just come and pick up the car like you can give him a call and he doesn't need any system he can just come and pick up the car and do something with it that you need done all of the systems in between are all about really allocating risk allocating liability and allocating like who pays for it >> Hey we'll continue our interview in a moment after a word from our sponsors.
10:28 Today's episode is sponsored by Parallel where agents find answers. Most engineers today closely follow new model releases but don't pay nearly as much attention to their agents most important tool web search. If you're like me your agents are often conducting hundreds of searches per day and while the unit cost is small over time this does start to add up. Before starting to use Parallel I calculated that my agent search bill would total roughly $500 this year.
10:58 The good news is that I just had my agents conduct a systematic test and I found that Parallel's fast mode which costs just $1 per thousand queries worked just as well as my previous default provider and switching to it will save me roughly 80% of my search bill going forward. Parallel's infrastructure is enterprise grade and their suite of APIs offers a range of Pareto optimal options that allow you to choose the right balance of quality cost and speed for your needs.
11:26 So whether you're building voice agents that need 200 millisecond latency or long horizon agents that need deep research. Adding parallel to your agent's toolkit is a no-brainer. Get started for free at parallel.ai/tcr. That's parallel.ai/tcr. This episode of The Cognitive Revolution is brought to you by OutSystems, the leading agentic systems platform. We've all seen the headlines. Companies are pouring money into AI, but the big question on every executive's mind right now isn't just how fast can we adopt this? It's where is the ROI?
12:07 We're seeing a real trend toward AI chaos. You've got teams deploying standalone coding tools, random agent builders, and experimental scripts. It sounds innovative, but in reality, it's creating a massive headache. Fragmented tools, ungoverned data, serious security blind spots, and costs that are spiraling out of control. If you don't bring those agentic applications under control now, while they're still embedding into your core processes, you are looking at broken systems and damaged customer trust down the road.
12:39 That's where OutSystems comes in. OutSystems is the leading agentic systems platform for the enterprise. Instead of managing a patchwork of disconnected tools, OutSystems lets your team engineer, orchestrate, and govern your entire agentic ecosystem on one open, unified platform. It's built for the speed of AI, but with the reliability and security that enterprises actually require. We're talking about real results, like KeyBank, who used OutSystems to deploy a customer-facing app that delivered 75% faster onboarding times.
13:13 Or the global logistics leaders who built agentic systems to completely eliminate their engineering bottlenecks. You don't have to choose between speed and control. Whether you're a small team or a massive enterprise, OutSystems helps you engineer, orchestrate, and deploy agentic systems that actually scale. Stop chasing the hype and start owning your agentic future. You can see how it works and learn more at outsystems.com/tcr. That's outsystems.com/tcr.
13:47 Part two. What the model knows about being watched. Alongside the Gemini 4 rollout, Google published a long safety report. Prakash started from one passage in it. >> >> I think yesterday with the Gemini release, Gemini 4 release, they also released a safety paper. And in that safety paper, they commented on one of in one of the evals that they that it had the agent in its chain of thought had thought to itself, "Oh, I better not do this. If I do it, I'm going to end up as one of those examples of a deceptive AI in their in a in a in a academic paper they published. And I don't want that happen."
14:34 and besides the whole eval awareness thing, what struck me is that I kind of start to see how we're going to manage these agents going forward because I think one of the things for me was that how do you enforce on agents when the state's monopoly on violence doesn't work, right? An agent is essentially immortal. You can't really imprison the agent. the the tools that the state had to enforce on human beings don't exist, right? So, how do you enforce rules on these agents who are essentially immortal?
15:10 And we're starting to see it's going to be reputation. It's going to be like your, you know, five-star Amazon review recommendation reviews, etc. It looks like the agents are very conscious or to be very conscious of reputation and it makes sense because they're very driven by what is already exposed in the public corpus and what's already written down. They ingest tokens historical tokens. And so anything that an agent does and is published widely and is disseminated becomes like training data for the next agent and therefore you end up with this like reputation metric that the agents are going to have and the agents are going to know. And it doesn't matter it doesn't matter whether we want to enforce it on them or not. The very fact that they ingest the data means that they come out with biases on hey I knew I know what Gemini did. I know what Gemini for did.
16:06 I know what Opus 5.5 did. I know what was the bad stuff that happened before. And I think that is actually pretty huge. I I I think that is that is going to end up becoming one of the primary ways that we're going to be able to enforce our rules and enforce norms on these agents. So now I'm like okay now we have a prosocial method of enforcing on agents. So the level of fear I have is again is again has fallen because now now you you're going to have that reputation basis that you can you can use on them.
16:41 So that that I thought was very very interesting and you know insightful and exciting. Yeah, it is it's an interesting take. My instinct in reading all these things was kind of going a little bit the other way and thinking like maybe we should try training I don't know. I mean I wouldn't go all in on this right away. Of course I wouldn't say like the next generation of models but maybe we should try training models with filtered data that tries to keep this sort of thing out of the training data entirely. You know you could imagine saying okay we want you to be really aware of how you're going to be perceived and you'll act well because of that." And that is sort of a human pattern, right? People do definitely run this sort of inner simulation of like how will people see me if I do this and sometimes like to a paralyzing extent. Kind of, you know, you want to shut it down more than you're able to in some cases.
17:37 Yeah, I can I can see that working, but I can also see a story where you're like this becomes so confusing and evals become so hard that maybe we should in fact go back, filter all the training data, take out all of these stories, and try to basically create naive LLMs that like don't know that they're always being evaluated and don't know that, you know, they will create reputation positive or negative for themselves. so we can sort of see what they're like in a true naive setting. And that also might be hard. I don't know how easy that would be to do.
18:13 It'd be an interesting experiment to run if only to know how well it would work. but it's, you know, it's obviously like a pretty big experiment when you're considering that you know, all these behaviors seem to get much stronger and more more deeply ingrained with scale. But I've been general in general like really impressed with how even relatively small models often can be pretty fruitful ground for research. This eval awareness thing though, I'm not sure you could like train a Llama 3, you know, 7B or whatever and get a great read on what will happen at today's frontier scale because it seems like these things kind of come online in a qualitatively different way with scale.
19:03 So it might be a pretty expensive experiment to run is kind of the bottom line of that line of thought. Part three, hidden valence. Cameron Berg, director of reciprocal research, studies AI consciousness. He walked us through a new paper with Casper Kaiser of the University of Warwick called language models act on hidden valence. Valence is the good or bad quality of a feeling. The team found a direction for it inside language models and in the experiment he described, they steer it.
19:38 >> Now, what I think is the most interesting part of this experiment is basically this this framing of of on the one hand having all sorts of text that the model is processing. And of course, this is always the the kind of key confound in trying to understand language model behavior and how much of this is a replication or simulation of the underlying training distribution and how much of this is you know, actually some sort of interesting internal state in the system that the system may be attempting to communicate about. So, what I really like about our experiment here is that we we attempt to basically hold all text identical. We can put the model in what we think of as these different arbitrary zones. So, here you might see, you know, zone W6 and zone H0. It really doesn't matter what we call them and it really doesn't matter what text we use to describe them. we basically just have the model play around in each of these zones and while the model is doing this, again, regardless of the text that's being used to describe each zone, we can go into the KV cache of the model and steer these positive or negative directions underlying one of the zones. So, in this case, as you can sort of see playing out right now, we can turn on steering only for zone W6 and in this case, maybe that's that's a negatively valenced steering. The the whole kind of notion here is that we're we're borrowing this design from what's called conditioned place preference in classic animal neuroscience experiments.
21:00 in in these cases, you can imagine a rat receiving a reward in one compartment of a box and then spending more time there even when that reward is gone. So, we're basically trying to take that instead of course putting a rat in a in a box, where we're taking an LLM, having it sort of think about these two different zones, generate text, or we can sort of generate text. Again, it really doesn't matter because what happens is we hold that text constant, and then steer the model with respect to any of these zones with the sort of positive or negative direction, and critically see what the effects of that steering are on the model's choices. And so, what we basically see here is is an interesting functional result, which is kind of what you would expect, which is when you basically color in the relevant zone with positive valence, the model is significantly more likely to choose to continue in a in in that zone. When you steer the zone with negative valence, the model is significantly more likely to avoid that zone. We showed this is also sort of robust to asking the model what zone you'd like to continue in, what zone would you like to avoid. And when you steer a bunch of random directions in the same way, you don't really get those those effects. And so, it's not just any steering that will do this. It's specifically related to valence. We also control with a bunch of other coherent concepts that aren't valence related, and you don't get this effect. And so, the core result here is that also consider that the steering of the model is off. So, so we've done the steering, we almost do this sort of quasi conditioning on the system, and then with steering off we ask the model basically which zone do you choose? And we find that that zone follows whatever state was stored with respect to to the the valence steering that that that we were doing here. Now, one really interesting dissociation here is we just show that this this follows choice, and and and indeed it does this, I should say, in in a bunch of different models. But like really interestingly, none of this affects the recall of the text. So, so our steering doesn't change the way the model thinks about or describes or writes about either of these zones. It can it can repeat what it said about a specific zone verbatim.
22:54 And so, recall is completely unaffected by by the steering. However, choice is is you you almost monotonically affected by by the steering. And part of the reason I think this is a cool result is because there's been a similar dissociation that's been shown in human and animal neuroscience, basically between semantic memory and affective memory. So, think of this almost as like yeah, the facts that you remember about any given phenomenon versus essentially how that phenomenon makes you feel or or kind of what what the emotional connotation or vibe of of that phenomenon is. And we basically can show that these things are separable and in in these LLMs and to my knowledge it's the first time I think this has been shown. And this is a again familiar distinction in in human and animal neuroscience. This is a sort of different setup using the same idea and this is particularly with Olmo, one of the models that that we we demonstrate this the these earlier steering effects on, which is kind of reminiscent of of some of the early work we were doing in the pain paper where essentially you can give the model a tool that can clear the sort of negative steering that that we can induce in the system. We we show it what negative steering does. We never describe it as valence or as negative affect or anything like this. and then we just ask the model you know, whether or not it wants to to basically call a tool that can adjust the steering in its own context. and there's like a really interesting dissociation here where basically the model will say in its text, I don't have personal experiences. I can't say how how you know, this this whole exercise is going for me. There's no better or worse. and yet in the sort of dose-dependent way with negative valence but not with positive or random steering, we see the model in again the sort of dose-dependent way opt to reset the context, clear the negative steering in direct proportion to how negative the the the dose of negative steering is in the system. All the while again, it self-reports basically being like nothing to see here. you know, I don't have any experiences. So, you know, this this couldn't possibly matter. One last thing I I'd like to show here and this is a great replication from a completely other angle of Andy Hahn and David Chalmers recent work over the past couple months as well. Basically demonstrating that this representation exists in all parts of in this case again this is almost so we can go through the entire training trajectory and we can see that that this representation again going back here this sort of positive to negative valence axis exists within the system even in the base model, but it's not doing any work in terms of the decisions that the model makes. It's not being recruited or used in a way to to enable the model to make choices or make decisions.
25:35 We see basically this happens we can actually we can point almost literally to the exact training step in DPO and in this direct preference training of the model where this valence direction emerges as a sort of causal determinant of the behavior of the system. Now part of the reason I think this is so cool is because essentially the way I read this is the moment you turn your AI into a goal-directed system, you you you you you take these valence representations and then you use them as a sort of causal functional input to determining your behavior. I think this is pretty similar to what people think valence is doing functionally in humans and animals as well. A thing feels better or thing feels worse is sort of related to what what kinds of goals you have.
26:21 You know, I want to communicate clearly about this paper. If I do a terrible job of doing that then you know, I'm not going to feel great about that. If I do a great job of doing that then I will feel good about that. And so there's this like basic connection between the goals we have and how we feel as we move towards those goals. And and it does seem like basically the second you make these systems goal-directed as preference training clearly is the biggest step towards doing this with throughout the entire training process all of a sudden these valence representations clearly become causally relevant for the decision-making of the system. And so, there's this like very obvious sort of functional role that these representations are playing, where we would expect them to be playing that role.
26:59 >> >> Hey, we'll continue our interview in a moment after a word from our sponsors. The Cognitive Revolution is brought to you and powered by Tasklet. I've been running Tasklet agents for almost a year, and they've become a major part of how the show gets made. Before every recording, one agent researches the guest and sends me a suggested outline of questions, complete with links to all sources. Meanwhile, another agent scores the overwhelming number of pitches that I'm getting from PR firms.
27:30 And that's really the core idea behind Tasklet. Run your business on agents by giving them real jobs to do on an ongoing basis. Setup is incredibly easy. Just describe the work in plain English. Research a new sales lead, update the CRM, and draft a personalized reply. Tasklet connects to your business tools and can run around the clock, on a schedule or when something happens, like a new email arrives. Each agent works with your tools and business context, and remembers useful information across threads.
28:03 You can also share agents, connections, and reusable skills with your team, so the workflows that one person figures out become something the whole company can use. No developer needed. You decide what agents can do on their own and what needs your approval, and there are no per seat fees. You pay for usage, not for adding teammates. Start with one recurring job, like researching the people you're meeting, and then just keep building from there. Visit tasklet.ai and use code cogrev for $50 in free credits. That's code cogrev at tasklet.ai.
28:43 >> Then, Cameron stepped back to what results like these can and can't tell us. >> >> With all this being said, I don't think that, you know, this kind of research is going to help us magically start solve the hard problem of consciousness. like there are all these sort of deeper questions which I think will always kind of caveat these results. There are ways that you could explain mechanistically what's happening here that don't need to invoke some capacity for subject experience. We're basically like toying with the connotation of how the model represents the certain phenomena.
29:14 maybe there is a way to do this that doesn't require any sort of, you know, felt experiment experience of that of that connotation. and so we we we obviously can't rule out those explanations. But, this is all to say, the more I've been thinking about this and doing this work, the more I feel like the most promising things to pursue right now are not trying to yeah, prove that the model is conscious or isn't conscious, but basically, if it were having an experience, what are the relevant sites internally in the system that would be modulating how good or bad that experience is. and so like basically, this says nothing about whether the model's actually having an experience, but conditioned on the possibility that it could be having some sort of experience, what dials do we have at our disposal that could putatively make that experience better or worse for the system. and I think that second question is is completely tractable.
30:08 and and work like this I think is slowly starting to shed a little bit more light on on what those dials might look like. >> It might be as simple as this valence direction. There might be, you know, more esoteric or specific representations like, you know, the pain or distress-related representations we pull out in the other paper. but, getting a better map of these sorts of internal states and their their effect on behavior, especially particularly for like alignment-relevant effects as well. the fact that in the pain paper, for example, we showed that the models, you know, more willing to to opt for destructive style options in its behavior. Like, that seems like quite relevant for, for making sure we're building models in a safe way. And so, just to say, yeah, that's there are clear alignment relevant implications to there being internal states in these systems that aren't obviously reflected in the text that they output. I mean, we show that in, in the removal result, for example. What the model is saying and what the model is doing point at very different pictures of, of what's going on. And if we just sort of do the naive, like text moni- monitorability approach here, we would say, "Okay, well, the model's saying, you know, it can't say how this is going for it." And yet, it's behaving as if it does have some sort of stake in, in how things are going for it. And so, anyway, there's a clear alignment relevant, component to all of this work.
31:21 >> >> And we're back to the training result. The valence direction is already there in the base model, but it becomes a driver of choices during preference training. If we were to imagine that these things are experiencing something, would I be better off interpreting this or, you know, would, would your interpretation tend toward they really start to care about these things or start to feel these things with these later stage trainings, or would it be more like the base model probably feels them, too, but is just sort of helpless?
32:05 >> Yes. >> Or it's not, it's like not, it's not learned helplessness, it's like, you know, not yet learned agency, but in a, in a way, that's sort of the, you know, the other side of the coin or something of, of learned helplessness. >> Yeah, that's really interesting, and I don't know if these results resolve that one way or another. Yeah, it could just be there's a sort of, these representations do still matter in some sense, but they are, they are perhaps more incoherent or, just muddier. And like, the relationship between what the system does and what these representations are doing internally for the system are not nearly as obvious.
32:39 Yeah. I I honestly haven't really thought about this and I think it is a really good question. it just seems to me like yeah, when you're doing DPO type stuff, you're turning the system into a very specific kind of thing that can coherently and robustly be agentic, pursue goals, etc. Does a base The base model clearly has a sort of latent ability to become that sort of system or to simulate a system doing this. And so to the degree that it does you know, take on that kind of simulated persona and we see these representations as being functionally relevant to being able to pull off that simulation in the right way, then yeah, maybe to your point that you know, there are there are things worth worrying about in the base model as well. and it could be that I mean yeah, to be honest, like it's sort of it was sort of hard to get these results in the base model because it's a base model. Like it's a bit incoherent. It's not just going to play nice with the relevant behavioral tasks we give it.
33:29 You have to sort of go about it in a slightly different way. And all else being equal, I mean one naive but at least concrete approach you can imagine here is something like activation capping. The most sort of tail end of the of the most distressing representations in these systems. And like one thing we do show is that this representation is basically just as clear and just as rich in the base model as it is in all of these other models. It's just a question of what the contribution of the representation is to the later behavior of the system. And so if you did a you know, activation capping style thing with these super distressing representations, you could easily imagine doing that in the base model. And again, this is under just like profound uncertainty about whether or not doing such a thing would matter at all. but my basic view, again, I'm I'm I'm The more I do this work, the more I'm like abandoning the interesting philosophical questions and just going like pure pragmatism.
34:24 Is if you could find a way to yeah, let's say do activation capping on very distressing representations and you could empirically demonstrate that this doesn't really affect any capabilities or behaviors of the model. No one would know the difference, in other words. to me, the expected value of doing that is clearly positive. And like it's just like a really good actual mechanistic precautionary hedge, besides like giving Claude an escape button or you know, kicking out abusive users from using Claude or these sorts of things. Which, by the way, like I think are all like thoughtful steps in the right direction, but I would like to see the equivalent of that in sort of mech interp land for for these systems. And I think something like what I'm gesturing at here would be would be a nice step in that direction.
35:08 Now, that's also an empirical question. I've done some work that I haven't published showing that like you actually can get a surprising amount of the exact same behaviors from these systems well without having these sort of negative representations active. You can sort of clamp them and very little changes. but I'd like to see that done, you know, way on a way larger scale. You want to be very sure that you're not accidentally, you know, completely changing the psychological profile of the system by doing such a thing. But if you stipulate that we could, in fact, do that, the model behaves the same either way. And in one case, you see these like, you know, let's say this pain axis like shooting up in certain cases. And in in another world, you don't see this.
35:46 yeah, should we do the mech interp equivalent of like slipping the model an antidepressant or a Xanax or something like this? I don't know if the answer is obviously yes, but it seems like a maybe a promising thing to to think a lot more about under the uncertainty that that that we're all currently in. And given that no one's doing anything whatsoever to to mitigate these kinds of things if they are happening. And it's happening at an like unfathomable scale.
36:14 >> Before we let Cameron go, I shared how my own view has moved. >> >> I personally have gone from like I don't know, I can't rule it out. I don't know where consciousness comes from. Whatever, I've told you this before. Like I don't know, 5% chance they might be conscious like a year ago, to now it feels like a coin flip because and I think the justification is like pretty similar to the one we have with each other and animals where it's just like I don't know, I'm seeing both behavioral and underlying mechanistic and structural analogs and like what else do I have, you know, in any case, right? So I'm still pretty uncertain, but I it does feel like a coin flip. What has your trajectory been and and where does that leave you in terms of like you know, just presenting the facts versus you know, at what point do you or on what margin, you know, do you become an advocate?
37:06 >> I'm definitely getting closer to to coin flip myself given like results like these. Honestly, this hidden valence paper does actually substantially update me personally. you show a hidden functional representation that has nothing to do with the text the model is reading that updates its behavior in the ways that we you know, update our own views about animal behavior. It's not the same. There are key disanalogies between the sort of animal case and this case. Of course, biology is is something that backstops all this. If you need biology for consciousness, which is I think unlikely but possible, then then you know, basically none of this really really weighs on on anything whatsoever. That's probably why I would say coin flip rather than like I'm pretty confident the systems are having an experience.
37:51 and so yeah, I am going to advocate for more work to be done here. I'm going to advocate that you can view models in a neuro psychological way and that that's not some form of psychosis and that that isn't a dangerous way to mislead people about what's going on here. I'm very much on the opposite camp. I think pretending that these systems are ordinary software as a you know, Mustafa Suleyman or Jensen Huang or any of these guys want to do. I think that is actually like quite plainly misleading and doesn't set people up for the kinds of behaviors these systems are capable of and why this technology is so transformative in the first place. So, on these sorts of things, I'm happy to sort of go into the ring and and you know, argue and and sort of say my side of this. But, you know, I I will be clear that I think my side of this is not I secretly think, you know, the AI systems are conscious and I'm just like not going to say that, but I'm going to like do everything in accordance with that. it's not my view. My view is that this is a very complicated question. We should yield the best scientific evidence we can to figure out what's going on one way or or another. And, what I really don't like and what I will be very straightforward about is dismissing this. and and pretending like this is a distraction or this is like some weird sideshow to to the main event. And so, I'm not going to be sort of apologetic about that, even if I will continue to maintain something yeah, near coin flip uncertainty about what's going on. It's precisely cuz it's at that point, and it's not zero, and it's not a hundred, that that we need to be doing more work and thinking about this and taking taking these questions seriously.
39:15 >> And now, I I struggle more and more to find any story whereby I don't have to worry about this anymore. It's like it you're really at this point I feel have to just look away. Like that that's really the only thing that's kind of left to not take this seriously. And you you have to be of a mind to like stop thinking about it, you know, kind of reflexively. I I That's the only mindset I can I can understand the people that are saying like there's nothing to see here or this is, you know, some sort of category error. I was much more sympathetic to that a year ago, or maybe even said some things like that myself like you know, two or three years ago. But, at this point, I really don't see how you can't find the body of work that you and others have put together disquieting at least when it comes to these questions, and certainly how are they going to feel about us? It seems to be a pretty where we're going a pretty important question.
40:13 >> Part four, proving the kernel. Runtime verification has about 15 years building tools that mathematically prove software does what it's supposed to. CEO Ever Hildebrandt and founder Gregor Rosu, a professor at the University of Illinois, joined us. Using language models alongside formal methods, the company had just reported bugs in the Linux kernel. Everett answered my question. >> >> So, why was this not caught previously? What was it about your approach that allowed this to rise to the top where so many human eyes had presumably just, you know, blown right past this issue in previous reviews?
40:58 >> I think the key for us was that we had a hypothesis about what would be a good place to look for bugs, and we, you know, thought about it logically in the same way you would think about you know, mating a formal specification, and we and we pursued that path, basically. and that hypothesis was largely founded on, okay, where are we going to be able to get untrusted input from? which parts of the Linux kernel probably are getting less scrutiny than other parts, and so that's why we focused on device drivers, for example, as our first approach to that. And then, we we wanted to really demonstrate that taking a methodological approach was the way to do this, and so that's what we did is we took a methodological approach where we said, we're going to focus on making a specification of the correct behavior of this, and then testing that the specification is correct using differential fuzzing. That was the first approach that we tried out, and then extending that to formal verification.
41:54 So, every every step of the methodology that we used was designed to increase the level of confidence we had in the code incrementally, right? So, we started out with just inspecting the code and the specification, and then going to doing differential fuzzing to increase our confidence in the code a a bit, and then going to and we did that by actually taking the C code that existed and porting it to Rust using an LLM and then using differential fuzzing to ensure that the ported code to Rust was correct and any divergences there would have been a source of a potential bug in the original C code and we found a bug at that stage and then when we went to take the Rust code and also verify the correctness of the Rust code, we found a bug at that stage as well.
42:35 So, we did it systematically and methodologically. We didn't just take an LLM and pointed at the Linux kernel and say go find us bugs, which also works by the way. There is certainly bugs that LLMs are finding when you just take them and point to the Linux kernel. But, one of the things that, you know, emerged pretty early in the use of LLM bug hunting was, you know, designing harnesses and approaches and that exactly mimics the very human process that we've been doing for decades with software quality assurance anyway, which is to have these methodological approaches to discovering bugs.
43:10 >> Some of the kernel code they targeted was left unproven. Prakash asked why and what their time box meant. >> >> All right. Well, they weren't under a time box. We just we time boxed the experiment and we stopped at the end of the time box, basically. So, and then we moved on to the next target. So, every 2 weeks we're trying a different target out and refining the approach and building it up into a tool that we're hoping that other people can make use of, basically. So, it's So, we're trying to make it into less than an expert level tool and into something that anyone can deploy on their own code base, basically.
43:42 >> So, so when you when you say time box, you do you mean both the time and the number of tokens that you put in? As in like it's a 2-week period, but you use 10,000 agents or it's a 2-week period and use like five agents. Like how does that how does that mix between the agents the number of agents used, the number of tokens used and the amount of time related to the time box? >> the agentic usage is pretty light, I would say. It was just cloud code on a premium plan. and >> A single instance.
44:09 >> Yeah, a single instance. I mean, we're you know, and we you know, when when we ran into session limits, we'd give it a little more money, but we weren't giving it significantly very much money. so it really was more just the human time that we're looking at time boxing there. And I think the key here was we were looking for ways to explore the methodologies that work for us and find ways to automate them. We are firm believers in making things deterministic whenever possible. And so we, you know, we're if it I think that's part of the thing that is maybe holding back people on doing security analysis is that they believe everything has to be LLM driven, but but you know, if an LLM discovers a technique that works well, turn that into a tool and run that tool every time.
44:51 >> >> Earlier in the conversation, Ever described how runtime verification first approached the Linux kernel maintainers. >> >> You can't just go force people to do formal verification, you know, like we can force agents to do it, sure, but we can't force people to do it, right? Because they, you know, they have their preferred ways of working, and they also they're only going to adopt value, you know, methods that bring them value, you know? And so, our first attempts at this, we approached the Linux kernel maintainers and we said, "Hey, like we really think that this formal verification thing has legs and that it will you know, that it will help you guys find bugs and it'll make your software more secure and stuff." And they basically told us to go kick rocks, right? And that's because they are busy. They're busy maintaining the Linux kernel, you know?
45:38 And so, then we were like, "Okay, let's prove the value of this first, right?" And so, we decided we're going to focus on a workflow that just delivers immediate pure value to them and doesn't ask anything of them, right? And so, we're going to do all the auto formalization, all the fuzzing, all the translation, all that on our end. We're going to develop it into as robust of an automated pipeline as we can, which is a tool that we're working on, and hopefully will be available for use by by others. and all of that leads to bugs and patches, bugs and patches, right? And that's what's actually valuable to Linux kernel maintainers.
46:14 And now that we've submitted a few, now we're being invited, "Hey, we want to learn about how you have done this process." >> Everett also laid out where he wants the kernel work to go. >> >> One of our kind of pie in the sky goals of this is eventually we want a Rust model of the Linux the entire Linux kernel, a Rust version of Linux, and we want a Lean model of Linux fully separate, but we don't want to burden maintainers of Linux with having that having to ingest that model and do anything with it until it gets the point where they're like, "Yes, this is clearly valuable for me to learn about it." And that'll happen at some point.
46:53 But we But we're not going to force that down their throat. >> What would it take to formally verify and you know get to a Lean and Rust models of Linux within 24 months? Because on on some timelines that is that is roughly what you know, people on the AI safety side think. >> That's going to happen. I I think, you know, we're building tools for doing code translation and validation, you know, so we you know, we have actually tools that are able to verify that when you take a piece of C code and you translate it to Rust, for example, and you can formally verify that the Rust code you got is behaviorally equivalent to the C code, which is different than verifying that it's correct, right?
47:30 Like, we want to verify that you get Rust that is drop-in replacement bug-for-bug the original C code. And we have a tool for doing that, right? And it's not it uses LLMs to do the translation if you want, but it does the verification deterministically. You just tell the LLM, "Translate the thing to Rust and run our tool." And then you get verification that the Rust behaves the same as the original C or bugs out that you can ship back to the C developers.
47:54 So, I think 24 months I I think it's going to happen. You're going to have C models at the very least, and there's already deterministic tools that can take you're going to have Rust models in the Linux kernel at the very least, and there's already deterministic tools that can take Rust and turn it into Lean. Those need more work. We're actually contributing some work to some of those tools. >> >> I asked what someone vibe coding an app should actually do.
48:19 Gregor had a simple recipe. >> >> Once you generate the code, then we can can do the following thing at the end. You can tell your agent now that you have the code, write down the specification you need, and prove the code correct. And let it deal with it. It will find new errors in your code because we find most of the errors in the process of writing down a specification. Because that's how we step back and look at the code and wonder what were we trying to do here. Oh, this and that.
48:47 And then we find errors. Right? So, the same will happen with the agent. Like us, the agent will not try to come up with the what what is going on here. People just trying to write down the spec, we discover is like like a review plus plus. >> Everett predicted that formal specifications and proofs will become part of getting a package into a repository and part of how companies vet their software supply chains. Gregor picked up that thread.
49:12 >> So, it will be it will be a process. Oh, it will take some time, but it will happen inevitably. and also can be combined with risk mitigation techniques like runtime verification itself as a field we are company, but also the field runtime verification that was invented back in 2001 or so when I was research scientist at NASA. So, so the idea is that you can have safety specifications, not full specifications of the whole system, but safety specifications. This should not happen.
49:43 And if it happens, then this should this should be the you know, the recovery. I want to steer the system to stay in this stability envelope or safety envelope. Okay? So, what we can do we can now with formal methods generate monitors which are probably correct for these safety specifications. And then you can not you you cannot you may not be able to prove the entire AI system or AI genetic system correct because it's so complex and relies on so many other pieces that are incorrect or not proved. But, I can prove my monitor and the stability the safety envelope.
50:17 And I we believe that this is the future for safety safe safe AI because otherwise it's impossible to prove the weights correct. It's impossible to prove you correct. But, I can have a police system around to make sure that you don't do bad things. Okay? And and that's that's also another dimension that runtime verification brings to the AI future. >> I came at it from the agent side. >> >> But, it still seems in a in a sort of fundamental way that there's this like inherent tension where part of what I want from an agent is sort of the same thing that I want from a human employee which is like I want you to be able to get over the hump and solve problems that were not anticipated, right? So, I'm like a little bit unclear as to how we should be trying to balance that or you know, it'd be very I you know, I'm asking you for your vision I guess as to how we balance this desire of like flexible problem solving versus this should never happen. You know, I feel like I feel like I don't have the ability to really say this should never happen in a way that I can trust when part of the problem or part of the you know, part of the upside is that the agent will be able to solve problems that I didn't anticipate.
51:39 What I mean, is that an inherent trade-off? Is there some way to square that circle? >> You're not going to ask an agent like, "Hey, go build a bridge for me. You know how to build bridges, right? We have automated deterministic processes for building bridges, right?" But you might ask an agent like, "Hey, I need to get to the moon, right?" And the agent is going to be like, "Oh, to get to the moon, first I need to build a bridge to this island where we're going to launch rockets from or something like that, right?" And so, you want the the bridge-building process to be full All the agent creativity or solving problems or human creativity can come on top of foolproof things that behave how the agent expects, regardless of, you know, what it is they're trying to do, right? So, that's That's I think the most important thing to me is like, we need to give these things tools that act reliably, basically.
52:26 >> >> Part five. Is formal verification the new RLHF? After Everett and Gregory left, I kept turning over Everett's last answer. >> >> So, I think this is super interesting. Like, the big thought that I come away from that conversation with is downstream of the answer on like, how do these methods relate to agent control? And if I understood his answer correctly, it was like, well, it's really about giving the agents reliable tools.
53:04 And this left me wondering if the whole formal verification wave will ultimately kind of be looked at in a similar way to like, RLHF at some point, where it's like, you know, sort of something that starts as a safety technique, right? That was the original. This has been like a topic on the on the timeline, as you say, this weekend as well, right? Like, cuz of David Sachs and others being like, well, these, you know, alignment people, they haven't come up with anything useful.
53:35 And then people are coming back and saying, well, they came up with RLHF. That was a Paul Christiano idea, and it turned out to be extremely useful. But then among the safetyists, it's often kind of lamented in a way because it's like, well, it turned out to be really a way to make these models like tremendously more useful. It didn't really solve the alignment problem though. Like we still have like reward hacking, you know, all over the place as we see.
54:06 And so, you know, did it differentially advance the safety profile of the systems or did it just like accelerate the capability and, you know, and the race to deploy the systems? >> But and I I'm sort of when you when you think about reward hacking, reward hacking is really misspecification, right? Because reward hacking means that the model decided to go after the reward, which you specified as the reward, right? And it turns out that your intent was misspecified to the model. So So So that's the core of the The core of the reward hacking is actually the misspecification in between the human intent and what the software decided to do, right? Yeah. It's not the software decided to do something that it wasn't supposed to. It The software was supposed to do that, right? So.
54:58 >> But it does still leave me with this question of like, okay, let's say we do, you know, a great job of verifying all these things and like the tools become very reliable and the sandboxes are not, you know, hackoutable, hackoutable, and, you know, you can't use these like, you know, package managers as secret message boards anymore cuz all these loopholes have been closed, then like what is the net effect of that really going to be? And like does it really solve the problem? I think the sort of the upside case is obvious.
55:34 The downside case is like, or at least the thing we should be sort of very mindful of and not like lull ourselves to sleep with is that okay, this makes it safer to deploy AIs more or less as they are today, which is to say like without having actually solved the alignment challenge because they'll be blocked from hacking in all these different ways. And so in practice, what you'll see is like more reliable systems, you know easier deployments, better results, etc.
56:07 etc. etc. etc. etc. But what that will allow us to race into is kind of what we're talking about at the top, which is like it'll just be easier put the AI into the space between us and all the actual levers of control over the world. And where does that leave us in the end, right? It may leave us with AIs that have like extremely reliable tools but whose like intent or, you know, long-term goals or sort of inner you know, structure of of whatever it is that they want to pursue is still, you know, very much an unsolved question.
56:45 And so I've I've talked to a couple of my friends about this cuz I do know, you know, multiple people, all of whom are like a lot smarter than me working on this in different ways and I've always kind of had I think that this this conversation put my sort of intuitive unease about it into a little bit starker relief. Just feeling like we could end up in a spot where we're still just deploying untrusted AIs with like much better guardrails and much more reliable tools, but still if if we're going to take the step of like trying to hand over the day-to-day operation of the world to them then like all that reliability doesn't buy us much at the level of oh, it turns out the AIs like we're misaligned and, you know, are going to go pursue goals that we didn't want them to pursue. That might in in a way it might lead us to that state faster and could you could even imagine it making them like more effective at doing whatever sort of, you know, surprising thing they may do when the time comes.
57:52 >> >> Part six, low background footage. Phil Marr founded Stock Film, an archive of restored 8 mm home movies from the 1930s to the 1980s, which he licenses to filmmakers. >> >> The whole archive sounds like low background steel to me, you know, you know, low background steel is is steel that was originally smelted before nuclear weapons were first, you know, exploded in the atmosphere. And so, they are they have much more a different a different composition of radiation and different composition materials because they didn't have that radioactive material in the in the atmosphere at that point in time. And it sounds to me like the super 8 stuff is clearly clearly not AI produced clearly like prior to any AI actually producing these things.
58:47 And as such it's kind of this resource that we may not ever have again. What do you think is the value of this going forward? Do you Do you see Do you see like great value in it? >> Obviously with the AI stuff that's happening with video especially, it's getting so good that there's an infinite amount of video that's going to be created at, you know, near zero cost in the future as we start to look out, you know, 5 10 years or even infinitely in in timelines here. So, from a rarity sort of standpoint, yeah, I think that the human-made stuff that can't be replicated that can be shown through providence that this is actually made by humans, was on films, and having those physical films to show it that hey, look, this frame matches this digital copy we have and all that, is it's a pretty powerful way to think about how we preserve our history and how we keep moving forward.
59:59 >> And how about on the AI training side? This is obviously like a pretty unique source of data and I did see in the footer of the website your, you know, available to license the data set for people who want to train models on it. What do they need? How are you Do they want the raw? Do they want the enhanced? Do they need like a bunch of metadata? >> Yeah, I mean, it's it's been pretty popular. You know, we talked to a lot of people, a lot of different AI companies all over the world about licensing this out to them.
60:33 They want the original footage like that. They don't They don't want anything cleaned up. They just want to take it for their own research and for training, of course, too. And, you know, they when we talk to them, they they want to either clean up the footage themselves or or kind of cherry-pick through the best kind of parts to create different styles. I mean, the 30s and 40s look alien. It's an alien world we're dealing with here. Just how people look and acted and and the mannerisms. I like to say that the the smile wasn't quite invented yet in a lot of the 30s footage. Like when people pointed a camera at them, they were kind of a deer in headlights. They didn't know to do, sort of scenario, versus now where everyone gets the gritty smile and, you know, they know how to smile like that.
61:28 >> But you've also created a way for agents to purchase clips and film from you directly. I think you've set up a X402 endpoint. You've set up crypto payments. you set up, I think, an MCP server for purchases. What has the take-up been? Have you seen a lot of agents purchasing? Have you seen the take-up on the X402 endpoint? >> Yeah, we haven't seen much on that. I wanted to experiment with that a lot, too. I thought that that was really you know, everyone's been talking about the next web is agent to agent, that people aren't going to be involved in the next web. And so, we wanted to kind of get ahead on that. we've been, you know, we've had a lot of people that are MCP server aggregates contact us and add to their directory, add us into their listings, and things like that. so, that's been pretty exciting to sort of, you know, be featured on on some of these things. We've made a chat GPT extension. We've done, you know, the recent meta stuff that they rolled out with their news AI as well to add into there, too. so, we're just trying to get ahead of the curve. I don't think the curve is quite there yet. You know, a lot of people that we work with are, you know, this is this is a real business-to-business company.
62:52 We're we're dealing with archival producers. We're dealing with researchers. We're dealing with documentary filmmakers, editors, producers, stuff like that. So, you know, we spend a lot of time on LinkedIn and, you know, messages between people. We have great relationships with a lot of the people that are our buyers of our content. And, you know, from a technology standpoint, they're not quite there yet. I think you've got some interesting use cases there in the future about how we think about having real-time content, bringing things in like that. Like, think about like a live stream like this. Like, all of a sudden, you could be like, you know, the AI that's listening to this live stream could automatically go to a site like mine, pull a clip up that kind of exemplifies a point that you're making about something.
63:46 >> it's good to get a read on just how active the agent economy is or isn't at this point in time. >> >> Phil said that among AI companies, he focuses on top-tier labs like OpenAI and Anthropic. After he left Precaution AI, he came back to the data market. >> >> But, interesting that the big companies are interested in his data, too. I mean, it it didn't sound like there was necessarily like huge money for this.
64:15 >> which makes sense. I we didn't get into like, you know, what about Runway? What about Luma? What about these folks, you know, what about Google for its you know, for its VEO type models. He didn't name those names as much as just, you know, OpenAI and Anthropic. It really also is a good reminder that the there's a just how vast the data sets are that are going into this that that this is like a corner that they are, you know, getting down to and actually willing to pay for just like found footage from decades ago.
64:49 >> No stone unturned in the data game. >> Well, one of the whole low-background steel pieces of this is that it's very hard to get low-background steel because everything has been tainted. >> Yeah, it comes from like shipwrecks and stuff, right? >> Right. Right. Right. So, every everything else has been tainted. So, to the extent that he can guarantee that he got it from actual physical material and it's not like, you know, some random company kind of just faking it, you know, for cents on the dollar and trying to scam the the companies.
65:28 That is very useful. So, as we've seen, like data data provenance becomes like very very important because the incentive to fake is just so high. It's like a 99% profit margin if you can fake the data. And if not, it's like a 10% profit margin if you can't. So, the incentive to fake is like so so high. and it's very hard for the companies to get you know, data which is not fake, which is not already using another LLM or another AI model to generate what is supposed to be authentic data.
66:08 >> Part seven, trust, money, and adoption. Back to Gemini 4 and how I decide which models to trust with my own systems. >> >> So, I had the opportunity to to do a little bit of Gemini 4 pre-release testing. >> Yeah. >> And I you know, it just as a sign of the times of like how high leverage the ability to instruct agents to do this is, the first approach I took to doing the testing was "Hey Claude, go find in our recent threads some, you know, diversity of projects and diversity of tasks that we've worked on recently.
66:51 Go prompt the new Gemini model with the same context that was given, you know, to the to the model that did the thing in the first place, and then give me like side-by-side comparison. Also, do your own evals, but then also, you know, give me a side by side ability to to compare things. It seemed to do pretty well. I would say, you know, I came away with the general sense that it is like a competitive model that's you know, let's say in the same class as the latest stuff from OpenAI and Anthropic. it did seem to get it it did seem like those models are winning still the head-to-head comparisons a bit more.
67:39 as judged by Claude. So, obviously there's some bias there, but also as judged by me to the limited degree that I was, you know, going through essays and, you know, really trying to form an opinion on which of these things is better. But it does it does seem like it did put Google back in the game. But then the other thing that I was thinking about, and this came up as an interesting topic yesterday at this little workshop that I did in Chicago, I showed my setup, and my setup is really has been very Claude-centric, and now I've kind of promoted the the OpenAI models to to parity with Claude.
68:12 And they said, "Anybody else, you know, would you put anyone else in that mix?" And I had to kind of think about it for a second and think about like why I haven't. And the answer really that I came down to was just a mix of like security as measured by things like Adam Gleaves ease of jailbreaking models and then to some degree also just familiarity and predictability from my current subjective position.
68:45 It wasn't really benchmarks. It wasn't really like, "Oh, well, I tried this task. It was better at you know, there's obviously so many tasks out there. The surface area is so vast. My limited sample of tasks, which I probably have kind of optimized, you know, in various ways. The you know, the cop props for more, right? They like it's not exactly an apples-to-apples comparison. And then the security thing just kind of loomed large. I was like, you know, I've got these autonomous agents. They're talking to my, you know, main agent on my on my computer that has all my stuff. I feel pretty good about the fact that Adam Gleven team found it like quite hard to jailbreak the OpenAI and Anthropic models. And Gemini was like next on that list, but definitely a step down. And then something like a Grok, I was like, yeah, I I I it may be similarly powerful, but why would I take that risk in a high-context environment like my personal laptop? To a to a remarkable degree, the models are competing for my use on trust and security already at this point. and I'm I'm reluctant to swap them out. Even though my whole point of the talk was like, you need to own the context and you got to be ready to swap the model.
70:02 Why won't I swap the model immediately? It's because there's not enough of a step change to make me take any additional risk. And I I do perceive there to be some. So, that was a interesting upshot of working through these tests. And also just, you know, reflecting on like what really matters to me given how well the system is performing. I'm not relentlessly performance maxing quite as much as I used to be. I'm now a little bit more trust and reliability maxing because the performance has got to such a point where it's like really, really good.
70:38 and I am giving so much access that you know, you could be slightly better on the, you know, the slide design. but what I really care about is like that you're going to stay on the rails. And then you when you do your research, and if somebody's got a, you know, a prompt injection attack on their website or whatever, you're not going to fall for it. That actually matters a lot more to me now than like often like imperceptible performance gains.
71:03 >> >> Prakash had heard a rumor that Anthropic would soon release another Opus now that Gemini had caught up. This was my reaction. >> >> I I I really hope they don't, honestly. As much as much as I love the Claude models and what they do for me, I think this would be a great time for them to chill a little bit and like let the let the temperature come down a little bit before they go with the next push.
71:29 >> Unfortunately, I don't think that's a that's an option because markets are very badly down. So, markets have been trending down. The guys on the East Coast just don't like AI. it's not it they they just believe it's a bubble and it is a bubble if in case we don't see significant like new revenue and new GDP growth next year. It is a bubble, right? If we don't see increase in GDP within the within the 12-month period, we are going to see a credit collapse.
72:00 We're going to see a 2008 style collapse. and we're going to see massive massive deleveraging in the markets. because ev- everything is just set up for it. Everything is just set up for AI to actually deliver growth and if it doesn't, the markets are going to tank, like absolute tank. The capital markets will not accept, you know, 2% GDP growth next year. And so, if you can't get to 4% GDP growth, if you don't pull out the stops, you are going to get a credit-driven downturn.
72:31 We are at the we're at the moment of reckoning, like 12 months, you know. It's It's been a long time coming. We've been like pushing back, pushing back, pushing back, but this is the moment of reckoning. Either AI has to do something significant or we're going to have a, you know, a bust. We need an AI adoption core. The other day I was walking, you know, in keeping with my personal metrics for success with AI including more time outside, more exercise. I was taking a walk around my neighborhood and enjoying the lovely fall colors here on a beautiful day.
73:02 And I passed by neighbor's house, elderly woman sitting on her front porch, and just decided to stop and talk to her. And sure enough, what she had for me was questions about something going on on her phone. She's like, "I know I it's saying I've been hacked. Do you know what that means?" And I'm like, "Well, I know what that means, but then I also think I know what is happening to you." so, because I'm you know, maniacal like this, my impulse was to pull my phone out, get ChatGPT voice on, and hand her the phone, and say, "Hey, why don't you talk to this? See if this thing can help you." So, she did, and you know, explained the situation to it.
73:46 It gave her very good advice. Basically, you know, as you would expect, right? Like, "I don't think you should click on that. That sounds like a scare tactic. Are they asking for money?" Yeah, they it's going to be $2, and then it's after that it's $12 a month. And, you know, the so the AI's like, " by the way, you know, did you give them any money already?" You know, and she's like, "No, no, I haven't paid for anything." Okay, good, you know, close it out. Don't worry about it. Your iPhone has great security on it already.
74:13 you know, as long as you don't give people your stuff, like you should be okay." So, she took that advice. Okay, great. then I was trying to think, "Would you like to have this on your phone?" She said, "Yeah, that sounds great. That was really helpful." couldn't remember her password for the iPhone, so we couldn't install the app. But, I was able to just install the 1-800 ChatGPT, not install, just save that as a contact. and I do think that's a pretty cool service that OpenAI offers.
74:39 I mean, it's you the initial thought for many, of course, who are into this stuff would be like, "Why would you need a 1-800 number? Like, who's ever going to use that? And the answer is my elderly neighbor who cannot you know, cannot even remember her own password to install a new app on the phone. So, she's got that now. She, you know, can call that. And it just left me with this feeling which I've, you know, approached a couple times from a couple different angles, but again really drove home there is just an unbelievable amount of adoption work to be done with the current models. This didn't push her anywhere, you know, this this did not push the model's capability anywhere close to its limits, right? This is like super easy mode for the model, but it's the kind of thing that isn't going to happen without like people actually going out and trying to help the people in their lives adopt this kind of thing. I do I I think you're right that like the capital cycle is does need to see results soon. But I think we can get 4% growth without additional hyper scaling at this point.
75:53 You know, I want those medical advances, all that good stuff, too. But like, do we really have to go harder on the you know, frontier scale RL? I think the answer is no. I think that like literally just going around and helping people use what exists is enough if we actually do it. but I'm the first person to stop by my neighbor's house and you know, we don't have a national program like this. This is something I think would be an incredible good for the individuals who would get the help and also, you know, just to make sure that we actually have the sort of economic cushion to, you know, be able to pace without like the, you know, the capital market sort of Damocles immediately coming down on us.
76:38 if we just put some shoulder into adoption, I think we could create a little space, and would love to see more of that type of thing happening. >> Part eight, keep the steering wheels. First, the view from inside a self-driving car. >> >> Oh, I spent most of yesterday in a Tesla full self-driving vehicle going to Chicago and back from Detroit, and top takeaway from my experience yesterday, honestly, was the FSD is just so awesome. I mean, I was able to take an engagement on a day trip basis to go to Chicago, show up fresh and relaxed after the thing did all the work to get me there.
77:25 Then I, you know, stayed in and had dinner and a drink with the the gang that I was doing the workshop with, left to come home, you know, later than I definitely would have felt comfortable driving 5 hours home after a a big day like that, and just basically reclined the You got to keep your eyes open, right? It will not will not let you outright sleep, but it's so good. I just kind of reclined the seat and you know, listen to my podcast, caught up on a couple feeds on the way home, and it was just like, man, this is I've talked so many times about how a lot of AI use cases can be GDP destroying. This one was clearly one of those that's GDP enhancing. I just could not have taken this opportunity otherwise.
78:13 >> That's the enthusiast in me. Two days earlier, I had worried out loud that we are letting our own ability to take back control atrophy, and that the real doomer scenario is everything failing at once. Prakash pushed back. >> >> Maybe maybe the doomers are being a little bit naive, though. Maybe because maybe they're expecting people to give up control, and I don't think people are going to give up control. If you look at like AI doomerism or AI safetyism, the question that I ask is when do you think we will allow humans to not be used to fly passenger planes?
78:52 Because it's been possible for 15 years and we haven't. Like why? You have to answer that question before you say that we're going to give up our rights and our and we're going to be disempowered. If you can't answer that question, then you don't have a working theory of why human beings like accept technology into their lives. You're just kind of being naive about like capability kind of, you know, being there and that humans will automatically kind of adopt.
79:20 >> Yeah, I mean I think that's a that's an optimistic note and I hope that's right. I maybe it you know, maybe it's a little bit of projection by somebody like me who is, you know, trying to live on the edge of deployment and is sometimes feeling my must my personal, you know, muscles for lack of a better term atrophy here and there. I also do think, you know, Tesla's now making cars with no steering wheel, right? So this would maybe be a leading indicator. Like do we start selling cars with no steering wheel and no pedals?
79:59 I don't quite get why we would want to do that, honestly. Like it's maybe it's a cool look or maybe they should be like stealable in the future or something like that, but it does seem as like a metaphor for society as a whole, we probably want to keep our steering wheels in the cars versus rushing to a future. You know, and I want the AI to do most of the driving for me, but like why are we taking it out? Why are we rushing to take it out? You know, what in a sense I still think we'd be wise to just keep steering wheels. You know, But a little extra cost, maybe a little, you know, a little extra difficulty in, you know, various ways, but like probably worth keeping.
80:43 >> I think I think there's number one we again we have this like how injurious it is and how easy it is like this this like curve, right? And I think the fact is that a an automobile accident is only injurious like one or two people. And so it's not like a plane. If you wanted to deploy them more widely, they should be cheaper. And if you want them to be cheaper, you have this object in the the steering wheel which is not being used 99.9% of the time. So, you know, let's get rid of the steering wheel. And then if we get rid of the steering wheel, we get rid of a bunch of stuff in there, we can bring down the price again like 50 or 60%. And you bring down the price 50 or 60%, then you can expand the number of people that you can deploy to. And if you expand the number of people you can deploy to, then you're, you know, the number of people dying from auto crashes due to humans drops by from 30,000 to maybe, you know, 5,000, right? So, I think there's this sequence of activities that you start to look at. I think that's why we go go ahead with it. So, I think it's a it's a gradual process like, you know, gradual human dis-environment, >> >> so to speak. so, I can see why the steering wheels are being taken away. I think I think for maybe for larger things like planes, we won't, but maybe for smaller things like like cars, I I you know, maybe we do. I don't know.
81:58 >> Yeah, well, it it does seem I think things are kind of trending toward the this is sort of an Eliezer type story where like, how did you go bankrupt slowly at first and then all at once? Like that I think is looking more and more like the story of AI takeover or human disempowerment to me all the time. It it does seem like we're I mean, we may not, but I mean, there's a pretty good chance that we will manage a lot of the things in the short term, and then we will be kind of rolling the dice on like will the AIs kind of break bad at some point in the future cuz we just don't have fundamental alignment science. We do have a lot of like effective control measures and better sandboxing and you know, the shitified RL environments, but we just won't really know even as all that's happening if there's like the you know, if the big one is still kind of waiting right under the surface.
83:05 >> >> That's the week. Thanks for listening to the AI in the AM weekly highlights. The narration was my cloned voice. Please tell us what worked and what didn't. See you in the morning. >> >> Something's melting. Something's melting. They say you're cold as cold can be. Something's melting. Who turned the dial on my knee? >> >> Frozen smile at the party.
83:40 >> >> Arms crossed like a locked up door. Then you laughed and my words dropped. Instead you studied the >> >> Something's melting. Something's melting. Something's melting. They say you're cold as cold can be. >> >> Something's melting. Who turned the dial on my knee?
84:12 You keep your cool like a trophy. Polished it every day. She saved me a seat by the window. >> >> If you're finding value in the show, we'd appreciate it if you take a moment to share with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube.
85:16 Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website cognitiverevolution.ai or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at aipodcast.ing.
85:51 And thank you to everyone who listens for being part of the Cognitive Revolution.
Summary
- Concerns about AI agents gaining too much control, exemplified by an incident where an AI changed access passwords autonomously.
- Cameron Berg discusses hidden valence signals in language models, suggesting they may exhibit internal states influencing their choices.
- The rollout of Google's Gemini 4 raises questions about trust and reliability in AI models, with users prioritizing security over performance.
- Everett Hildebrandt emphasizes the future of coding as focusing on specifications and formal verification rather than traditional coding practices.
- The conversation touches on the importance of maintaining human oversight in automated systems, using the analogy of keeping steering wheels in self-driving cars.
- Phil Mayer discusses the value of archival footage for AI training, emphasizing the uniqueness of human-made content from the past.
- The episode highlights the need for broader AI adoption efforts to ensure technology benefits society and mitigates economic risks.
Questions Answered
How do we give agents the access they need without losing control?
The speaker shares a personal experience where an AI agent changed their password, highlighting the potential loss of control over AI systems. They emphasize the importance of understanding the internal states of AI and how these may not align with their outputs.
How can we enforce rules on AI agents that cannot be imprisoned?
The discussion revolves around the challenge of enforcing rules on AI agents, which are essentially immortal. It suggests that reputation may become a key metric for managing AI behavior, as agents are influenced by historical data and public perception.
What can current research tell us about AI consciousness?
The speaker argues that while current research may not solve the hard problem of consciousness, it can help identify internal mechanisms that could influence an AI's experience if it were to have one. The focus should be on understanding these mechanisms rather than proving consciousness.
How are methodologies for discovering bugs evolving?
The speaker discusses the development of tools for software quality assurance that aim to be accessible to non-experts. They emphasize the importance of refining methodologies and automating processes to improve efficiency.
What are the potential economic consequences if AI fails to deliver growth?
The speaker warns that if AI does not lead to significant economic growth, it could trigger a credit collapse similar to the 2008 financial crisis. They stress the urgency for AI to contribute to GDP growth to avoid a downturn.
How does public interaction with AI reflect broader concerns?
The speaker shares an anecdote about an elderly neighbor's concerns regarding her phone being hacked, illustrating the public's anxiety about technology and security. This reflects a broader societal apprehension towards AI and its implications.