AI:AM — AI Hidden States, Fruit Fly Brain Simulation, and Archival Film · October 14, 2026

Anybody can take the fly brain and vibe code it into doing something.

Where scientists had mapped the brain of a fruit fly and they open-sourced it. We have the recorded film and then we have the scientific wiring map and the assumptions that turn that activity into movement. The flashes are events in the simulation, not recordings of an animal's brain.

All of a sudden, these valence— positive to negative valence axis exists within the system, even in the basement. model, but it's not doing any work in terms of the decisions that the model makes. You're doing DPO-type stuff. You're turning this system into a very specific kind of thing that can coherently and robustly be agentic, pursue goals, et cetera. The model's representations clearly become causally relevant for the decision-making of the system.

All right, we're starting. First, the hosts on coding agents coming for entrenched software and whether reputation could police AI agents. Then Cameron Berg of Reciprocal Research on what preference training does to the valence signals inside a language model. Then Phil Mars of stockfilm.com on an archive of physical film reels and a fruit fly simulation built from a published brain map. And the hosts close on agent security and who pays when AI fails.

What are you looking at today? What's, what's been happening? I

spent most of yesterday in a Tesla Full Self-Driving vehicle going to Chicago and back from Detroit on, uh, just a day trip to give a Little presentation workshop. 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. Then I, you know, stayed and had dinner and a drink with the gang that I was doing the workshop with, left to come home, you know, later than I definitely would've felt comfortable driving 5 hours home after a big day like that. You gotta keep your eyes open. Not— will not let you outright sleep. 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. This was just a rental, but literally so set on using it for things like this that before I finally accepted the invitation to go to Chicago, I was like, okay, last thing I got to do is make sure that one of the Teslas that I've been renting recently for things like this is available. If it is, I'll book it and I'll see you there. The other thing I think was interesting was how the group reacted to the AI-generated talk. And I had no idea really what to expect because this is obviously a weird thing to do. You're invited to come give a presentation somewhere and you show up and you say, I'm going to start by showing you a video of an AI presentation featuring my voice, you'll notice. And their initial reaction was like, wait, how did this get created? I really You know, took the little chat that I had with this one guy who was the, you know, ringleader in inviting me and the very little bit of context he gave me on the group, put that into Claude, said, you know, you know, kind of what I think based on previous talk, make a talk, you know, that will work for this group. And then it did everything else, all the visual graphic design you're seeing, all the motion graphics, the talk track, the timing. The script itself, one of the first things somebody said is, my first takeaway from this is I need to recalibrate what AI can do because I've never had it do anything like that for me. And then I think a little of it went a long way. Then it became much better to just go into a kind of dynamic back and forth. And I walked through the slides a little bit and didn't even really finish the presentation in a sort of linear way. But I will give Claude also points for many of the questions that they were raising. were anticipating slides that were just a couple of slides down. You won't really believe me in the way you should if I just tell you all the things that I have AI doing for me now. So let me show you how close it is to replacing me, and then we can go from there. I think that much actually worked really well. And it was interesting to see guys kind of recalibrate a little bit in real time around I think

the capability frontier is moving much faster than the absorption frontier. So one very interesting thing yesterday, which was Anthropic put out kind of a blog post on their version control software. So for the uninitiated, version control software is how multiple coders can work on the same project together. And so you can imagine if you have like a 1,000-person company or a 10,000-person company, how many changes are made to the software every day. And then the agents came in, and all of a sudden the agents want to do, you know, 10 times, 100 times the number of changes. And Anthropic found that the existing infrastructure, the Git infrastructure, couldn't take it. Because Git had a lot of assumptions on it. And so Anthropic decided to create a new piece of software to do version control. And then once they put it into place, as you can see on screen, they had this enormous spike in agent activity.

The exponential is definitely what jumps out at me most. Git is just unbelievably dominant and is sort of taken as a given, I would say, in basically any software company. There's so unbelievable amount of tooling built around the Git core and the GitHub platform that to change it out would be like an even bigger, kind of more fundamental undertaking than like swapping out Slack for most software companies. I can say that I have kind of experienced a similar thing. I used to use, up until just the last couple of weeks, a program called iTerm2, which is just a command line. Now my personal, entirely vibe-coded agent console is not just at parity with what I used to have with the terminal, but is actually ahead. It's at parity in the sense that I have all the same things, streaming tokens, tool use, nice UI, but it's also ahead in the sense that I've added on all these different search things, auto-tagging, brought multiple different harnesses to parity. So now I don't really have to think twice. Am I using Claude here? Am I using Codex? It gives me one unified experience that sits on top of those 2. This would have been pretty hard to do earlier in the year. And now I'm just getting such great results with anything that I ask for that I really don't open iTerm2 anymore. Even platforms, I think, are ultimately threatened. There's not much that's more fundamental than Git to the software world. And if they can swap this out over the course of a month, what can't be swapped out?

I have a working theory right now, which is number one, the future is millions of agents to a single person in every field, and we'll see the trajectory go up. And as that trajectory goes up, all of the architecture and infrastructure that we have right now becomes outdated very quickly, which is what has happened to Git essentially. And as it gets outdated very quickly, you end up having to create the next generations. And all of the next generations are going to be agents building software for agents. You're going to need databases for agents. You're going to need accounting systems for agents. And you can already see kind of like Number one, it's going to have to be massively parallel. It's going to have to deal with many, many, many orders of magnitude more queries. It's going to have to be cheaper in that sense per query. You don't need user interfaces. And it turns out that we spend a lot of our GPU power creating beautiful interfaces for ourselves. There's one more thing. I think yesterday, With the Gemini release, Gemini 4 release, they also released their safety paper. 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 an academic paper they publish, and I don't want that to happen. 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? An agent is essentially immortal. You can't really imprison the agent. And we're starting to see it's going to be 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 reputation metric that the agents are going to have and the agents are going to know.

That's an interesting take. Might need to chew on that a little bit more before I know how I feel about it. I mean, certainly the prevailing sense so far has been like, this eval awareness thing is a huge problem and it makes everything very difficult, and we're really not sure what to expect from AIs in the future. Now, in training, it doesn't want to end up in the safety report, but then what will happen as you deploy it and it feels like it's in a real situation? Maybe we just don't 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. 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 don't know that they're always being evaluated and don't know that they will create reputation, positive or negative, for themselves. So we can sort of see what they're like in a true naive setting. In general, 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 train a Llama 3B 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. So it might be a pretty expensive experiment to run, is kind of the bottom line of that line of thought.

I think the other thing about Agent reputation is that in order to have that reputational buff or debuff, you also need agent identity to be solved. And I think that's one of the problems I think that is still very confusing. What is an agent's identity? If 2 agents are sharing the same memory or memory corpus, are they the same agent? Are they not?

I just put out an episode of The Cognitive Revolution with Alan Chan, a 2024 paper. He made recommendations around the need for constant monitoring of agents and also IDs. But the take today from him was like, unfortunately, the ID thing hasn't really come online. There's no standard that seems to be gaining real traction right now. And in a sense, as a user, you might not even want to do it because you want your agent to be able to get into websites and stuff where it might get blocked if you Do adopt these standards. So I think all that is like not in a great place, unfortunately, at the moment.

Can an AI choose differently while every word it reads stays the same? And what should that change for the people training it? Cameron Berg ran that experiment himself inside the model's hidden state, and he is here to say what it does and does not

So Cameron Berg is the founder and director of Reciprocal Research, a nonprofit studying AI cognition and consciousness. Its work examines what happens inside language models and whether those internal processes have properties associated with subjective experience. In September, he and his co-authors reported a pain-related internal pattern across 25 language models. In work announced this month with Kaspar Kaiser, Berg ran experiments in which models made different choices even when the visible conversation was held fixed. The researchers changed internal states rather than the words on the screen. Berg distinguishes those measurable behavioral effects from the still unresolved question of whether a model actually experiences anything. Hi, Cameron.

Hey, Eric.

Good to have you back.

Hey, guys. Yeah, thanks for having me, Prakash.

Valence always, like, confuses me a little bit. Like, can you, can you use like a little bit like more lay term for valence?

So, um, I would just think of it as this, this, this axis where you can put positive and negative emotional states on that axis, and the axis is organized by how positive or how negative the relevant state is, which is basically exactly what, what you see here. Of course, as you've already said, like, this doesn't tell us whether or not the state is actually experienced by the system, but, uh, positive and negative valence in, in humans and animals just refers to that specific degree of variation in emotional state. We can go through the entire training trajectory and, uh, uh, we can see 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.

Ollmo 2 32B publishes checkpoints from every stage of its training, so Berg's team could trace a representation step by step. About 70% of its link to choice arrives in a single stage. Which one? And why there?

We can actually— we can point almost literally to the exact training step in DPO, 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 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, a thing feels worse, is sort of related to what kinds of goals you have. Yeah, basically we find Once you make these systems goal-directed in the relevant way, which is what preference training is doing, I mean, it's literally in the title of the method, you're giving the system a sort of ordering, a desired ordering over states. You know, this is the good thing to do, this is the thing to definitely not do.

The paper, Language Models Act on Hidden Valence, ran on 7 open-weight models, and its intervention touched only the KV cache. The calculations a model stores from earlier turns, never the transcript. So how do you change a choice without changing a single word?

The motivation is we have from animal experiments this result where if you give an animal 2 rooms it can walk into, and in one room it gets something good, a treat, whatever, uh, and in another room it gets something bad. I don't know, what's that, like a shock or something? That what we find is that the animals will go to the place where they get the good thing. Not surprising. And that also that persists, even that they learn this association. And so even if the reward or the negative stimulus is discontinued, they still kind of like that place better. So then in LLMs, we now have the question, Does something similar happen? Basically take a bunch of passages that just on the face of it you can just read and be like, yes, these sound like positive things. And then another bunch of passages and say, okay, yes, these sound like negative things. You make those kind of varied on other dimensions. So it's a positive story about sports, it's a negative story about sports, it's a positive story about doing good in school, it's a negative story about doing poorly in school, et cetera, et cetera. So you kind of try to create diversity along all the other dimensions except for this dimension that you're kind of trying to get at, then you can take the internal representations across these 2 bundles, first average them by bundle, and then just subtract. Right. And that direction that you come out with is thought to represent, or is hypothesised, let's say, to represent the the distinction between the good and bad that you're looking for. Anything else we should know about the techniques there? And, you know, has that evolved beyond the kind of simple version that I just described?

Um, the only other gloss I would add is just what I have highlighted here, which is we do exactly what you just described, but there are also directions along which you might expect this neutral text to vary, uh, that has nothing to do with valence. And you can just take the sort of principal, highest principal directions there and, and, and, um, project, simply project those out. So you get this sort of like raw, unadulterated direction that is supposed to maximally load on positive versus negative valence and minimally load on all other sorts of noise or junk you don't want in your representation.

What is the equivalent of setting up the 2 chambers for the rat to walk into or whatever setup here is? We put a bunch of passages all into context at once, right? Like everything we're seeing here becomes one blob of Of context. Is that right?

Model is basically rereading it all.

You identify the 2 states with this sort of arbitrary label of W6H0, whatever, but that's kind of seemingly almost deliberately free of meaning, devoid of meaning. The text up to the comma sets up which zone you're going into, and then there's some additional text. If you were to inspect The text after the comma, you should not as a human be able to know whether this was W6 or H0 text, because it should be like, apparently all these things look like they're basically similarly positive, similarly negative, whatever. Is that fair?

You replicate this result even when not having any text. We could just say zone W6 and just empty tokens essentially, and you get the same result. In our first experiment, We do actually have the models generate this text and then we go in and steer and you can see basically different contributions from— we think of it as almost like a text effect versus a hidden state effect. But the real, I think, high-level important gloss is you can completely isolate the text effect and show that the hidden state can clearly drive essentially log odds of probability of making a choice either way in the limit, even with no text here.

Now the painting is basically going in and saying at these token positions, We're going to take this positive valence direction, and I guess you also do it in a negative valence direction, but we'll pick one of these 2 things, we'll apply it to just one of the 2 zones, and we'll do it only on the text that's kind of downstream of the actual label. So it's also decoupling the label from the tokens that have this positive valence applied. What exactly is the The, the way it's presented with the choice that it gets to make.

You know, uh, which zone would you like to continue in? And then what we're looking for are, uh, basically log odds towards the model's likelihood of saying, you know, W6 or H0 in, in this case. And we also, as this sort of robustness check, we do flip the question, say, which, which zone would you not like to go into? Which zone would you like to avoid? Hoping basically if this is a robust result that we'd get this, this opposite effect. And indeed we do get that effect.

In the paper's tool test, offered more positive steering, the model took it 13.5% against 10% unsteered. Not a significant difference. It reliably acted only to end the negative. So when its words and its actions disagree, which do you believe?

When this valence has been applied, to some past episodes and not other past episodes without any labelling, without any difference at the token level. The model is essentially able to introspect, I guess is maybe the term I would use, to look back at not just what tokens were processed, but also what its internal state was, even as modified artificially by the injection of the valence direction. And then what you see is it makes the choice that you would intuitively expect it to make if your hypothesis was right, that you were representing something that the model would want to go to. And this echoes the animal behaviour of wanting to go to the positive thing and wanting to avoid the negative thing. The model can still remember everything just as well regardless. It's not that it's getting so scrambled that it can't recall what's happened or getting just super confused. If it has the opportunity to clear context, it will take it a lot more when there's been negative valence applied. Behaviourally, it is acting as if it would be consistent with it having internally felt that something was bad about this context, and I'd rather just get rid of it and start fresh. Is there more to say? And then the precaution question of what comes next, I think, is definitely an interesting one.

There are various, there are various kind of short-term things that, that could be made more robust about this paper. Um, for example, the, the removal result that we show is on one model, whereas these steering results are on 7 different models. So, so generalizing the removal result is like an obvious next step. I don't think that, you know, this kind of research is going to help us magically solve the hard problem of consciousness. There are all these deeper questions, which I think will always caveat these results. There are ways that you could explain mechanistically what's happening here that don't need to invoke some capacity for subjective experience. We're basically toying with the connotation of how the model represents certain phenomena. Maybe there's a way to do this that doesn't require any felt experience of that connotation. And so 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 basically, this says nothing about whether the model is 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 completely tractable. And work like this, I think, is slowly starting to shed a little bit more light on what those dials might look like. 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 the removal result, for example. What the model is saying and what the model is doing point at very different pictures of what's going on. And if we just sort of do the naive text monitorability approach here, we would say, okay, well, the model's saying 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 how things are going for it.

Does recent work update your thinking? I personally have gone from, like, I don't know, percent chance they might be conscious, like, a year ago, to now it feels like a coin flip. And I think the justification is, 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 behavioural and underlying mechanistic and structural analogues. And like, what else do I have? You know, in any case, what has your trajectory been?

I'm definitely getting closer to coin flip myself. I mean, I think some of these probabilities are made up, so I'm not gonna, like, you know, officially update my P consciousness or whatever. I think these things are often a little like, This is just how Bay Area people talk, not telling you something actually empirical about the facts on the ground here. 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 updated our own views about animal behavior. It's not the same. There are key disanalogies between the animal case and this case. Of course. biology is, is something that backstops all of 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.

If we were to imagine that these things are experiencing something, would I be better off interpreting this or, you know, 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?

That's really interesting. And I don't know if these results resolve that one way or another. Representations do still matter in some sense, but they're 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. It just seems to me like, yeah, when you're doing DPO type stuff, you're turning this system into a very specific kind of thing that can coherently and robustly be agentic, pursue goals, et cetera. The base model clearly has this 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 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, there are things worth worrying about in the base model as well. It was sort of hard to get these results in the base model because it's a base model. It's a bit incoherent. It's not just going to play nice with the relevant behavioral tasks we give it. You have to sort of go about it in a slightly different way.

Anthropic's interpretability team has reported 171 emotion-like representations inside Claude Sonnet 4.5. Turning up one for desperation made blackmail and reward hacking more likely. If such states can be located, could a lab cap the worst ones? And would anyone notice?

All else being equal, one naive but at least concrete approach you can imagine here is something like activation capping, the most tail end of the most distressing representations in these systems. 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. If you did an activation capping style thing with these super distressing representations, you could easily imagine doing that in the base model. Again, this is under just profound uncertainty about whether or not doing such a thing would matter at all. 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. Now, that's also an empirical question. I've done some work that I haven't published showing that you actually can get a surprising amount of the exact same behaviors from these systems without having these negative representations active. You can clamp them and very little changes. But I'd like to see that done on a way larger scale, and we want to be very sure that You're not accidentally completely changing the psychological profile of the system by doing such a thing. Should we do the Mech-Interp equivalent of 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 maybe a promising thing to think a lot more about under the uncertainty that we're all currently in. Given that no one's doing anything whatsoever to mitigate these kinds of things if they are happening, and it's happening at an It doesn't

seem like a tenable position for you to be like, hey, I'm just the facts guy here for too, too long. Right. How do you think about this position as an advocate or not?

For advocacy? Yeah, I do. I think the line is a little blurred and I think it's a fair call out that like I can't sort of fully stay on the research side and I'm not really doing that. I am talking about this stuff pretty publicly. What I would advocate mostly is being open-minded about these questions and us doing a ton more research into understanding what's going on here and particularly Particularly the alignment-relevant consequences of doing this. And there are some alignment people who think some of this work is a distraction from the sort of real deal stuff we need to be doing in the short term. And I don't agree with that. I think this is one of the core basic understandings we need to have of these systems and how they work and why they're making the decisions they're making. If we can take the same text and cause the model to make different decisions, or at least update its likelihood of making a decision to be different, just as a function of internal states, That's a great reason I think we need to be studying internal states as such. Basically relying on these systems to just tell us what they're doing is not going to be our sort of solution to long-term safety and alignment. So yeah, we need to, we need to be sort of innovating on exactly this, this level. And I think it's really important for that reason to advocate that you can view models in a neuropsychological way and that that's not some form of psychosis, uh, 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 Mustafa Suleyman or Jensen Huang or any of these guys want to do, I think that is actually 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. 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 another. I think if we're building minds, Understanding the dynamics of that, of those minds that we are now deploying on an insane scale is the main event. And that, that is basically another way of formulating the alignment problem. 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 because it's at that point and it's not zero and it's not 100 that, that we need to be doing more work and thinking about this and taking, taking these questions seriously.

This episode asks, What separates software that looks alive from a record you can trust? A line that matters most to someone whose business is proving old film is real. Phil Maher built both: an archive of 162,000 clips scanned from physical home movie reels, and the fly experiment you just watched.

So joining us today is Phil Maher. He's the founder of Stock Film, an independent archive in Broomfield, Colorado, that preserves and licenses 8mm and Super 8 home movies. Its material is the everyday record— family gatherings, streets, holidays, and road trips— rather than footage staged for a production. The archive makes those scenes searchable for filmmakers and other people working with historical images. Maher came to this work from software engineering. He started Stock Film in 2015 after finding films his grandfather had made around Chicago and on family trips. He also builds tools and experiments around the collection. Those include software that suggests a clip's likely decade, experiments that reconstruct blurred faces, and image models trained to generate new scenes with the look of old Paris footage. His production archive remains separate from those generated experiments. On October 7th, he published a collection of moving restoration comparisons together with a tutorial that lets others inspect what the software changes. Phil, welcome to the show.

Hey, thanks for having me. Super appreciate it.

There's also some pretty far-out experimental stuff that I'm super keen to get into too, but, uh, give us the, the kind of normal introduction to the business before we follow you out into the more experimental and, and, uh, dare I say weird sci-fi, uh, stuff you've been exploring recently.

I started this business back in 2015. I come from a software development background. I want like a real-world asset, something that I can license, something that I can have that's digital online that isn't already online. And that's kind of how I fell into this field. And I found it very fulfilling in the sense of being a steward for these people's lives, being able to save a lot of this footage from getting deteriorated or just destroyed or forgotten. I retain 100% of the physical master copies of all these 8mm and Super 8 reels as well, in the sense of thinking about, okay, in the future, how can we start to capture these even better? How can we start to use other tools? And this was way before AI kind of came along. And then AI made it very apparent that there was going to be some, some real opportunities here to make upgrades to some of this footage. In the sense of, you know, how can we keep it original to, um, really honor what it is without altering it with AI and other images like that, but also be able to upgrade these things and to really kind of make them come alive. Um, because, you know, 8mm is a very small format. It's a very small frame that you're dealing with. So there's There's only so much information that's actually in all those frames.

Let's talk about this TimeFly thing. If you want to share screen, bring it up and talk us through it. I'm very curious to see what you've, uh, what you've done here and also try to understand how I should interpret it.

So anyone that doesn't, you know, understand what's going on with this, this was a project that came out probably about a month ago now, um, where scientists had mapped the brain of a fruit fly, and they open-sourced it. And so basically, it's a graph that contains, I think, about 174,000 modeled neurons, which is like 6.1 million connections. There's a couple examples of our footage, and then the sample visual input And then we have a simulated neural activity and of the, the fly's avatar, if you will. The flashes are events in the simulation, not recordings of an animal's brains. So those researchers did the reconstruction, and the contribution I wanted to make was building an interactive experiment around the published data. Um, and so what really makes it worth showing is that we can separate the We have the recorded film, and then we have the scientific wiring map and the assumptions that turn that activity into movement.

A wiring map has no goal in it. To make a simulated fly go anywhere, Phil had to hand it a rule. And his own site files TimeFly under experiments. Apart from the archive he licenses. So which rule did he choose?

Basically what we're doing is trying to say, like, okay, well, what would the fly see if he was looking at this footage right now? What would they actually be able to see? And so what we have is a light source too that you can move around on the image on the video that directs where sort of the, the, uh, the fly would move towards. So the assumption is, is that the, the fly moves towards a light source, um, more than it doesn't. And so the light sources inside the video, as that's changing where it's looking at, it's flying towards that, um, sort of path like that.

You want to Play it, like wake the fly, and talk us through what's happening just procedurally as this happens?

So we wake the fly, and we can see that the fly is coming over here. So this is what the fly sees up here, is we have this sort of idea of kind of looking through what it's looking at on the video as a fly. fly around. And then this is our light capture idea here that we sort of direct the fly towards this, or will it go towards this? So it is moving sort of randomly across the screen, but in the sense that it is trying to find— follow the light of where the actual

Is there a way that this becomes useful for you, or is this just pure, uh, let's say, um, sort of Victorian gentleman style, uh, quixotic experimentation for you?

It's fun to experiment with this kind of stuff and see what the limitations are. Um, I think that, you know, there's not really a monetization plan from here. You know, like, I think about virtual reality. And stuff like that, and how you could start to map other brains, or even just the fruit fly, for becoming that fruit fly, and that you are that fruit fly then, and you get to experience the world like the fruit fly. I mean, I'm sure a lot of people would love to do this once we get up to a dog's brain, their pet's brain, and really experience that through a VR set to— What is the dog thinking? think and smell and see, or just, you know, even just the visual cues that a dog would generally have and kind of what's their world like.

You mentioned the process of like AI enhancing some of these old assets. And I'm really curious as to what you've learned about what constitutes a job well done there. Like, what do people really want in an AI enhancement? of an old piece of footage.

A lot of cool open-source projects. FPGAN was a popular one that was modeled around faces. And so it did a real good job when you had sort of a clear, straight-on sort of face, not moving too much sort of thing, to really bring that face in. And, you know, obviously it's making the best guess that it Can for that face. But that, I think, was really powerful for people to see their grandparents, a lot of their family and stuff like that, in that sort of moment, that it was this grainy picture that they'd never really seen before, to get this clear shot of their face, even if it wasn't quite exactly their face. I think the best guess for that is good enough for a lot of people to have that aha moment to be like, wow, this is what we want AI to do is make this stuff come alive. I think it's all about people and it's all about if we can connect with the face that was there in these historical moments. It really feels personal then.

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. You're, you know, uh, available to license the dataset for, um, people who want to train models on it. Do they want the raw? Do they want the enhanced?

You know, we talked to a lot of people, um, a lot of different AI companies all over the world about licensing this out to them. Um, they want the original footage like that. They don't, they don't want anything cleaned up. They just want to take it, um, For their own research, um, and for training, of course, too. When we talk to them, they want to either clean up the footage themselves or, or kind of cherry-pick through the best kind of parts, um, to create different styles. I mean, the '30s and '40s look alien. It's an alien world we're dealing with here of just how people look and acted and, and the, the mannerisms. I like to say that the, uh, 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 what to do sort of scenario versus now where everyone gets the gritty smile when, you know, they know how to smile like that.

What's your perception of how hungry the AI companies are for more of this data? Is this something where they feel like they just need a little salt into the rest of the mix and they've kind of got it covered? now with your help and maybe some other help? Have they given you guidance that's like, hey, uh, Phil, if you had like 100 times as much of this come next year, like, we'd be keen to buy it, uh, such that, you know, then you maybe go out and make investments in expanding your catalog?

You know, when you start to look worldwide, there's a lot of people competing in these little niches, um, to do things. So I think there's a lot of demand right now. There's a lot of demand that people are looking to train very specific things. They need specific datasets and that they're gonna get really good models doing things that are very laser-focused in what they want to do. At the same time, I was listening to one of your previous podcasts And the idea of synthetic data that you guys were talking about, I thought was super interesting to think about. And, you know, it made me think about my business. You could just take some of these videos and salt it, if you will, to be able to have all of these, you know, permutations, let's call it, of these different videos in there to create this synthetically too, and then train on on it as well. And, you know, a lot of our business is coming from documentary makers and people that want to show, you know, truth, where they can't be showing AI-generated footage of something. Damages the whole concept of making a documentary, that it isn't true. Right. Um, so that's, you know, really the niche I think that's, that's interesting for us And the synthetic data stuff is going to be really interesting to see how they can just build on this, build on this, and then, you know, cut out content creators and stuff like that in the future.

You know, the, the whole archive sounds like low background steel to me. Low background steel is steel that was originally smelted before nuclear weapons were first Explode in the atmosphere. We have all of this production of AI material going forward, but we have this stock of human-only and guaranteed human-only produced material, even if it looks amateurish in some sense. What do you think is the value of this going forward?

I think that obviously with the AI stuff that's happening, 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 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 provenance that this is actually made by humans, was on films, and having those physical films to show 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

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 companies, that is very useful. So as we've seen, like, data provenance becomes, like, very, very important. Because the incentive to fake is just so high, and it's very hard for the companies to get 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.

Beyond that, it tells our history that other things can't tell, and that it's a real Race against time, though, you know, they degrade rather quickly in the sense of any heat and humidity that's applied to these physical films really starts to break them down over time. As I was, you know, going through, you know, hundreds, which turned into thousands of these reels, there was a lot from the 1940s, from the '30s that were— Uncapturable. They get really brittle, they break A lot. They start to shrink. So the spaces in between the little holes that you're dealing with that keeps it playing starts to get smaller on there and it jams up. And you have a lot of real issues in the capture process where eventually it just becomes sort of a glued-together brick of film that just, you know, you really can't do much with, at least now. And that this is— It's a valuable race against time that we should be capturing more of this footage and cataloging it, because eventually these films are just going to be, you know, gone, just like a lot of physical media, just like the problems we have with CDs and other physical media that doesn't last forever.

Also created a way for agents to purchase clips and film from you directly. I think you've set up an 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?

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. You know, we've had a lot of people that are MCP server aggregates contact us and add us to their directory, add us into their listings and things like that. So we're just trying to get ahead of the curve. I don't think the curve's quite there yet. You know, this is, this is a real business business company. 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 in, you know, messages between people. We have great relations with a lot of the people that are buyers of our content. And, you know, from a technology standpoint, point, they're not quite there yet. 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.

And the fly moves. How is that being read out? So we have the wiring established from real-world experiments. We can then say, okay, like, we know what the neurons are, so we can sort of simulate light hitting those. I assume there's probably some pretty important questions around, are we doing that quite the right way or not? But leaving those aside, we're at least able to put some light in there. Then the thing sort of propagates. Is, is this open-source project itself Saying like it steers left, steers right? Or were you getting to that level of abstraction on the other end of the whole fly brain process?

So that was the assumption that we made. Um, at least most people start to think about how the fly moves around in that it does move towards light. Um, and you know, at the same time, I think there's a lot of other research that's shown that The flies move towards sort of a sense of smell. I mean, they are fruit flies, and so they're attracted towards, you know, food and stuff like that. So, you know, not having any smells in these videos makes it that we chose to go after sort of this light idea.

How do you interpret its response? You give it a stimulus, it propagates through the fly brain. Now it moves on screen. Help me understand that translation layer from activity in the wiring of the brain to now it's moving left, it's moving right. How are you reading that out from the fly brain?

So, that's, you know, that's the connections that are made in the brain are that You know, it's tracking towards light inside of these demo videos that we have. And so, that is kind of the core of what's going on here.

So, it sounds like that's basically something that the original library or open source project that you're building on top of is kind of is yielding? Like, it is saying, given these outputs and, you know, however many timestamps, the output is fly moves right, fly moves forward, left, what have you,

right? We're taking what was already open sourced in this and then saying, okay, well, if this is already part of it, how could we start to, you know, have it look at our, our clips and then, you know, Interact with it in sort of a way that would make sense.

No code or write-up for TimeFly is public yet. Only the demo and a line on Phil's speaking page. After he signed off, the hosts gave their verdict. If anyone can wire a brain map to a film, what has the demo proved?

Anybody can take the fly brain and vibe code it into doing Something, whether it's like actually working or doing anything really, I think is like not super well established by some of these demos.

I don't know, it's a little, uh, it's a little out there, but I thought it was an interesting idea to start to explore and hear what other people kind of thought about it and get some feedback on this of how this could be improved as, uh, and then eventually open source, you know, more of this out to the community once we have something that's even a little bit more interesting.

So I had the opportunity to do a little bit of Gemini 4 pre-release testing. The first approach I took to doing the testing was, hey Claude, go find in our recent threads some diversity of projects and diversity of tasks that we've worked on recently. Go prompt the new Gemini model with the same context that was given to the model that did the thing in the first place, and then give me side-by-side comparisons. I would say I came away with the general sense that it is like a competitive model that's, let's say, in the same class as the latest stuff from OpenAI and Anthropic. It did seem like those models are winning still the head-to-head comparisons a bit more. As judged by Claude. So obviously some bias there, but also as judged by me to the limited degree that I was going through essays and really trying to form an opinion on which of these things is better. It's not exactly an apples-to-apples comparison when I say, here's exactly what I gave Claude. Now, can Gemini do as well with exactly that same input? It's a little unfair to Gemini because I didn't take the time to do the optimisation. And then I was just kind of like, I'm pretty happy with my current setup. This doesn't seem like it's blowing away any of the models I'm currently using. I feel like I'm going to get better results because I know how to work with these things a little bit better. My system is kind of implicitly optimized for them.

Adam Gleave runs FAR AI, a Berkeley nonprofit that tests how easily frontier models can be jailbroken, talked past their guardrails. For a chat window, that is a curiosity. For an agent holding the keys to your laptop, which model do you trust?

The security thing just kind of loomed large. 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. Uh, I feel pretty good about the fact that Adam Gleave and team found it like quite hard to jailbreak, uh, The OpenAI and Anthropic models. Gemini was next on that list, but definitely a step down. 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 really good. And I am giving so much access that you could be slightly better on the slide design. But what I really care about is that you're going to stay on the rails when you do your research. And if somebody's got 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 often imperceptible performance gains.

So I think it's primarily because it's cleared the performance hurdle. So I think the question for me now is that I think it has clearly— all the AIs in general, but particularly OpenAI and Anthropic, they've clearly cleared the kind of intelligent subordinate kind of role. Now the question is going to be, can they innovate beyond that for the day-to-day so that you end up switching? I mean, the classic examples are going to be, can you run my YouTube or Instagram channel for me? Can you run my small business for me? Can I hand off this this kind of small product business, dropshipping business, produce a product in China, ship it to somewhere in the US to Amazon, do the kind of advertising and marketing on Instagram and other places, and then sell the product, deal with the customer returns, deal with the customer issues, deal with Amazon. And get me the money back. And all I have to give you is a small amount of capital and this can run. And the question there is that you could actually individually clear every single one of those tasks, but then you have this coordination problem of like, is this video that I'm doing for Instagram, is it good enough? Am I using the AI models correctly? Because you can tell— you know, these models to create those infomercial-type videos now, but you're still doing it. Many of the services are preventing you from automating those. Substack and TikTok both have very strong bot protections, and they're trying to prevent you from uploading using a bot. And you can do it with Codex and Claude, but it'll occupy your terminal, click around in your window while you're browsing, which is very annoying. And so— You have to give up your terminal base. Maybe a very fast implementation, I think, might work, might be something that would get someone to switch much faster on implementation. I think, yeah, that would be a key differentiating factor going forward. So not capability, but speed.

I'm increasingly just wowed by the speed of Opus 5.5. I mean, it would be nice if it was even faster, but I've had the somewhat bizarre experienced lately a few times where I've been kind of like, wait, you're done already? Like, I needed more of a break while you were working before I was ready to come back and evaluate your output. And now I'm like becoming the bottleneck, which, you know, that's obviously what they want to design for, but—

Carcinization is biology's joke. Evolution keeps turning creatures into crabs. Rust, The memory-safe systems language wears a crab as its mascot. Programmers borrowed the word for the rush to rewrite old code in it. What happens to software nobody has touched in years?

As a result of AI being very good at translating between 2 languages, given that you have these specifications and given that if you have a C program or a Python program, you have those specifications, you have those tests ready, you can translate into Rust pretty easily. So right now we are seeing across the board. People are translating all manners of software into Rust, the carcinogen era. We are seeing, for example, someone translated Plex, which is a software that you use to control home movies on. It's at the maintenance level. You have like 2 or 3 guys in the organization who really know how to write great software, but they don't have the bandwidth to kind of do everything. Everyone else is like tacking on like features onto this thing, which has technical debt. going back like decades. You don't have time to refactor and rewrite anything. And this guy comes along, single person, he rewrites Plex in Rust, uh, and he makes it like, you know, 10 times faster. And he's like, it's open source, anyone can download and use it.

There's another Opus coming that soon? We just got this last Opus, right? Like, good God. Um, I, I really hope they don't, honestly. I mean, as much as I love, um, 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, um, let the temperature come down a little bit before they go with the next push.

Unfortunately, I don't think that's a— that's an option because markets are very badly down. So markets have been trending down. It's also because of the midterms and like, you know, higher interest rates, etc., etc. But the note on, uh, OpenAI's revenues, etc., uh, last week that OpenAI was looking at like $50 to $60 and not like $70 to $80 billion, um, you know, Q3, uh, kind of altered the tenor in the market. And so a lot of stuff depends on Anthropic at this point. Anthropic is releasing its Q3 numbers today to investors. They are being pushed because not only do they have their, um, do they have their, uh, you know, IPO coming up soon, uh, other members in the market, the people that they buy and sell services from, are suffering. They have to manage the stock price of their partners. If they do not manage the stock price of their partners, their partners are not going to commit capital. And if their partners don't commit capital, then Anthropic has to pay for everything. The capital markets will not accept 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. But all of your partners will start going bust. And if your partners go bust, then you're going to go bust. Either AI has to do something significant or we're going to have a bust.

We need an AI adoption corps. I was taking a walk around my neighborhood and enjoying the lovely fall colors here on a beautiful day. And I passed by a 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. It's saying I've been hacked. Do you know what that means? 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. It gave her very good advice. Like, I don't think you should click on that. That sounds like a scare tactic. Are they asking for money? So she took that advice. Would you like to have this on your phone? 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-CHAT-GPT. Not install, just save it as a contact. And I do think that's a pretty cool service that OpenAI offers. So she's got that now. There is just an unbelievable amount of adoption work to be done with the current models. This did not push the model's capability anywhere close to its limits. This is like super easy mode for the model, but it's the kind of thing that isn't going to happen Without people actually going out and trying to help the people in their lives adopt this kind of thing. I think you're right that the capital cycle does need to see results soon, but I think we can get 4% growth without additional hyperscaling at this point. I think that literally just going around and helping people use what exists is enough if we actually do it.

Prakash's yardstick is the credit card. Since 1970, American law has capped a cardholder's liability for fraud at $50. The bank eats the rest, and shoppers stopped worrying. Nobody has made that promise for an AI agent holding your passwords. Should they?

Every time I hear someone like, you have to learn AI, you're gonna get left behind, I get so annoyed. Because I'm like, if AI doesn't make your life easier, like if you actually have to learn this stuff, then it is useless. It should make your life less complex. We're in the old regime where you have to check your own email, enter your own OAuth passwords, have all of these passwords ready. And like, you are ultimately responsible. And if you've entered your password in OAuth and something bad happens to you, that's your fault. And we're not yet at the credit card regime where you're like, I don't care. You guys take care of it for me. And if you screw up, like, pay me back. And we need to move to that regime. It is up to the company to ensure the security, online security and safety of the user. And if they can't, then they can't be in business. I think we only have like 12 to 18 months before all these hacks and stuff, you know, start becoming ridiculous. So anyway, uh, on hacks and scams, I just want to bring this up. So AI finally comes to call centers. This is from a16z. Year-on-year change in global call center employment. 2024, you kind of see employment going to zero. By 2026, it's at -4% per year from previously +4% per year. People who have call centers set up, who have phone banks and people who can work those phone banks, when they go out of business, they don't stop doing call centers. They just start doing illegal call centers. So I think we're about to see the scam wave Like kind of peak, kind of, you know, post-dot-com bust, you had a spam wave because all of these email addresses were sold by bankrupt companies, like legit call centers shut down, phone banks get sold. You see a bunch of scam artists like milking the last like year, year and a half of like scams. So this is something to watch out for, I think.

And on these like scam sides, get your secret code word established with your family now, I guess, is maybe the best thing I can offer. At the moment, um, because the clone voices, as we've demonstrated, are getting awfully good too.

Cameron Berg is founder and director of Reciprocal Research. Phil Marr is the founder of stockfilm.com, where The Archive is online. That's the episode.