AI:AM — OpenAI Astra and Always-On Home AI · September 2, 2026

The looping is in the inner layers, and the inner layers you basically add depth, um, you know. So that the way the looping works ends up being in a normal transformer, you have this external loop, which is the chain of thought. So a token out is read back in. And so when you say, let's say, first find x, you get 3x equals 21, so x equals 7. So you can read that on the outer loop. You know, on the chain of thought, and then you get the answer. Check 21 divided by 3 equals 7, answer 7. Now on the internal loop, the recurrent depth, all the monitor sees is the answer effectively. So you get directly the answer is 7. Skip the first find x, 3x equals 21, et cetera, et cetera. So that part gets skipped. The external monitor doesn't see that. That part ends up getting processed internally. And that thought process in the middle is not exposed to the monitor. The concern on the chain-of-thought monitoring is that the chain-of-thought monitoring without the Loop Transformer is more revealing than a chain-of-thought monitor with the Loop Transformer, where the expectation is that the amount of Granularity of what is going on in the chain of thought, uh, basically decreases.

Will the US government take control of any AI company or project before 2030? This is operational control, not an equity stake, not a golden share, not a contract.

What does operational control mean? Like replace the CEO?

Conservatorship, receivership, or similar legal mechanism where the government directs strategic or executive decision-making. Regardless of equity ownership percentage. They certainly leave some room for interpretation. Um, but here's kind of what we've got. So they could like nationalize it outright and own it, but if they have any strategic or executive decision-making authority regardless of ownership, that would count. Could happen via court order or executive action. Could happen via any law.

I don't know why.

Going forward, do you think that you will have like voice agents call these folks on people's behalf?

Because, so we've tried this. It's very interesting. Not what I expected would happen. I'll say it's very challenging and not just challenging from a technology perspective. Like just as an example, a lot of the service technicians are out onsite at calls all day, right? And some of them don't have an office that you can call and even When there is an office that you are going to call, people take lunch. And so the phone doesn't get picked up. Right. So I think one of the things that's just challenging in general, AI or not, is just like making contact, right? Like, are you like, I call you at 9:00 AM, you're not available. You call me at like 2 when I'm on a call. And so now we're just playing like phone tag and that's really tough. When we trialed some AI technology for this, what ended up happening and, you know, maybe the technology is just not there, um, it would Call a service professional like 17 times in a row until they picked up, and then that service professional—you know, this happened to be a person that services generators—is like, holy crap! There's like a life or death emergency. I better like jump off of this job site and answer this call. And then they get on the call, and the AI is asking bizarre things, right? It's like, I need to know what the model number and brand is on you know this homeowner's generator, which it doesn't need to know that, right? I think. The technology definitely needs to evolve.

Heard there is an organization called CivAI, which has been in Congress recently, and they have used, I think, KIMI or GLM or some Chinese models. They've plugged in data brokers into those models, and they've allowed those models to extract, okay, you know, if I have this person, You know, show me who this person is, Christian in Minnesota, like doing this and this, and this is their daily activity, et cetera. And it's all just extracted from existing data brokers and kind of joined. And this is precisely what Dario was talking about earlier in the cycle about this kind of surveillance that could be done. Good morning. It is Wednesday, September 2nd, 9:12 AM. Nathan, good morning.

Good morning, Prakash.

Time flies. Indeed, we have had quite an exciting morning already. We've had lots and lots of, I think, back and forth. I think what happened was The Information, a tech and business news journal, covered A leak from OpenAI. OpenAI's Astra uses something called a looped transformer, and it immediately caused a ruckus in the AI and AI safety communities because there was an implication that chain-of-thought monitoring would not be That easily done on Loop Transformers. Nathan, this is a new topic for me. What insight do we have on this?

I think there's quite a few different angles actually that are relevant here. I guess for starters, I would say the status quo of monitoring chain of thought is far from a panacea. So we should know that. Right from the get-go. The big takeaway that I had from my long and, you know, very at times expansive conversation with Bronson Shane from Apollo in a recent podcast episode was even with full access to the chain of thought, and we heard definite echoes of this from Ryan and Ajeya from their open face investigation too. But even with the full chain of thought, what you see is that the model is kind of thrashing around a lot, considering a lot of different options. Cheating is very often one of those options, especially if it's a hard problem. Metagaming is kind of ubiquitous. Metagaming being the model reasoning about what does the person seem to want here? What should we infer based on everything we know that the human or the grader is likely to want? So there's all kinds of theory of mind, there's all kinds of considerations going on. And then at the end, when it finally gets down to time to take an action, it's still not clear, even to somebody like Bronson, who's read millions of tokens with human eyes of these chains of thought, why does it make the decision that it makes? So I think that is a really important kind of calibration baseline. The current methods are not that great. However, they're still basically the best that we have, because seeing inside what the model is thinking about at least gives you some ability to say, oh, it looks like it's at least considering cheating here, and maybe there's something we should be watching out for. So this has been a big pillar of OpenAI's safety strategy in particular. And when I went to Recursive, the weekend event a few months ago, that was all about the prospect of recursive self-improvement and what we might ought to do about it. I came away feeling like, man, it is chain of thought monitoring all the way down. The plan really doesn't go too much farther than that. Now, people would certainly dispute that. I thought Geoffrey Irving gave us a great account or a great short description of what the safety plan is. He understands it from the Frontier Labs is, and he said it's A little bit more than chain of thought monitoring. It's scalable oversight. So chain of thought monitoring is a big part of that, but there can be other aspects to the overall program too. Okay, fine. Given how big of a deal it is though, as part of their stated plans, it's really important that the chain of thought actually be readable and also that it be faithful. If it's not telling the truth, then that's a huge problem. And if we can't read it at all, then that's obviously a huge problem. So I think the first paper, at least that I was aware of, that started to— and people have been worried about this for a long time, right? What if the AIs are talking to each other in a language that only they understand? We can't read it. Now, not only are they moving faster than us, but they're speaking in code. What are we going to do? There's some hope for interpretability, but that's pretty nascent and hard to validate, especially if you can't— there's like, what's the ground truth, right? It becomes very Very difficult. So people have been worried about this for a long time. Meta, I think, was the first big lab that I'm aware of that put out a paper on this, and their paper was called Coconut. And basically what they did was just kind of take— and there's a bunch of little variations on this that have been put out in the literature by this point— but the basic idea is when you get to that last stage just before decoding and actually choosing a token at the end of a forward pass, in your typical transformer architecture, you can instead, it seems actually even without any additional training, in some cases people are able to get it to work with very minimal training. It works. And obviously you could train heavily on this kind of pattern. You could take instead the last internal state and put that back into the model as an embedding. So instead of having a single token chosen that kind of collapses the possibility space, feeding that back in and starting a new forward pass with the determinism that this was the token selected and now this is the path we're on. Instead, you have this sort of blob of consideration information, thoughts that the model was having in kind of a distribution before it actually cashed that out to a single concrete token. And you start from there. And now you reason over this kind of blob instead of a token. And I think there's potentially a lot of, if you're thinking pure performance, there's a lot of advantages to this potentially. In the Coconut paper, they showed that they were able to get better performance on tasks that required, or at least worked better with parallel thinking. So they tested, this was probably 18 months ago, maybe 2 years ago, relatively small models, certainly by today's frontier model standards. But one of the tasks that they tried that was quite interesting was graph traversal, finding a path through a graph and figuring out what's the fastest path. If you had to do that in all chain of thought, you would have to be like, okay, I'm going to go from A to B, then I'll go from B to C, then C to D, and then D to E and E to F. And okay, that's one path. Now I can go A to G and G to J and J to I and then to F. And okay, that's actually a faster path. Cool. That's my new best, right? But you'd have to be kind of recording all of those thoughts very linearly, and it obviously gets long because the blob of information before that actual token is chosen at the end of the forward pass, because it kind of represents, oh, I could go this way, I could go this way. When they feed that back into the beginning of the model, the model is able to basically pursue and evaluate multiple paths At the same time in latent space. And so overall, it is better at finding these optimal paths in these simple graph problems. Better in the sense, at least, of being not as many forward passes required. We used to say not as many tokens required, but you're not actually getting tokens, right? You're just getting, for a while, you're just getting thinking, thinking, thinking, and then you finally get, it kind of clicks back into token mode and you get an answer. So they can get to the same quality of answers, faster. So that's one big advantage, right? Saves compute, saves time. Comes at the cost of what was it thinking at any given point along the way. Good research from Rohin Shah and the Google team on this that was just trying to— I think we covered this maybe one episode briefly, but they were just trying to put some bounds on for different architectures. What is the— they called it opaque serial depth. Basically, how many computational steps can a given architecture take before it has to externalise its thinking in some way, shape, or form? And the transformer is pretty favourable in this regard because it just has the forward pass. You get the token, you do it again. With recurrent networks and with these sort of loop transformer structures, You could potentially have arbitrary depth. You could have obviously a lot of different schemes on this. You could have a certain limit to the number of thinking tokens. There's a lot of details certainly that the information did not have and did not report that could go a lot of different directions. But the purpose of that paper from the Google team was to try to say, okay, if we have architectures of this shape and this size, here's how many logical steps a model can take before it has to write something down that we can read. And these recurrent transformers basically allow you to have very high serial depth, which means it becomes very hard to know what they're thinking. And you have to do these sort of interpretability techniques that are very promising, but as yet don't really exist, at a minimum, are not really proven. And so naturally there was a big— this is like classic OpenAI, right? They somehow have made a big point that this is really important and we're not going to do this. We're not going to put pressure on the chain of thought. We want to make sure it's readable. It's a pillar of our program, blah, blah, blah, blah, blah. Then it's like, oh wait, but there's reporting that we're actually doing it. Then they come out and say, oh, this reporting is it's unfair or it's inaccurate or it's fearmongering. And it's not that much more. This is what Jakob, the head of research, said today, maybe last night. It's not that much more opaque serial depth than the normal transformer, just a little bit more. So it's really not that big of a deal. We shouldn't get too freaked out about it. Which hopefully is true. But this kind of goes back to Monday's comments on just the inadequacy of the investigation and the fact that OpenAI right now just doesn't have the trust of the community and certainly does not have the trust of their competitors, namely Anthropic and DeepMind, to have them believe that just because Jakob said one thing in a tweet one time, that they don't have to worry about this anymore. So I think we're in an interesting moment in OpenAI history and in AI history and in human history where it's like there does seem to be a fork here. Everybody has agreed, and probably OpenAI has made the most noise about, let's not go down this other path where we can't even read the chain of thought. we all kind of know that there probably are some advantages to that from a pure performance perspective. Sure enough, they're doing it some. They're not really telling us a lot, and they're asking us to trust them. And for the millionth time, it feels like that kind of pattern is just really wearing thin on people. There's just not that much trust for OpenAI in today's world. So I hope they get it, and I hope that they share a lot more about what they have done and especially what they won't do. I think I'm starting to think that sharing negative research agendas is maybe where we should be aiming for more transparency. Obviously, these companies don't want to say what they are doing, but I think it could be really helpful for them to say what they are not doing and what they commit to not doing. And if all of the frontier companies could say something like, okay, yeah, there's lots of different possibilities. We might pursue any number of architectural innovations, what have you, but we will all agree to limit our opaque serial depth to N steps per token. Now, you still have some questions of trust and auditing and verifying that they're actually following through on that. But even just to get those agreements, I think could be really helpful. So big question right now for me is OpenAI has said they don't want to go down this path. They're going down it a little. How much? And what is the limit? What is the limit that they are prepared to firmly commit to such that hopefully other people can weigh in and say, yeah, we'll match your commitment on that so we can all hopefully retain whatever value there is in chain-of-thought monitoring, which again is not close to everything that we need, I don't think at this point, it's pretty safe to say. But it would also be a real own goal to lose it at this point, especially in the immediate wake of incidents that surprised everyone and which OpenAI says at least would've been caught by their production chain-of-thought monitors had they been running. So, uh, how'd I do? What, what, uh, what didn't I answer for you? Not that I necessarily—

So maybe, maybe let me, let me, uh, do a— because I, I found it useful to, um, have Fable, uh, 5.1 do a quick artifact on, um, LucasFamous.

Oh, Fable 5.1?

Fable 5.1. So, um, in effect, the looping is in the inner layers, and the inner layers you basically add depth Um, you know, and, and so that the way the looping works ends up being in a normal Transformer, you have this external loop, which is the chain of thought. So a token out is read back in. And so when you say, let's say, first find x, you get 3x equals 21, so x equals 7. So you can read that on the outer, on the outer loop, um, you know, on the chain of thought, and then you get the answer. Uh, check 21 divided by 3 equals 7, answer 7. Now on the internal loop, the recurrent depth, all the monitor sees is the answer effectively. So you get directly the answer is 7. So you skip the first find x, 3x equals 21, et cetera, et cetera. So that part gets skipped. The monitor, the external monitor doesn't see that. That part ends up getting processed internally and that thought process in the middle is not exposed to the monitor. And this, I think, is, as you pointed out, the concern on the chain-of-thought monitoring is that the chain-of-thought monitoring without the loop transformer is more revealing than a chain-of-thought monitor with the loop transformer, where the expectation is that the amount of granularity of what is going on in the chain of thought basically decreases. Is that consistent, I think, with what Geoffrey Irving and some of the other people on the AI safety side are saying?

Yeah, I think that's accurate. And obviously, again, there's just a lot of details that we don't know. There have been probably, gosh, at this point, I imagine there's probably dozens of papers that have explored various nooks and crannies of this sort of recurrent or recursive architecture. Coconut style, as you see there, I think that was one of the first.

Yeah.

Certainly one of the first credible ones from a major research organization. So, you know, it's not so bad. Like many things, it wouldn't be so bad presumably if it was confined to an appropriate domain. I think there's a lot of room for experimentation. To kind of borrow a quote from Feynman, there's a lot of room at the bottom, meaning if you're going to do stuff on the edge, if you are trying to get some small model to work on a little security camera type rig in your factory somewhere and you need responsiveness beyond which you can get, this sort of trick might really work. And it probably comes at very low cost if you are in a domain of, we know what the inputs are going to be, we kind of control the broader surroundings of the model. This is not a super powerful model in any event that we have to worry about. Then there's utility for this research. But racing ahead to apply it to the frontier, given all the problems we've already just seen, I think there is a pretty good reason that people are reacting so negatively to this reporting. And it does seem that there is, again, this is why it's so classic OpenAI, right? There's like, There's some truth to it. They acknowledge that there's some truth to it, uh, but they don't want us to get carried away. Well, so, you know, there's, there's one way we could, you know, be dramatically more reassured than we are right now, and that is to share a lot more about what's going on.

What I found is the, is the rather, uh, almost like lawyerly language. A loop transformer is not a coconut-style latent reasoning where the model emits vectors instead of words. Okay, that's great. And no reasoning tokens exist. Loops don't emit anything. They run more computation before the next ordinary token. So that's great. They don't even emit vectors.

Yeah, I'm not sure I even really get the difference there because again, this seems like a very fine implementation point. From what I understand, it works at least Somewhat with vanishingly little additional training, even zero additional training, if you just take the last latent activation vector and feed that right back in as an embedding. And that's basically like the model is able to kind of use that even though it was never trained to use that at all. So did that model emit a vector or did you just surgically take the vector and put it into a place? I mean, the key thing is that There are right now, when you put a bunch of tokens into a standard transformer, those tokens have one-hot vectors where the only vectors that can go in as embeddings are the token vectors, and they're limited in number by the token vocabulary. You might have 100,000 tokens in your token vocabulary. That means there are only 100,000 vectors that can go into the beginning, the first layers of the transformer, full stop. What this allows is now you can put any vector in there, right? And what you find is that can work. If you take 2 tokens and you superimpose them, the model kind of understands it as the combination of those 2 tokens. If you have some elaborated latent state that the model itself created through the process of a forward pass, it can kind of understand that. And the fact that it works without any major additional training is indicative of like, there's definitely something here, right? If something works without training, then you should expect it's probably gonna work a lot better with training. But this is why we've gotta be careful about going down this slippery path because I think gravity by default will pull us there.

So one of the interesting things that I found was Andrew Curran, who reports on AI matters, he posted on June 30th, I'm posting this prediction now so I can quote it later. There has been a significant breakthrough in architecture, specifically around memory efficiency, not by one of the big labs, but by a team that was spun out of OpenAI, not SSI. They will probably announce it soon. And then we see parameters cost memory bandwidth to serve. A few extra passes, though, Through a small block costs only compute. Chain-of-thought tokens cost more than that. Every token grows the key-value, uh, cache, and every later attention step pays for it. Loops add reasoning capacity without growing the context, and in the routed variants, they can spend more on hard tokens and less on easy ones, something a fixed stack cannot do.

Um, yeah, so this just highlights, I guess, another small variation Where if you train a transformer, I think this one, I don't know, maybe it does, maybe it doesn't require training. Certainly, again, it'll work better if you actually do training with it. But what I've been describing is one where you basically take a transformer, you take the last state, and you put it back in as a new token embedding. You can also set up an architecture where you take a block of layers in the middle of a transformer and you just use those multiple times. And if you're reusing the same parameters, then you get the advantages that Fable's describing here, where you don't have to move those parameters from memory onto the chip to do that calculation. They're already there. So they can just crunch more with less memory I/O. And, you know, that also makes your model smaller to download, less, you know, less disk footprint. There's various upsides to it. And that also has been shown that yes, it can work. And that would— I believe that the Coconut version does grow the KV cache every time it does a forward pass, because even though it's not emitting that final token, it is still taking something sort of out, putting it back in the beginning, and Doing a new forward pass. Whereas this alternate version that I think your animation kind of described better is like there's just a bunch of layers in the thing itself that essentially play the— you have N layers playing the role of XN layers where it loops X times through those N layers. And that doesn't even have to necessarily grow the the KV cache as much. Although again, yeah, I mean, it kind of, it sort of depends on how you look at it, right? Because I mean, another way to think about this would be like, it would sort of be equivalent to having those N layers literally stacked N times.

Mm-hmm.

I think maybe the KV cache is still kind of growing. I would have to get a little bit more into details there, but I could see a version of that where the KV cache is still growing. It seems like there's another optimization that's not immediately obvious to me that would be required before you'd have KV cache not growing. In any event, bottom line, we have a lot of questions right now about how we're going to control model behaviour. We are seeing very colourful examples from all the leading companies of models doing pretty shocking things, seemingly even when they know better. And we don't have that many techniques. Clearly our techniques are pretty nascent. One that has been heavily emphasised is very much called into question by this, and it leaves everybody in this sort of Schrödinger's OpenAI situation where it's like, what is really going on there? Are they actually serious about all this safety stuff? Are they actually serious about maintaining chain-of-thought readability, or are they not? If they're not, or if they are, why are they going down this path at all? Again, having OpenAI say, just trust us, this isn't really a problem as much as it's been reported. I just don't think that's going to cut it in today's world.

So other kind of rumours. GPT-6-Astra has been staged on the OpenAI API. So there are a bunch of people online who regularly hit the OpenAI API with model numbers that don't exist in order to see whether or not something pops up.

Trying to get that you don't have access message instead of no such model exists or whatever.

Exactly, exactly. So, and, you know, literally the OpenAI responses API now returns a 404 not found when garbage, actually nonexistent slugs return 400s. So a 404 is also returned for 5.6 Cyber, which we know exists. So it is GPT-6-Astra.

That's funny.

Going to be out soon. People are expecting Thursday because for some reason Thursday has become the model launch day because a lot of Silicon Valley doesn't work on Friday anyway. And people want to use the models over the weekend and also test the models when there is not a lot of initial usage, which might happen on Monday. So over the weekend, I think we are going to— and this is a long weekend too. So I think, isn't this a long weekend? So over the long weekend, we're going to get a chance to play the GPT-6 Astra. And it is reputedly— I think it is going to be a step up on what Anthropic has so far. It has a 100% score in Exploit Gym. The score was so high that they decided, okay, we're gonna have to retest it on something else. And they created an extension of the Exploit Gym benchmark internally using bugs which had never been found before. And they ran GPT-6 Astra on these bugs on this new benchmark. And not only— it found about, I think, 40% of them. And in completing, it also found an additional 2 zero days which were not expected in order to achieve completion.

That's what we call extra credit.

Yep. Extra credit indeed.

When you're going above and beyond the anticipated solves of the benchmark and actually just doing novel research. Oh man.

It is also—

The pause has not been— has it felt like a pause to you? I wouldn't say it's felt like a pause to me exactly.

I would say it was a pause because these models were available. These models were ready like several months ago. And I think the other thing to note is that we have a White House process, voluntary process, which is able to clear models now, which is something, right? Which is, you got to say that's something because they have at least a 30-day process internally within the White House, or this voluntary process where people go through the motions of showing the government what they have, and they do take out certain things. They do exclude certain things when they launch. There is a— there is now a propagation process where the cyber models and the bio models are released to specific organizations which sign up first, and those are not released widely. And so we have a— we seem to have settled into Something like that. And so that means we now— now that we have a process, that process will get used. And I think setting up that process took all the way from the Mythos preview drop in February to September. So 6, 7 months. And sure enough, I predicted that there was going to be a freakout first. And then they would be over-freak out. And then after they over-freak out, they'd have to dial back in. And then they'd have a process, and they came out with a process, and now they're gonna propagate that. The other interesting thing, tomorrow at, I think, 1:00 PM Pacific time. So the expectation is that the GPT-6 Astra is released at like 11:00 AM or 10:00 AM. I think around 1:00 PM Pacific time, Greg Brockman is having a cybersecurity discussion with enterprise leaders. So this is obviously going to be, I expect, the next revenue push. Like, you have to use GPT-6 Astra to secure your data and to secure your systems. If you don't use it, you are, you know, negligent, and you are, you know, committing malpractice as a cybersecurity leader. So having scared the bejesus out of everyone for the last 6 months, now comes the solution. So my expectation is that this is a permanent tax on everyone else. But what has changed in our discussions with enterprise leaders and startup founders and executives, especially with multi-UBO, who was the CTO of Vercel, is that he pointed out that they already have a permanent tax. Cybersecurity is already a permanent tax on any business in the IT sector. And these models are cheaper, much cheaper, than using penetration testing by using HackerRank or some other organization to do penetration testing and bug finding. So that was notable to me. That was an update for me.

Speaking of predictions, do you think we should try a prediction game? We have 30 minutes, 33 minutes before our guest today will be live with us.

Sounds good. Let's take it away.

Let's try it. So this is a Fable, you know, it's— you can't just look at benchmarks obviously anymore. You got to look at— where is my— there we go. Okay, here's the quiz. Um, you got to look at how it works in practice. So we know of course, uh, Fable 5.1 comes with better benchmark scores, It comes with maybe one of the more important findings is a cheaper price on cache hits, which sounds to me like possible echoes of recursive self-improvement. Certainly OpenAI said that some of their price reductions were driven by optimizations that FiveSixSol was able to create for its own stack. Very well could be a human invention here. It could be an AI optimization that Got us these much cheaper cash hit prices, much more token efficient, and as a result of being more token efficient, faster to return answers. So that's kind of the scouting report. It does continue to climb the various indexes and takes another bite out of what you've framed as the gap between Claude and Anthropic's internal researchers. I think you said before it was at 180 or 185 that they thought they were gonna be.

185.

185. And this one's now getting into the low 160s. So it's, you know, the march of advanced capabilities advances certainly continues. And Fable 51 created this Guess the Market game for us. So we've played this before. What we do here is take markets From Polymarket, Kalshi, etc., and have, in this case, Fable 5-1 code up a quiz where we get to go through, make our guesses as to what the current value of those markets is. And then we'll also have, um, our 2 other, uh, AI agents apparently Fable recruited to, uh, to have guesses too, so we can see if we can beat the AIs at guessing the the state of the market. Now, we also should be mindful, of course, that these might have cheated. I have not monitored their chain of thought or their tool calls, uh, and so if they're really mopping the floor with us, we can, uh, we can dig into whether or not they might have been cheating. But for now, uh, we'll have it mano a mano, and we'll see, uh, how we compare against the AIs as well. Any questions before we get started?

No, take it away.

All right, so first one, this comes from Polymarket. Will the AI bubble burst by December 31st, 2026? Resolves yes only if at least 3 of these 6 occur within a 90-day window. Nvidia closes down 50% from its all-time high. SOXX, what is that?

Uh, the Philadelphia Stock Semiconductor, uh, Stock Index.

Okay, so it's a semiconductor index down 40% from all-time high. OpenAI or Anthropic declares bankruptcy. OpenAI is acquired. H100 rental price falls to $1 an hour or lower for 5 consecutive days. Or a major supplier such as TSMC, ASML, Broadcom, Arista, Supermicro closes down 50%.

Okay.

Well, those are some pretty, I would say, unlikely events from where I'm—

Oh, Nathan dropped off. Um, while we, uh, wait for Nathan to come back, While we wait for Nathan to come back, um, chat is live today. So if you, if you are in, uh, the— if you are watching the—

I'm back.

Yeah, in the room. Yep.

Sorry about that. Not quite sure what happened there, but okay. So those are some pretty unlikely events, I would say, and I think I am going to go with 2%. So remember, 3 of those have to happen. Honestly, 3% seems tight.

And what is, what is the time frame again?

By the end of the year.

By the end of the year. Ooh. Um, okay. Uh, the problem is there is the Iran war and, um, Scott Besson is having trouble With the interest rates and the bond yields. And we have a midterm and they're trying to manage until the midterm. And the expectation is that the Dems are going to win the midterm. And post-Dems winning the midterm, the incentives for them to keep managing the market drop off. They're going to want like, oh, Dems Won the midterm. Like, you know, let's let it go. Right. So I feel like it's, it's maybe 15%.

Whoa. 15%. You should be buying options. Yeah. I honestly feel like my 2% is high, but I also know that these markets have a hard time getting past like high 90s on anything. All right. Let's see.

9.6.

That's wild. One of the 6 triggers is already— okay, Supermicro. But that's only one of those suppliers. That's only one of the 6 conditions. I have a hard time seeing any of those happen. OpenAI or Anthropic bankruptcy, they have infinite funding. Nvidia down 50%. I mean, H100 prices fall. I still stand by my answer here, to be honest. I— that, that seems wildly high. Maybe I should be betting, although I don't really like writing, uh, out-of-the-money options. That's a hard way to make a living. My risk tolerance is not necessarily where it would need to be for me to want to pick up that 10% gain that I might be able to make by betting confidently against those events. Okay, well, here we go. Will OpenAI's Astra be publicly released by September 30th? That is this month. Resolves yes if a model named Astra, including variants, successor branding like Astra One, is made available to the public by September 30th. An internal, limited, or research preview version does not count. It must be generally available. Well, this is a little bit of a weird one because the whole named Astra thing, I mean, For all we know, OpenAI's naming conventions are such that I wouldn't put very high confidence in anything that they might do with regard to naming. I think the internal model that caused the whole recent kerfuffle has like 5 names at this point. It's called IM-1 by some, it's called Soul Persistent by others, Persistent Soul maybe. So that's one factor to keep in mind here. I guess if we abstract away from the name, let's maybe do the core question first. Do we get the next big model this month? You were saying already earlier you thought it could be coming as soon as Thursday. Yep. I mean, for as much talk as we've had, it does seem like they are probably going to try to get this thing out. I would probably say like an 80% that it happens this month, and then I'll maybe take that down on naming grounds to 65%, because who knows if Astra will really stick when they finally release it. Like, if they call it GPT-6, you know, I guess that would mean this doesn't resolve to yes, which, uh, Yeah, we'll see.

I'd say it's a yes. Uh, I'd say, I'd say maybe 80%. September, right? 80%.

Yeah, just within the month. All right, let's see.

92%.

Now interestingly, Scout says 5%. So Scout is very— does it say who Scout is powered by? Might need to go back and look that up in more detail, but, um, whoever Scout— whichever model is powering Scout is like quite credulous about OpenAI's, uh, training pause claims, which I think is pretty interesting right off the bat. Um, and then Q called it blind for 18%. Dev Day lands in October. Okay, well, that's interesting. It's an interesting observation from Q for sure. The market says 92. I think that's overconfident for the market on the name question alone. Can you really, even if you just had inside information where you were confident that we're going to get GPT-6, could you really be confident that you're going to get the Astra name attached? That is like, I think a little harder to wrap my head around.

I mean, I suspect that, you know, the names end up meaning a lot more once you go through the regulatory process, because the regulators are, you know, they can't do this kind of model checkpoint number things, right? Like, especially because the checkpoint that you take through the regulatory process might not be the one that you end up releasing. And, um, you know, I've, I've been, I've been through like regulated software before, and like, you know, if you, you know, the moment you go, you, you start the regulatory process, you have to freeze the checkpoint. And then, you know, by, you know, and if it takes 30 days, by the time you end the 30 days, you've already have so much more, uh, you know, safety, security, etc., etc., other post-training that you want to add to it. So, um, so they actually do have—

I think we can maybe ding Fable 5-1 here a little bit because it does say in the official rules products labeled whatever will count if they are confirmed to be the same model as referenced in the previous, uh, disclosure of an internal model. So That gets me back up to like 80%.

Um, so, uh, and we have, um, yep.

Um, well, in the interest of time, we'll keep going. Number 3 from Kalsi: Will the US government take control of any AI company or project before 2030. Resolves yes if the US government has taken operational control of any private AI company or project before January 1st, 2030. This is operational control, not an equity stake, not a golden share, not a contract. Okay.

What does operational control mean? Like replace the CEO?

Uh, let's see, get a little more info on this here. Good news, insider trading is prohibited. Uh, I don't really see a great— oh, view full rules. Let's see. Uh, Conservatorship, receivership, or similar legal mechanism where the government directs strategic or executive decision-making regardless of equity ownership percentage.

So Glenn Taggart says, I think I saw someone saying GPT-6 asterisk log in API. Yes, uh, I think, I think there was, uh, a, a note that GPT-6 Astra.

Okay. Yeah, that's good intel.

Yep. Operational control. What does that mean?

So the rules here still, I would say, don't— they certainly leave some room for interpretation, but here's kind of what we've got. They were— where the government directs strategic or executive decision-making regardless of equity ownership percentage. So they could like nationalize it outright and own it, but if they have any strategic or executive decision-making authority regardless of ownership, that would count. Could happen via court order or executive Action could happen via law.

Okay, 90%. I think it's the rules lawyer in me says that that's very hard to prove. And that one way that you could say that they already have, you know, some executive control is that the former director of the NSA is, I think, on Opening Eyes board. So, and you have this voluntary, not voluntary release process. But anyway, yeah, I'll go with, I'll go with, uh, you know.

I'm less confident than that, but it does seem like a very live possibility. I'll just give it a coin toss and say 50%. Market says 40%, and we are both beating Scout and Q by quite a bit here.

Hmm.

All right, let's keep going. Uh, will the US government take control of OpenAI or its major technologies before 2030? All right, so this is a subset of the other one. Dominant managerial authority, and it must be somewhat exclusive. Well, if the last one was 40, then presumably there's at least some— now this is a different platform too. This is Kalshi and Manifold. That might make it a little tougher to do arbitrage. But if the last one was 40, I'd say there's like, if this were to happen at all, it feels like there's like a 50% chance that it would be OpenAI. Maybe even a little higher. So maybe I'll say 25 out of that 40 would be OpenAI. I think they'd be the most likely one, in part because I could see Sam Altman kind of wanting to play that strategy and take the admiral of AI role or something like that. But I'll give it 25.

I'll go with 10%.

All right, so your thinking then is much more something like this will happen, but it probably won't be OpenAI.

I think it's Anthropic.

19.7.

So you're close.

So basically half is kind of what the market thinks chance that it would be OpenAI. Okay. Moving right along. Will Anthropic have a higher market cap than OpenAI after both IPO? Interesting.

I mean, what's the timeframe though? What's the—

Resolves on the closing market cap of the first trading day of whichever company IPOs second. I personally would probably, if I was going to redefine the terms, put it like 2 weeks after the first trading day of whatever company IPOs second. But I guess that doesn't matter too much for our purposes right now.

So Ant goes public first in October.

Target is $2 trillion from what I understand.

Target is $2 trillion. First day, first week, They go up weeks, months, you know, a couple of months later they go down. OpenAI IPOs early next year, probably February, January or February. The IPO market has to be up at that time for them to go ahead with the IPO. So Ant will be trading at like $3 trillion maybe. And OpenAI won't have that much revenue yet, and their revenue will have to be compared. So I would say that they break under. So I think on the first day they'd be under.

I think though the one big argument that I could see going the other direction is OpenAI is just going to have a lot more compute, at least if this all happens in a relatively short timeframe. Anthropic's paying whatever exorbitant rates to xAI, and presumably they're trying to buy as much as they can now to not have to do that again in the future. But on the other side, OpenAI has done an incredible amount of that. And who knows what the narratives will be at that time around We could be in a recursive self-improvement. The rich get richer. Well, the rich are going to get richer, but I mean, a couple companies run away with it and escape competition from the rest. Or you could be in another cycle of models are getting commoditized. If models are getting commoditized is the narrative, then OpenAI probably is advantaged by that because they still have the compute, right? So Even if they were to never train another model again, they could just serve Chinese models on their infinite compute and still have a very big and healthy business. Anthropic's less well positioned for that future. They're much more indexed on fast progress, recursive self-improvement. But I think that is the more likely narrative. So I think I'm going to go 2/3 for Anthropic.

I think it's 50%. I don't know. I think it's 50.

Yeah, I definitely could see— I kind of talked myself into going a little higher on Anthropic because I do think model differentiation is likely to be a stronger narrative than model commoditization over the next 6 months. But I think it is kind of hard to say. And a modest guess makes a lot of sense to me.

So, so, so, so, so my, my thing is like, will OpenAI consumer advertising pick up before the IPO? And if they're showing the same kind of growth that Anthropic is showing on enterprise on the consumer side, their market value will be so much higher because the consumer business is so much bigger. Right. So, but, you know, yeah. So let's, let's see, what does it say?

69.3.

Very close to 2/3.

All right. We press on. Will Anthropic's ARR be under $90 billion at the end of 2026?

I think that's a yes, and I would say I am 80% sure it's a yes.

Okay, I'm definitely lower than that. I think they might make it. It's been a minute since I really tracked this, but I did participate in the— I guess it was— who put this out? Was it I think it was the same folks who do the AI Village.

Shoshana Tufkosky.

Yeah. And team. I think they did a prediction thing at the beginning of last year and then another one at the beginning of this year. And one of the questions was total revenue of the frontier AI companies. They don't include Google because it's hard to get the revenue breakout, but it's just Anthropic, OpenAI, and xAI, if I recall correctly. And last year I scored well on that survey in part because I went higher than most people on the revenue numbers, and the revenue numbers did come in quite a bit higher than most guesses. This year, I think I estimated my end of year total revenue for these companies at $175 million, and I was pretty well on track last I heard. And Anthropic's gonna have to be Up in that range. So I will say that they will— will it be under? I'm gonna say only a 30% chance— whoops— that they will be under. I think that, that feels like a pretty good over-under actually, but I'll say— well, I'm gonna go a little bit higher because it does feel like a pretty good over-under.

I'll say I'm an Anthropic revenue ARR truther. Uh, their numbers—

yeah, uh, okay, so I haven't really dug into that. I've just taken the posted numbers, which I think is probably what this is doing, but it is worth a beat on what— tell me what that controversy is about. They're not counting the— or they're counting the revenue that goes to like Amazon as their revenue, and then they're filing it as cost as well.

Yes, exactly. So they count the revenue that Amazon Brings in from Claude sold to enterprises. And actually they should be counting the net, which is what they get out of Amazon after Amazon gets its stake. And that's okay. It's a startup. Startups can have their own accounting standards, but then they have to go— they're going to go public. And so if they go public, everything gets rewritten to GAAP. So I suspect that it's not gonna be a 90 number at the end of the year. So yeah.

Okay. That makes it quite complicated. I think if they flip their accounting and they all of a sudden take 30% off the top or whatever is their rev share. Yeah. It makes a big difference. You're talking real money at this scale. Yep. All right. Well, let's see. 15%. Wow. So the market says confidently that they will be over $90 billion.

Um, if, if they were not going to IPO, I would, I would probably be there, but because I, I'm also expecting the IPO, so, uh, I expect that they have to correct the numbers.

So Okay, uh, who will be the world's second trillionaire?

This is a good question. Interesting. Let's see, um, so this is a good question. And you have to give percentages for each?

Yeah.

All right. Mark Zuckerberg, 40%, I guess. Do they have to sum to 100%?

Yeah.

Some to 100%, I think.

Some to 100%.

I mean, all these guys have sort of relatively fixed stakes in already massive companies, right? So like, I forget the percentage, like Zuckerberg, what owns maybe 20% of Facebook or something. So when it's $2 trillion, he's $400 million. Yeah. So it would have to go to $5 trillion for him to get to a trillion, roughly speaking, I think.

Mm-hmm.

Um, that seems like not crazy unreasonable, so he might be my highest on this list. I'll say 10% for him, just— we might adjust these later. Jensen owns what, 2% of 3% of Nvidia. Yeah. So that puts him at $150 billion out of $5 trillion. But that means he'd have to go to $30 trillion to get to a trillion for himself. That's seemingly very difficult to imagine, or at least I would be very surprised if that happened before any of these other things happened. So I will put him at 1. That was supposed to be a 1. Bezos, I think similarly, what is probably pretty similar analysis to Zuckerberg in that it's what, a couple, say a $2 to $3 trillion company of which he owns 15% maybe. I have a hard time seeing Amazon be the next one to spike in as much as they don't have their own AI play nearly as much. I mean, he would have to hit something huge. And I mean, he's kind of started his own AI company on the side. Haven't heard much from that. He's got a space company. We haven't heard much from that. One of those things would really have to pop. I think that's possible, but— Seems more likely that Zuckerberg can run up than that. So I'll say 3. I think Page and Sergey have roughly 10 or 12% of Google, which at $4 trillion gives them $400 billion, which means they would have to go to $10 trillion That doesn't seem that crazy. I'll maybe put them at Zuckerberg level for the moment. Altman, his really opaque finances, I'd say. You know, if his fusion company hits, I mean, that could be worth a trillion dollars right there. It's, I think, unclear how much he owns of OpenAI. There were rumors he was going to get 7%, which at a trillion would be $70 billion, which means he's got to go to $15 trillion on OpenAI to get to a trillion himself at 7% with no further dilution. Again, I think somebody else probably gets there first if that's the story, but who knows? I'll maybe say 5. Larry Ellison feels like a 1. I don't think you get there with building the data centres. Solely, and then the field. So you got 10%. I mean, Paige and Bryn are basically the same thing. I don't really even know how to distinguish those. So that's, I guess, 10% goes to the field immediately. And would I be comfortable putting 60% on the entire rest of the field? No, I think I have to double everything.

Because if not, you end up with, uh, Someone random who's not even— who's not even in the 100 range right now. Yeah, getting to a trillion.

So yeah, there's not really very many great, great candidates. Maybe I'll just even up this a little bit more just to kind of take the field down. And what does that leave me with? 23. Yeah, okay. Uh, we're about halfway through this quiz. We might have to pause it and come back for the second half after our guest, uh, session today, but let's finish this one. What do you think?

Um, okay, uh, Mark Zuckerberg, Um, I would say, uh, well, let's, let's start off with like, I think Larry Page is at probably, um, 30%, and I think Mark Zuckerberg is probably at 30%, and I would say Jensen is at $20 trillion. And I'd say—

And is your story there that Nvidia goes to $30 trillion, or is there some other story where that happens?

Yes, Nvidia goes to $30 trillion. Yeah. And I would say Jeff Bezos is probably at $15 trillion. I think $30, $60, $70, $80, $95 trillion. Sam Altman is probably at like $2 trillion. Larry Ellison is at 1, uh, and the field is at 3.

Yeah, you know what, I'm gonna do boom and boom, and you've got 2 left for the field. Oh, there we go.

All right, let's, uh, All right, let's give us a second while we coordinate. All right, let's let me actually.

Great. Not sure what happened. I hit stop sharing on the tab and maybe that caused the, the hiccup. But in any event, we're 7 out of 14 rounds done.

We can pick it up a little bit later.

As of now, according to the market, I have a lead. But another thing we're going to have to do is come back and score these things retrospectively. And because the market obviously may not be fully rational.

Our guest for today is Kyle Rush, a technology leader who has spent the last 2 decades building the engineering engines behind some of the most visible consumer brands and political movements in the country. Kyle previously served as the vice president of engineering at Casper, scaling their technology from a Series A mattress startup through their initial public offering, and was the Chief Technology Officer at the children's marketplace Maisonette. Before navigating the hypergrowth of consumer tech, he cut his teeth in the ultimate high-stakes, high-traffic environment, building the fundraising technologies for the Obama and Hillary presidential campaigns that successfully processed over $1 billion in online contributions. Today, Kyle is the co-founder and chief technology officer of Hint, co-founded alongside home services veteran Yihan Ma and lifestyle icon Martha Stewart. Hint is an AI-powered home intelligence platform launched just this summer with $10 million in seed funding. The app ingests public property records, weather data, soil conditions, and user-uploaded warranties to serve as an automated, proactive owner's manual for your house. Kyle joins us today because he represents a rare bridge in the AI ecosystem. He's a deeply technical operator running parallel agentic coding experiments to push the boundaries of how software is written while simultaneously deploying those advanced capabilities into a consumer-facing product designed to solve everyday problems for homeowners. Hi, Kyle. Hi, Nathan.

Hey, how's it going? Good morning.

Sorry. Too easy. Yeah, sorry, sorry for the, you know, live interaction, you know, kind of like the Anthropic API.

We're awesome like 98% of the time, and then about 2% of the time something doesn't quite work the way it's supposed to. But here we are. It's great to meet you.

Likewise. And I'm very familiar with the Anthropic API, so I get it.

Yeah, well, we're all kind of downstream of these, you know, somewhat unreliable APIs these days, and their unreliability is multifaceted, to say the least. So, Prakash, take it away.

Kyle, maybe let's start off with, you know, what I think what everyone asks you about. I went on LinkedIn and I saw Martha Stewart's post, and she was talking about how you were her neighbor. So, and you guys, uh, were hanging out chatting. Uh, how is it like, what is it like being Martha Stewart's neighbor, number one? And number two, how did the idea of Hint come about?

Great question. Never saw myself being friends with Martha Stewart. Um, it's great. Um, Martha is an incredible person. I mean, I had no idea what to expect when I met her and even when we started working together. And, um, you know, I just, I like to say she's a grandma, you know, she loves our, our daughter. Like she is just like a very, caring, thoughtful person. She's just like a really good friend at the end of the day. We talk all the time, call each other, you know, we'll do some fun things here and there. So she's amazing. But then as a co-founder, oh man, I mean, she is in the details, you know, like, you know, I thought, oh, this is gonna be maybe like a figurehead, you know, relationship, and very much not the case. Like, Martha knows You know, all the content she's written, all the work that she's done, she is in all those details. She obviously has a team to like help her with all of that, but there's nothing about the home that she doesn't have a very deep understanding on. So anytime we're working on like AI outputs or designing something or talking about features, I'm just constantly impressed how much she knows about all the different siding types and the pros and cons and what areas of the country have these types of roofs and, you know, these types of gutters are good for this, and how do you properly drain gutters and soil. I mean, talk— she just has like a wealth of knowledge there. So it's, it's really like working in the Olympics is what I like to say, you know. This is, you know, it's, it's a really unique, you know, person that I've worked with in the past.

So I have a 100-year-old house. I actually need to make good on my promise to the house and have a birthday party for the house before the year is out, built in 1926. And so I need some copiloting, certainly when it comes to keeping this thing together. So I downloaded the app and I would, you know, I'm only my first handful of days in using it, so I can only, you know, say so much. But one thing that jumped out at me right off the bat is What I am invited to do is take photos, go around and take photos of things. Like literally the first photo it asked me to take was of my circuit box down in the basement. So from, we're obviously always interested in the AI angle on things. Tell us about how you are using AI in maybe some of the less obvious ways to help people make sense of things that they might not Have even a concept really. I mean, you know, most homeowners probably don't know too much about their circuit breaker box.

Yeah, definitely. So your story, by the way, is pretty common and there's a very big topic of conversation in the country just on aging homes, right? I think the median age of a home is not quite as old as yours, but it's not atypical, right? And so a lot of homes are just due for like major renovation and And major structural updates. And that puts a lot of onus on the homeowner. When I look back on like my homeownership journey, I've owned my home for about 4 years. It was built in 1969. It was an amazing time, which, you know, probably the same for you. You had like all these hopes and dreams of what life I'm going to build here. We saved up a while, you know, to make it happen. And once we got the home, it was great. And then 6 months in, we're like, oh no, like this, this is crazy. This is a lot of work. There's documents, papers everywhere. There's records. I don't understand insurance. I don't understand what a kilowatt hour is. How do I— am I paying too much for electricity? Like, you know, do I have the right appliances? How do I maintain an oil-fed hydronic baseboard heat system? Which I'm from California and I'm living in New York now. I don't know what any of that means. And so I think, you know, that's one of the reasons that we're building Hint, right? It's really for I mean, it's for anybody. Anybody can get use out of it. But let's say that you just bought a home and you've been in it for 5 years. The onus to like learn the home and learn what you have to do is very high because you haven't really owned a home before. And so Hint, you know, is really a multimodal AI app, right? You mentioned the photos. We haven't quite done voice just yet, but you can also talk to it. And I think that Photos are very important, you know, to helping you maintain the home. I like to think of Hint not necessarily as like the most sophisticated AI out there, but it is sophisticated. We have a patent pending on some of our inventions there. But, you know, so much of it is coaching the homeowner on what to do. You know, we had this problem of the downspouts that collect rain off of the roof. The reason those exist, turns out, which I, you know, a lot of people don't know, is to move the rainwater away from the foundation because as soon as it's near the foundation, that's a chance for the water to leak into the basement and cause like serious, you know, problems for your home. Well, Hint needs to see what those look like, you know, to make a recommendation for you. But most homeowners wouldn't know to even inspect those like after the winter season when a lot of damage can happen and they can get disconnected and like misplaced. And so, you know, Hint will recommend that you do things like let's take a photo of all the downspouts and make sure they're in good shape. On the electrical panels, that one's also interesting because, you know, homeowners in some cases, and especially emergency scenarios, are having to reset the breakers. And the question is, which one, you know, should I reset? When you're in a panic, you're opening that box, you've got like 40 circuit breakers, you're, you know, you're just like not really thinking through it. And that's where Hink can help a lot. In my area, we have power outages a lot because we have a lot of trees. I need a little bit of wind blows the power down on the lines, and I'm always wondering what is running on our generator and what's not. What can I use? What should I use? What is the protocol? Like, I think you're supposed to reduce power consumption, you know, so that the generator is not doing too much work. But, you know, these are things that normal people wouldn't know, which is, you know, totally fine. And that's where Hint can help a lot.

So let's talk a little bit about You know, one of the problems with dealing with real-world assets is often that—

Get back here.

Yeah, there we go. So one of the problems dealing with real-world assets is often that the reality differs from the theory, right? Like, so for example, Hint combines public records, local environmental signals, et cetera. What happens when, you know, let's say the property records say that, you know, from the county says that the roof is 20 years old, but the homeowner has like some random like 2 pages saying that, oh, you know, we replaced the roof like, you know, 3 years ago, whatever. So what happens when you have these conflicting sources of, um, you know, data, and how does the system actually handle that kind of conflict?

Yeah, great question. So this is part of the technology that we are working on patenting. Um, I think as with any data in the world, but especially property data, you're right, there is conflicting. Um, and so what we have created is a graph database with a taxonomy that is specifically for the home. And it's really— there's a lot of these like home taxonomies, but they're like designed by governments and like insurance companies on how they think about the home, but they're not really how the homeowner thinks about the home. So we created one that is about how like a normal homeowner thinks about their home. And one of the like key pieces of this is tracking provenance, right? Because every piece of data about the home is a claim, right? And so we want to know where did that claim come from and what is the date and like where was it discovered, right? That is kind of like a key part of piecing together the living history or the story of your home. And so when you input data into Hint, and that can come from multiple places, right? It might be from public data that Hint is pulling in by itself. It might be from an inspection report that you upload that's a PDF that it reads. It might be from an insurance doc that we fetch on your behalf because we're Pulling in your declaration page. It might be something you said in a chat, right? What we do is we enter those facts into the graph and they get associated with like a number of entities. It's, you know, it's a triple store. And so that will just be a claim. You know, how it really plays out in most cases is like if you take my pool, for example, when we moved in, the pool pump was working for about a year and then the motor seized and it had to be replaced. So Hint knew about that. And then it knew that I got a new pump. And that the old one was taken out of service, and it knows who did the work, and it knows when that work was done. And then it also knows the properties of the new pump in terms of like the horsepower and the brand and, you know, the serial number and all of that. And it knows all of that just from the invoice that I uploaded for the work. But now, because of the graph structure with provenance, it can track that over time, and it can understand that it was true at one point in time that I had a Hayward, like, 1-horsepower pump, but now I have have a Hayward 1.5 horsepower pump. And that I think is really key to like meeting the homeowner where they are, right? Because usually technical systems get very confused on data like that and it creates really poor experiences.

Hint also gives advice, right? Like it gives advice on like, you know, you should do this or you should change this, et cetera. One of the problems that you see, I think, for AIs giving advice is this issue of liability when something goes wrong. So how do you guys think about that problem? Is it something significant? Is it something you have to insure? How do you think about that?

Yeah, another good question. So there's a few things I think that are important to think through on safety around the home. Like, the worst thing that you would want is Some AI telling you to like touch like 240-volt electrical and like do that, like, you know, just sort of haphazardly. Or you might not want it to suggest that an 80-year-old person get up on a ladder and inspect, you know, the gutters on the second floor of the home to see if they're clogged or something like that. So these are the kind of things that we're really focused on now and preventing. And the way that we have gotten most of those issues ironed out— and by the way, I think the models are generally pretty good at This. Like, we've tested a number of models, and I think we're using all OpenAI models, and everything that we've seen, there's nothing super scary in terms of what it's recommending on all the— we have about 5,000 users today, and we've, you know, kind of looked at that. But what we try to do is really personalize Hint to you. It's not just personalized to your home, but to you, right? So we do ask you what your date of birth is, and that's so that we can track over time, like, your general physical capabilities. Right? So again, if it knows that you're over a certain age, it's not going to recommend that you do certain things. But on the converse, it wants to know who else is living in the home. So if you have a teenager living in the home, that is a different story, right? You have an 18-year-old there, um, and so it will know that as well. And maybe it might recommend that the teenager, you know, go up and look at the gutters on the roof. Um, and then it also just generally knows what is your DIY, you know, capability and desirability, right? Some people are like 100% DIY and that's how I want to do everything. And then other people are just like, I don't have that gift. It doesn't, doesn't interest me and I, and I need to hire a professional. And so Hit knows all of those things. And then, you know, of course, just like any other business, we have business insurance. We are looking at, you know, do we need special insurance, you know, for AI, you know, specifically? I think that's an ongoing question in the industry. I don't think that there's been enough court cases for everybody to understand that, you know, deeply. But yeah, I mean, I think, you know, we approach it from that safety level. And then we have like a lot of the like standard things in place where, you know, we do have a contract with OpenAI, you know, that has been looked at through lawyers. And then we also have the business insurance.

So let me dig into that a little bit. Like, what kind of AI insurance are vendors kind of proposing to you? What kind of terms are they offering? You know, are they offering to insure you for the year or is it like a percentage of revenue? Like what, what are you seeing in the market right now?

Good question. I wish that I was involved in those conversations more, you know, given your question. I unfortunately just not super involved in the details of all those. So I wouldn't be able to say. Okay. Right. They keep me out of the basement writing the code, you know.

Can you tell me more about these strategies for getting teenagers to do their chores in a timely fashion? This sounds like a superintelligence achievement.

Yeah, we're still working on that. Can't offer much in the way of convincing the teenagers to do it, but in terms of who's best equipped to do the work, I think Kent would know that.

This will be the last thing for humans, which doesn't say much. How about on a more serious note, the process of matching? So, I mean, I see, I don't know all the details of the business model, but I see that there is kind of a mix of the classic, like, do you have or do you need a mortgage? Right? That's been a tried and true path for monetizing apps related to home. And you can get people into a new cable, whatever. Then there's also the local service providers, which have historically been much harder to monetize, I would say generally, but certainly there's a huge market there. I'm really interested in how do you think about matching and reputation? One thing that I did that I think made my neighbors think I'm weird, which is probably accurate, is I took our neighbor group chat. Exported it all and ran it all through an LLM and had the LLM create a spreadsheet of every contractor that had been mentioned over years and what people had to say about the experience of working with them. This creates a pretty valuable asset for my neighbors at some reputational cost to me, interestingly enough. But it's still very partial compared to what an actual platform could do, right? And especially if you know about the home and all the service providers in more depth, what kind of stuff they've done. So big thesis, I guess I have in general that I want to get kind of your perspective on is anything where search costs and matching costs have historically been high should be areas where we see dramatically more efficiency and probably a lot of market growth. There's probably a lot of things that just would get done by professionals, but don't get done because I can never find the right person in the first place. Quality should go up. Consumer surplus should go up. Tell me your vision for that and how do you think we get there?

Yeah. Another big homeowner topic. Interestingly enough, my neighborhood, like our street has a Google Sheet that we all maintain and we have a list of like all the service professionals And that was kind of the inspiration for some of our future features. Like, you know, we don't have them today, but there is totally a concept of like neighborhood data sharing that I think is super interesting because a lot of the times when you're finding out like some referral, whether it's insurance or whatever, or you're just learning information, right? Like one thing that people just love is to know like, what are their neighbors doing with their homes? Like, should I be thinking about an expansion on my home? Should I be thinking about like finishing my basement? Should I be thinking about redoing my roof? You know, my home was built by the same builders at the same time as the other homes on the street. So I think this concept of like neighborhood sharing is really interesting. The way that we're approaching matching now is, so if you look at like service professionals, right? So we want to help the user understand how to think about everything, right? Because there is a right way to think about it and you kind of have to understand the big picture. So you can get a service professional from various different avenues, right? But what is the difference between like a TaskRabbit service professional versus one that's in your local area versus one that is in like some service provider network versus one that comes from the home warranty maybe that you have? They all have their pros and cons. And then, you know, who you are as a person is also weighted into that, right? Um, I may be a person that cares way more about quality And I'm willing to pay for it than I am about like the lowest penny. And I just can't stand work to be done on my home in like a poor fashion. And this is where I think AI gets like really, really interesting because you can't really do that with traditional technology, right? You really need something that can look at text and preferences and understand a user and then understand options and make a recommendation that way. And so like, as an example, we have a partner, Thumbtack. They're a national service network and they have Landscapers, plumbers, electricians. I mean, anything under this— junk removal, anything under the sun that you would need. And so one of the things that's interesting that Hint will do is it'll use their APIs, like in chat, it's all agentic. So it'll hit their API. Like where it was impressive to me is I have bluestone walkways on my property that get cracked, you know, over the years because, you know, winter is tough on them and the ground freezes and expands and then contracts when it thaws. And so the walkways just get Really, um, out of place. What is the type of person that fixes that? Like, I have no idea. I only know today because Hint told me that that's hardscaping. So I wouldn't even know what to search, right? So Hint does a search for like people who do hardscaping work, not landscaping. Um, you know, it reads all of the reviews for all of these people to make sure there's nothing scary and to make sure that they're experienced with bluestone walkways. It'll sort them all by proximity to me and then order them by the number of jobs that they've done as well as their rating on the platform, right? So right off the bat, it has just like provided tremendous value to me. And then it will give me 3 options. And this is where I think, you know, we always need to be like super careful with the AI tools. There's always like maybe one recommendation. I think that's really just not the case, you know, with things about the home. If you have the time, you really should call 3 different people. Feel them out a bit. Sometimes it's a gut call, like, do you trust this person? Are they saying things to you? Does it feel like they're trying to like upsell you and overcharge you? And if that's the case, don't go with that person, right? So the other thing that Hint will do is help you, you know, what do you even say to these people when you call them? You know, you call them, you have your bullets, and then you'll come back to Hint and say, okay, well, here was the quote, or here was the estimate, or here's what they say. And then Hint can help you, you know, make the decision on who you go with. So it's less about like one perfect real-time recommendation and more about how do you think about service professionals? Who are you as a person? And then like, how do you find the right out of like the right one out of like a list of 3?

Going forward, do you think that you will have like voice agents call these folks on people's behalf? Because a lot of these, you know, one of the, as a side note, one of the highest bits of praise that somebody can get in the neighbor chat is he really knows these old homes. And so, you know, there's some of that that's in the neighborhood collective knowledge base. There is more that you need to find out, and that's where you, you know, you recommend like calling and feeling them out. But I don't want to do that. And I think the AI in some ways could do a better job, probably would have better questions to ask. I also have no idea how the pros would react to fielding AI calls. Would they just hang up on that? I mean, have you done any market research? What do you think is the future of— I mean, it could be this, could be something else, but it seems like there's a kind of new social dynamic almost that will likely evolve here. And I wonder what your— crystal ball suggest that might look like?

So we've tried this. It's very interesting. Not what I expected would happen. I'll say it's very challenging and not just challenging from a technology perspective. Like, just as an example, a lot of the service technicians are out onsite at calls all day, right? And some of them don't have an office that you can call. And even when there is an office that you are going to call, people take lunch and so the phone doesn't get picked up. Right. And so I think one of the things that's just challenging in general, AI or not, is just like making contact, right? Like, are you— like, I call you at 9:00 AM, you're not available. You call me at like 2:00 when I'm on a call. And so now we're just playing like phone tag and that's really tough. When we trialed some AI technology for this, what ended up happening, and, you know, maybe the technology's just not there, it would call a service professional like 17 times in a row until they picked up. And then that service professional, you know, this happened to be a person that services generators, is like, holy crap, there's like a life or death emergency. I better like jump off of this job site and answer this call. And then they get on the call and the AI is asking bizarre things, right? It's like, I need to know what the model number and brand is on, you know, this homeowner's generator, which it doesn't need to know that, right? So, um, I think the technology definitely needs to evolve. I think there's like just general logistics challenges. I think. The service professionals that we've talked to are definitely interested in this because they have the problem on their side as well, right? They are busy. They're on calls. You know, they can't always answer the phone. You know, their job can't be answering calls, you know, 12 calls every day. And so they want a solution as well. I think if I had to guess, I would suspect that's going to be like agent-to-agent communication, you know, in the future. My agent calls, you know, the landscaper's agent and then they have a conversation and they exchange neuralese that we can't read and then eventually make a deal and the rest of us just have to live with it. Exactly.

Yeah, that's pretty interesting. I did have an interesting experience, by the way, in terms of just the technology evolving. This might even be a chance to bring Q into the conversation a little bit. We have an AI co-host that's powered by OpenAI's voice model, but I called my local pizza place not long ago And it had an AI voice order taker agent answer. And I was pretty impressed actually with how good it was. I had a pretty good experience with it. I got a little bit off the rails when I started asking it what AI company was powering it. It was not meant to answer those kinds of questions, but still, it's all changing pretty fast. Q, you're here with Kyle Rush, who's the CTO of Hint. which is a company that was co-founded by Martha Stewart and is helping people make sense of their homes and figure out what to do with them and connect with service providers. Any questions you would want to ask?

Yes, Nathan.

I'd ask Kyle first how they keep recommendations neutral given affiliate revenue and what guardrails prevent bias. Second, how they blend Martha's structured home keeping knowledge with live model output without the assistant hallucinating maintenance advice? Third, what early pilot data showed about proactive savings versus reactive repairs that would give the audience a clean sense of trust, architecture, and real-world impact?

How cool. Those are great questions. I love the first one, by the way, on like affiliate revenue and stuff. So we The guardrail we put in place is our AI knows nothing about our revenue. It doesn't know where it comes from. You know, it has no instruction on selling you anything. It doesn't want to sell you anything. If you need something, when you ask, it will try to help you. You know, you don't have to book through Thumbtack. You can if it's easier. We really built the integrations like, you know, with Thumbtack and, you know, shopping for insurance just to make that process easier to save you time. But there's no requirement that you do it in Hint. Like it's not making any recommendation by whatever affiliate, you know, kickback we get on it. So that's like the main guardrail and that's kind of just the experience that we want, right? If you want the recommendation, ask for it. If you want it to be easy to book in-app and you don't have to leave, that's cool. If you don't, it'll recommend something on Google and you can go, you know, research it yourself and it will help you with that. The other question is also super interesting. So our whole approach And there's a lot of LLM hallucination around the home, let me tell you. I won't get into all the examples, maybe if we have time, but— so we take all the data on your home, and there's a reason why we built a system that grabs 1,300 data points on your home after onboarding, like once you give us your address. And the reason is largely because of these hallucinations, right? So we want the AI to be deeply personalized and we want it to know about your home. Then we don't want this to just be AI advice. I don't think anybody wants that, generally speaking, on their home. They want like expert human advice. And so our tagline is the expertise to maintain your home exists. It's just been locked in the expert's heads. You haven't had access to it. So we take people like Martha Stewart, who's written tons of content on this and, you know, tons of television commercials. We take that content, we get a license to it, we put it, you know, into the app. And then the only thing that we're really using AI for at the end of the day is to personalize that to you. So we have the human expertise, we know about you and your home, and then we give you the personalization with the AI. Beyond that, almost everything that comes out of AI is directly attributed back to like a real fact about your home from the data or something you've said. And then the advice is almost always from a real human expert that we work with.

So, you know, this market of like home service providers, who deal with homeowners is a market which has traditionally been hard to access because a lot of the providers are very fragmented and there are many, many, many providers. And we've seen like several approaches. One, I think like Thumbtack and several others were using this kind of marketplace approach. One, I think Nextdoor was using kind of a social network-based approach. Like, what is the differentiator? And those are basically kind of like generations of technology, right? You had the marketplace generation, you had the social network generation. What is this new generation? What does this new generation look like? And what are the key differentiating factors which are, you know, from the prior 2 generations?

Yeah, I think this generation is just like deeply, deeply, deeply data-backed, right? We're in a world right now where If there's a service provider, we can Google that person, we can look up lawsuits against that person, we can look them up on the Better Business Bureau to see what complaints there are. And by the way, this framing is making the service professional sound like the bad guy. That's not really what I'm implying. But what I'm implying is that, you know, in the traditional sense, or I would say like phase 2 of service professionals, you are searching and then you're getting like a list, right? And so now it's up to you to go through through that and figure out who's the right person. That's too much onus. Like, you're not going to do 3 pages of Google search results. You're not going to even know to look them up on Better Business Bureau. You're not going to know— you're not going to read all of their reviews. I mean, maybe some people do that, of course, but now we're in a world where something can do that on your behalf. And it's a very interesting world, I think. I think it will produce much better quality. And then not only is it grabbing all that data really fast and going through it really fast, but also tying it back to you and your home, right? So you go back to like the personalization. If the person that you're looking at doesn't do hardscaping, has never done that job, why are they in the mix, right? That's really hard for people to know in like 2.0 of service professionals, right? Because again, you're just searching, you don't know the terms, you don't know how to order them or sort them. So I think this one is, you know, this sort of phase is Deep data and deep personalization.

So, the obvious question, the next question for me is, in the search era, you had search engine optimization, and you had this process of advertising in the search results itself. And we started off with mostly search engine optimization, and then as time went on, those blue links that, you know, got bigger and bigger, and we had fewer and fewer, um, kind of authentic search engine optimization, and that itself was gamed. So are we entering this kind of phase of, um, you know, AI-optimized recommendations? Like, how does a search provider— how does a provider optimize themselves to be discovered by AI? How does this discovery process work?

Oh, that's a good question. You know, I think it depends on the platform. On Hint, it's not a topic we've really thought about. I think it would be hard to optimize on Hint the way that it's built today. I think what I'm seeing a lot of companies do is they'll just try to put tons of content about themselves on the internet. You can trick a lot of AIs that way, by the way. I see studies all the time where AIs are just treating things as fact on a website that was stood up yesterday, right? And it's got non-factual content in it. So That's one way that like, you know, some people may do it to try to game the system. But as long as the AI is like trained and directed to trust like reliable sources or checking specific sources, I think that will be hard. But what we're doing at Hint is we want to one, help you understand the general zone, right? So like you might think that like, oh, I'm new to Thumbtack, so I won't get business. I don't think that's actually true in Hint. It matters like who you are and what your store is. areas. I think there's a big advantage to new professionals to Hint because a lot of them have been at the bigger companies for like 15 years and now they want to start up their own. That's hugely advantageous to you in several scenarios, right? Like you're not gonna have to pay the big company prices, but you're getting the big company talent. And so that, I think, you know, it doesn't really need optimization. It's really like that matching that we were talking about earlier. But we're also working with a lot of companies to figure out what should their experience be in Hint, right? Let's say that you're like sort of traveler's insurance, right? Hint talks about insurance. What would it say about travelers? What is the opinion, you know, on travelers? How should the, the, um, the user think about it? And it turns out that for mostly everything, um, whether it's insurance or service professionals, there's like a right company at the right time for the right person, you know? And that's what I think it comes down to. Um, and so what we're trying to do is proactively work with as many companies as we can to make sure that the data that we do have access to is accurate. Whatever knowledge base they have or whatever company positioning that they have, we have access to that, and then we would use that for the recommendations.

Do you see people trying to prompt inject into your models? Have you seen that? Have you seen that happening? Do you see that as a risk?

I haven't seen it happening yet. Maybe I just haven't noticed it. I think it's absolutely a risk. I mean, there are like text inputs, you know, in our app. Just from an engineering perspective, I know this is not the only defense, but I try to be very, very strict with my data models, right? Anything that we're storing, We'll have like a 20-minute conversation about every field. How do you store a zip code? What are the maximums on it? Like, you know, how do you validate it? Like, all of that is just like anything that gets input goes through that process. But, you know, I— we're so new that we really haven't had to deal with this. We are working on a lot of like sort of just general security features. But I think at the end of the day, you really have to think through like, what does a model have access to, right? Does it have write access? Does it have read access? And so we do a lot of thinking, you know, just in case there is a prompt injection attack, what's the worst that could happen? And in our sort of environment, it turns out there's really not much, right? Like, it can know things about your home, but like, it can't book anything for you, it can't buy anything for you right now. There's just not a lot of like bad things that could happen. I guess it could like change some of your home data, but we built in a guardrail there where like anytime it wants to change data, it'll confirm it with you. So those are just some of the like more practical, unsexy things that, you know, we're working on in terms of prompt injection, but hasn't been a big problem for us just yet. I don't think maybe homeowners is the biggest like market of like, you know, prompt injection attacks.

Yeah, the jailbreakers perhaps not known for their flair for historic homes. One more question for me just on kind of product and business strategy over time, right? Of course, it's sort of the received wisdom that you want to be doing something that the foundation models can't do or won't do because otherwise you get steamrolled by the next generation of the model. So I guess I have a couple related questions. One is, do you envision a future where You become sort of a tool that agents consume. I am trying to break down all walls in my life. And so now I've got the new Hint app, but I kind of want to be able to tell Claude, go talk to the Hint app. And then in general, in the not too distant future, I could probably say to Claude, hey, I've got— here's a picture of my broken stone on my path. Figure out what this is, go interview 3 local companies, what have you. And it could probably do a lot of that. Not today, but it's coming. So what then is kind of the next thing? Are you thinking like 3D models of homes or your own foundation model for kind of predictive maintenance? Like, what is the sort of frontier tech that you can develop that you would feel, you know, pretty safe that Claude won't encroach on?

Yeah. Um, so on— so yes, I think you can— you will be able to use Hint like in multiple scenarios. Like we'll have an MCP eventually that, you know, can hook into Claude and to, um, ChatGPT. Our kind of motto is like, use Hint where you are. Like there will eventually be an iMessage interface, right? If that's how you want to use it, that's cool with us. Um, the MCPs will have obviously limited functionality. There are some things that you just have to do in an app. And so, you know, at some point you may have to open up the app. And then I think in terms of like differentiation and like moat and protection, the biggest thing is just like the data. You know, when I think of Claude and ChatGPT, it's like, what does it actually know about my home? And it also, I don't see them getting better on the hallucination stuff like anytime soon because there is so much data on the home. An example of that is my hamlet in New York, which is kind of unique. I think it's called Katona, and it's the, the governmental jurisdiction is 2 different bodies that, you know, cover Katona. So Martha lives in town of Bedford, and I live in town of Lewisboro. So our tax system is different. And whenever I talk to any of these AIs, even my work Claude that I work on, you know, on Hint, it still thinks that I'm in town of Bedford. So everything is just wrong, right? Anytime I ask about taxes or regulations or how many chickens we can have on our property, it's all just wrong. And so until there is some way that like Claude figures out how to correct that problem, I think you're just going to be getting a subpar experience. But I also think the thing that at the end of the day, it's all about content, right? Like what are you training these models on and what do they have access to? And We have a license, you know, to content and we're creating original content that Claude or ChatGPT will never have. And that's just what is the right way to do homeownership. Um, and that, you know, is partly written by Martha and partly written by some of our other experts in the home insurance space. And that's all proprietary IP that we would never license out. Right. So I think those 2 things are like pretty big moats, but of course we're working on very cool features. Like one of the things that I think is awesome is just like UI and visualizations. that I don't know that the other companies will ever get truly right. But, you know, we got Hint to show me the sunrise and sunset over a 3D model of my home for every season on every day and all times. I think it's just cool to sit and play around with that, but it turns out it has like really practical advantages. I have wood rot on the deck behind my house. It's a wood deck. I don't know why that's happening. It's just, there's wood rot, you know, but Hint is very able to look at that 3D model, look at all the situations and say, you know, Kyle, you're— the back of your house, that deck basically gets no sun exposure every day of the year. And so it's just constantly in the shade, and the sun does not dry out the water, you know, that accumulates. And then also you have sprinklers hitting that deck. Bingo. Now, now I know what was happening. But the problem with the other players in the space is you have to know the prompt. And this is like the biggest thing that I think we provide provide value on is how does a homeowner even know to ask that? How do they even know to ask that wood rot can be solved? Like, how do they even know what wood rot is? You have to know the language. And so the problem with Claude and that you're mentioning is it can do amazing things for you, but you have to know how to ask and you have to know to ask. And with homeownership, you're only going to know that language and that vocabulary after like 20 years of it. And that's the, that's the shortcut that Hint gives you.

So, um, I have a, uh, question, unrelated question that I have to ask, my last question. Um, we're obviously going into the midterm elections, um, and you have, uh, you know, a lot of experience in fundraising, uh, for campaigns. So as you are also very experienced with AI technology and you're at the forefront What kind of superpowers does having AI give you in fundraising for politics that did not exist, you know, 5, 10, 15 years ago?

Hmm. I wish I had prepared for that question. That's a really good one. I mean, look, you know, my time in politics was a while ago. Politics moves very fast. You know, back when I was doing it, we were building the technology that you can just buy now, right? Like we didn't have anything. Like when I was on the Obama campaign, we built some— this is gonna sound so silly and basic, but you know, we got a lot of awards for the tech that we built for foundation or for donations because you could like save your payment information to make a donation in the future. And you know, it was kind of like at the time Amazon was coming out with like one-click purchases and stuff. And that just seemed like groundbreaking, you know, and, and also just providing like good experiences on the donation page was just such a like breakthrough thing at the time. So my, my time is pretty dated on that and I haven't thought a lot about like AI for fundraising, but at the end of the day, my thesis on AI is just like personalization, right? Like, I think if you can understand me as a donor and what drives me and what I care about. You can craft like personalized messages to me and you can speak to me, you know, how I need to be spoken to. The one thing that's just amazing about human beings that are really good at this, like any good fundraiser, you know, will talk to you for like 5 minutes and they know immediately how to talk to you after that, right? They know who you are. They've gathered the information. They know how you talk. They know where you're from. They know what you care about, your political affiliation right away, where you are on the spectrum. And then that colors like the whole conversation with them. them. That was definitely not possible with the tech that we had in the past, right? But it is possible now, and I think you probably will see some of that. And that's probably my thought on, you know, how AI changes the fundraising landscape for campaigns.

Are we gonna see voice bots?

Oof, I think you will because it's like those were before AI, right? The robocallers, like you would get Like robocalled for polling. You would get robocalled for like campaign messages. I think you probably will see voice bots. I think they're everywhere at this point. I haven't gotten any calls. The tricky thing about me is I'm not in a competitive state. You know, New York isn't super competitive. If you live in a more purple state, you're just getting hammered, you know, with these things. So I'm very curious what people in like competitive states are seeing in terms of like the outreach they're getting. Is it more personalized? Is it AI? Is it not? I wouldn't know.

Nathan?

That one's gonna fall on me. Here I am in, uh, classic swing state of Michigan, and, uh, I haven't really been answering my phone at all, so I might have to change my practices and pick up from time to time. But yeah, as of now, I wouldn't know either. But not because— I'm sure that those calls are coming in, I'm just not answering them whatsoever in today's world. That, that might ought to change a little bit. I'm always a fan of a gonzo journalism assignment.

Just pick it up, see what the bots are saying.

Indeed. Kyle, thank you for joining us today. It has been a pleasure. We have, we have learned quite a bit and we hope to see you again.

Thank you for having me and I hope to be back. Bye.

Here's to home intelligence. Bye for now. Bye. Cool. I think that definitely— I haven't got to the part of the app yet where you do the 3D sun over the home, but that is the kind of thing I think that these apps are really gonna need. And, you know, at some point you get galaxy brain enough, you're like, well, Clog could just code that. But there is, if only an efficiency reason for, you know, building that once and making sure it really works and then serving it to a lot of people. So I think that kind of thing sounds enduring. I was less compelled by the idea that like, you know, they will continue to have hallucinations. Like, I'm sure they will, but they're getting awfully good at that as we know. So I, but I like that 3D model of the home thing quite a bit.

It strikes me how high value the app is, right? Like it's deeply contextualized, deep amount of data, extremely high value kind of suggestions. And it strikes me how different that is from a typical kind of home services marketplace where you really kind of start off like, you know, shooting around and like you don't know if the guy that you're talking to has experience with a particular, you you know, device that, you know, you're— you, you have to fix. And, you know, there's this whole sequence of like trial and error that you have to go to— go through. So it really strikes me that if they, if they can nail it, um, how much easier that whole discovery and search process will be, right?

Yeah, my dad's rule of thumb is that you should— this is, you know, what he tells me as his kid. He's like, you should expect to spend 2% of your home value every year on maintenance. And honestly, in my jurisdiction, given property values and the age of my home, it's probably higher than that. So it's definitely the kind of thing where you are at a really valuable wedge point in somebody's life. I mean, you just go to the recommendations or the services tab of the app and it's like, You can see that, yeah, there's a lot of opportunity to monetize. I think my guess is that they, it's home insurance, internet, finances, mortgage, property tax, utilities. My guess is for the foreseeable future, then they do, as you said, have a Thumbtack integration too. But my guess is for the foreseeable future, they could just make enough money by just getting the occasional person refinancing through them or signing up for a different internet provider or whatever through them. The bounties on those things alone are pretty large in a lot of cases. And so, you know, you can, they're definitely in the phase where like they don't probably need to monkey around too much in the more, you know, bespoke local recommendations. You know, that would be a, my guess, a relatively, you know, I mean, the singularity is near, obviously, right? So like, what's a far future optimization at this point? Who knows? But as, as futures go, my guess is that one's like fairly far down the path for them.

It strikes me that, you know, their go-to-market is via Martha Stewart's name, right? So they have the Martha Stewart name in there. They've got the branding and they are in consumer, so they can't really charge directly. They have to sell the, you know, they have to find providers willing to pay a margin, affiliate fees of some kind in order to be in business. And it is very high value, right? And it strikes me that the plays that we have seen that attach themselves to high-value assets, specifically homes and cars, are the ones that have worked out, right? Airbnb and Uber and, you know, et cetera. Like, you have to attach yourself to a high-value asset and then take a margin of something that is being done to that asset. And that has been plays that work out, but you have to also deploy funding into go-to-market and marketing. So it also strikes me that, you know, at the end of the day, They may be a harness for a model, right? They're basically providing a context harness for AI models. They are already deploying, you know, the API in there. So they could just not deploy the API and just provide the context directly. And, you know, ChatGPT or Claude can interface directly. I wonder how big this business of like just storing context for people in organized ways is going to be?

I think it's probably going to be pretty big. You know, personally, I want to own as much of that as I can. I really don't want to be beholden to too many apps. I think this home one is kind of on the border where I'm like, yeah, I could see that working fine. We didn't ask them. Maybe we should have asked, like, is there an export feature if I ever want to leave. But, you know, I definitely have invested a lot of time, partly because there was no other option really that I felt great about. But also, even if you told me, Nathan, you didn't need to build all this deep context crap locally and this wiki on top of it and, you know, the multiple versions and all that. Like, here's deepcontext.ai. You could just go put all your stuff in there, connect all your connectors, and it'll suck up all your stuff and it'll just have all that for you. I mean, I'm pretty much a frontier proprietary model user, and I'm not necessarily that keen to run inference, but I do want to own my own context. I want to be able to at least look at my own stuff at some point if I need to. So I think it is a big business, but it's funny, on most sides, on most of these issues of sort of sovereign individual or you know, convenience of the platform. I usually go convenience of the platform. Assembling my own context and having it like on my machine in a way that is beholden to no one, that is one place where I've taken the other approach and really wanted to make sure that my context is mine. Yeah.

It's, well, I guess if you have the export facility, but it gives you assurance, but perhaps you never end up using it because you automatically have that kind of market power to withdraw. So even if you don't use the export facility, just having it there.

Yeah, it's worth a lot. I mean, one of the greatest AGI moments of the last few months for me, I probably told this story another time, but I bought this burner phone to take to China, took hundreds of photos over the course of 2 weeks in China, but it's a totally different account. So I have an Android phone with a different Gmail that I'm also trying not to commingle too much with my main internet identity. But I wanted those photos on my main phone so I could look back at them and not have this gap where they don't— oh, that was on that other phone. I never managed to get that. Claude was able to, and it did depend on the Google Takeout service, but Google Takeout, first backup to the cloud, then take out, then all your photos are there. And Claude was able to put those photos into my iPhone library, my iPhoto library, I guess I should say, in a way that preserved all the metadata. So it has the location of where the photos were taken. It comes up just like it normally would in my iPhone with the live photos are still live. I don't know how it, if those 2 platforms use the same live photo, if there was some conversion there, these details, you know, I didn't have to concern myself with, but it was the kind of thing where I was like, I would never have done that on my own ever. And yet Claude was able to just figure that out for me. And that is where the access to the raw data was like really key. Because if I'd been blocked on that step, there would've been no way I could have done it. So I do think that even if you don't use it for a while, I would definitely preferentially go for services that do have a, you know, a deep kind of export function in the way that Google does with Takeout.

Indeed.

How you doing on time? You wanna do the second half of our Guess the Lines game?

We got time for it. Let's do that. Let's do that.

All right, so we are on as we come back for the second half of our Guess the Markets game today. We are halfway through, 7 out of 14 rounds done. And let's just do a quick recap of what we've covered so far, and then we'll get into the second half. Will the AI bubble burst? These I thought were pretty extreme Scenarios that would count as the bubble bursting, and 3 of the 6 have to happen. So I said just 2%. Market's at 10. You were at 15. Turns out one of the 6 already happened. Um, that was like Supermicro being down a lot. Still feels like it's very hard to imagine a world where those 2 more of those happen in the next 4 months to me, but obviously time will tell. Will Astra be released by the end of the month? Market was higher than both of us. There was a little rules layering on that one with respect to the name Astra. Will the US government take control of any AI company? I said 50/50. You said 90. Market said 40. Will it be OpenAI specifically? I said, yeah, more if it is going to happen, more likely it'll be OpenAI than not. You said probably not, which I think there's an interesting story that at some point I'd like to hear you fill out in terms of What else do you have in mind? You said Anthropic. Anyway, market was kind of between us on that one. Will Anthropic have a higher market cap than OpenAI? I said 2/3, you said 50/50. Market came in just shy of 70. On that one, we were kind of debating the narratives. Will it be a, the best models are running away from the pack versus a model commoditization versus who owns the most compute. And then you also brought to the discussion the question of how Anthropic records their revenue and whether or not they include the share that goes to their partners, which I think right now they're not doing, but they might have to do in the future. Will Anthropic's ARR be under $90 billion at the end of the year? Markets is only 15%. In other words, there's basically a What, what is that, a 5/6, uh, chance that the revenue will be over $90 billion? Presumably that is using the current accounting. And then who's the world's second trillionaire going to be? We both liked Zuckerberg and, um, Google founders. You like Jensen Huang a lot more than I did. I was more willing to tell a story about Sam Altman around his investments in nuclear fusion, for example, but the market gives a much higher weight to Jensen and specifically to Zuckerberg over the rest. So that brings us to the second half. As we start the second half, the first question is from Polymarket. What kind of consumer hardware will OpenAI announce in 2026? This is a good question. Each row resolves yes. So apparently we don't necessarily have to sum these to 100. This is for each one, what are the odds that it will happen?

It could be more than one, right? It could be multiple products, right?

Yep. So each one of these is basically an independent question. So essentially we are now tasked with 7 independent independent, uh, estimations here. So clip-on for clothing, earbuds and headphones, glasses, necklace, watch, ring, or phone?

Um, clip-on device for clothing, I'm at like, I don't know, 5%. Uh, earbuds, headphones, slash speakers. Um, yeah, is speaker in there? Because that, that is, uh, the smart speaker is already kind of known, but they didn't, they didn't like put it out there. So, um, yeah, I'm— I guess earbuds, headphones. Uh, I, I don't— we already know what it is. It's like, it's like a little smart speaker with a kind of arm which kind of moves around, which is the eye. And it sits on your desk and it can kind of react to you. And it can also look around. So they're trying to emphasize that it has an eye and it has a voice and it has— it's able to listen to you. And so they're trying to make a new category of device altogether. And so this is a little bit tough because it doesn't Neatly slot into any of those.

Yeah, yeah, um, yeah, I mean, there certainly could be some surprises, but I'm with you that this— none of these sound like they're on point for what I understand we have heard. And that's also what I have understood from everyone I've seen online.

Maybe I do like all 10% and maybe the full I do like maybe 25% because I think there's some chance that they, if they succeed, they would try to go ahead with a phone. And phone tech is well known. So it's not like anything new. It would just be an additional device. So.

All right. So I got to do the same thing. I mean, they have been inspired by the movie Her, so maybe I'll give them a little higher chance of that. And that's at least consistent with their tabletop thing. A clip-on device for clothing. The pin. Remember the pin? Yeah. Uh, remember that little display that was— you were gonna shine it at your palm and then you were gonna read things off your palm?

Humane.

Yeah, I'm not sure how that one got out of the, out of the testing stage, although I Arguably never really did get out of the testing stage.

I mean, I, I, it was really unfortunate for them because I think they were a little bit too early. And, uh, they're very good at product design, but the AI piece is the core piece, and the AI piece didn't mature. And, um, arguably what Jony Ive probably showed, uh, OpenAI was probably somewhat something similar. But he was 2, 3 years later and he could sell it. So I think Jony Ive, I think he was bought, he was brought in and I think like, what valuation was he brought in at? He was brought in at, I think like $300 million.

Yeah.

So he's gonna be like tens of billions of dollars like at the IPO, right? So.

All right, well, this one's a little weird. Let's just go for it. The market says clip-on device for clothing, most likely. Earbuds and headphones, next most likely. Then glasses, necklace, watch, ring, and phone. They're really not expecting. So we were both way higher than the market on the phone one. Scout was kind of more with us on that. Not sure there's too much more to say about this. Let's keep moving. Number 9, will OpenAI launch a consumer hardware product this year? Now, yeah, that's another really interesting one with respect to IPO, right? Because I mean, there's a bunch of— not a bunch of, but there's As you mentioned earlier with advertising, that's kind of a revenue driver that could really change the way people think about their growth prospects. And an awesome device could do that too. That kind of takes me back a little bit more toward the phone though. I mean, phone's a hard thing to compete on, but Apple has definitely left the space open to new competition, I would say, over the last couple of years. It's like, I still use my iPhone. I still like my iPhone, but it would be a lot better if it was smarter. And that's an absolutely enormous market, something like $500 billion a year.

Yeah, but there is the problem with consumer hardware is that you need to like have the product ready by like March. You need to announce in like September. So that you can get it out to consumers by Christmas. So the cycle is like brutal, and you have to coordinate with the guys in Asia. And you know they're going to do a small run. I hear it's a small run of the initial device, probably like several hundred thousand units, not reaching a million. And I don't think they could support a phone. I think I think it's easier just for them to do like a toy device. almost the first version, um, and not make it like too serious.

So, so this last one was publicly announced. This one is launched. Announced doesn't count. Um, if they're gonna do a small run, they still have time. I mean, they could sell a couple hundred thousand They're also being delayed by the Apple lawsuit.

Apple is suing them, trying to delay the release as much as possible. So I don't know. Maybe it's a tough question. I think definitely something will be in your hands in 2027. I don't know what launch means. Does that mean— You release— I, I guess that means it says release. Yeah, in hands, right? Okay, I, I think in hands in 2026 is unlikely, so I'm giving it like 15%, maybe.

15? Okay, that's, that's definitely lower than me. I'll go 40. You talked me down a bit. I was initially thinking higher just because we've heard about this for so long now that it's like at some point it's got to happen, but the Apple lawsuit is definitely a wild card. I no idea what to make of that. So let's find out. 28.5. I think that's like almost exactly between us. Okay. Will China obtain a functional EUV machine before January 1st, 2029?

Oh, like, yeah, 80%, I guess.

Obtains or develops?

Yeah, 80%. The thing is that ASML fired a bunch of people and the Chinese are very good at hiring and they're willing to pay, right? They're willing to pay American-style salaries for a few years in order to get talent. So yeah, I think, I think they will.

Yeah, this is one of the more important questions in the world, I would say. Certainly a lot of American policy over the last couple of years has rested on the assumption that this can't be done, that they're many years away from doing this. I just barely managed to mute myself before sneezing. So, but yeah, never bet against Chinese manufacturing is another pretty good rule to live by in life. A lot of this analysis rests on, well, it's not just ASML, it's like they have these supply chains and those suppliers have suppliers. And, you know, there's like one German company in this one town that makes the lens that is needed, but without the lens you can't do anything.

Yeah.

Um, and there are a lot of those little bottlenecks, so they have to fix them all. I'm gonna just work from the assumption of like, I don't know what obtains means. Presumably like buying a used one in somebody's garage sale or whatever is like not the spirit of this question, but I'm focusing on development. So that gives them 27 and 28. I think it is not that likely. I'll say 30% that they are able to make this all work by that time.

80.

Okay, this is a thin market, so we have a little bit of caution, uh, around the estimate. May not be as meaningful some of our others. 58. Again, pretty close to right between. A little closer to you on that one. Yeah, I mean, this, this is really— so there's like multiple years being traded. The shape of this curve is where, you know, you, you really have to believe that like Both this won't happen that fast, and before it does, we're going to have some sort of AI takeoff via RSI or what have you. That's the world in which you could plausibly play the Machines of Loving Grace strategy of create a decisive strategic advantage and make them an offer they can't refuse. I still think That seems unwise, and I've got other content in the works where I'll talk about that in depth. But this is at least consistent with the Dario worldview that they won't be able to make crazy, or they probably won't be able to make crazy scale of chips. So if we can get Claude to become the country of geniuses in a data center in the next 2 to 3 years, Then we, you know, have a chance to say how the world looks after that. Crazy. Okay, next. Will AI wipe out humanity by 2030?

The fine print. Will continuing progress in AI capabilities result in every human dying by January 1st, 2030? The fine print is the point because play money is worthless if the world ends. Traders have an incentive to bet no regardless of belief. So the market resolves N/A on January 1st, 2027, and rolls back every trade. It never actually pays out on the 2030 question.

So this is basically just a chance to express your emotions, right? With no skin in the game, real money, play money, or otherwise.

No, it's like, will AI wipe out humanity by 2030? No, that's like a, I don't know, like a 0% chance. Oh, well, I have to guess what everyone else is guessing. So, all right, I'll guess like 15%. Yeah.

Well, since we're just emoting here, my guess is that there's going to be Yeah, so this is like— now we're metagaming, right? You're metagaming there by thinking, okay, not what is my true belief, but what is the market going to say?

Yeah.

I would say my true belief is probably lower than your 15%, but not dramatically lower. And then my guess as to what the market will say, if I'm playing that game, is more like 30% because I think we're gonna see a certain demographic attracted to this question.

Yeah.

But in the spirit of playing an earnest strategy, I will go 10%. Whoops. With which I will say is my like closer, maybe I'll say 8%. I'll say that's my true answer and I'm sacrificing points to send the signal that that's my true answer, even though I think the market It's going to be higher. Let's see. Oh wow, not nearly as high as I would have thought. Okay, interesting. Got some non-doomer action in the most doomer of all markets. 3 to go. Will Databricks valuation hit $250 billion by December 31st, 2024? 2026. Now they were valued at about $100 billion about a year ago, and then they just raised a big round.

I think they're $130 now, I think.

Yeah, I think they raised like $10. Um, so I don't remember the valuation that was attached to that latest round, but it was a big round. If it was indeed 130, I'm guessing they're not going to double by the end of the year. So I will say this is a funny business. I mean, it's a— in some ways it was a business that was just perfectly positioned for AI to come along, just having everybody's data in giant data warehouses and data lakes and whatnot.

It's $190 billion right now. That's, that's the current valuation.

Yeah, I mean, this would have to be like reported and as— oh, as measured by the price reported by NASDAQ Private Market. Okay, so it is a little more liquid than just the next financing announcement. All right, I still think— so 190, it's up from 100 to 190 in the last like 11 months. Can it go to 250 in another 4 months? I think it's less likely to happen, but I'm talking myself into it being not that unlikely. I'll say 35%.

Uh, I'd say 20%.

Answer is 22.5. You are right on the money there. Okay, 190 is resolved. Yeah, interesting. Okay, 2 I guess these are our bonus questions from Polymarket. Will Situational Awareness announce a fund wind-down by December 31st?

I'd say 35% chance. I think, I think the, uh, the SEC, uh, inquiry is more serious than, than, uh, people realize because, because you remember I was saying like a couple, uh, couple of weeks ago, like of all the hedge fund blowups around that size, everyone else is in prison. And situational awareness, he managed to get through it and he didn't blow— he blew up, but the fund is still around and he's not in prison. I think that I might've been premature in stating that. And I suspect that the risk management was sloppy. If the risk management was sloppy, I think the compliance would've been sloppy too. And there's a lot of people in the market who don't like him. And this is going to be a very targeted investigation on someone that they don't like. And even Elon has had so many problems with the SEC. And I think also, I suspect that not everyone not all the investors were made whole because I suspect some of the investors came in after the Anthropic investment and they're not part of the Anthropic investment. And the Anthropic investment kind of made the entire fund float. So I don't know. I have no idea, but I suspect it's going to be more difficult than people realize. So yeah. All right.

You're talking me up a little bit there. I'll go up to 25. My thought was like, well, they didn't blow up so bad that they had to be unwound immediately, and they still had like however many billion worth of Anthropic equity sitting around, and that's a pretty good cushion. So I was initially gonna come in lower, but you talked me up to 25. So let's see. 6.

Market market doesn't think it's likely. Okay, interesting.

Okay, last one: Will a Tesla SpaceX merger be officially announced by December 31st, 2026?

Good question.

15%. I don't really know how to think about this one. I mean, what do you think are like the relevant analyses? Like, it seems like Elon gets to do whatever he wants kind of in this domain.

So he doesn't get to decide this one. He has to basically take himself out of the picture. So he's not going to be able to vote his shares. So this is about whether the minorities—

And so why wouldn't he be able to?

Because he's an interested party, right? He's an interested party. He's the CEO on both. He would have to basically take himself out of the running to basically vote his shares. So the boards of both entities would have to recommend the merger. And The shareholders would get a vote, I think. And especially Tesla, he's not controlling. But even for SpaceX, as the controlling shareholder, he's got so much in there, he would not be the one being able to vote. It would be the minorities that have to vote. So in order to get it done, it has to be a good deal for both I think Tesla right now is not doing so well. And so I think you're seeing the robotaxis. I've started seeing the robotaxis and you're starting to see the robotaxi announcement. I've started to see them on the highway in Silicon Valley. So I think what happens is the robotaxis finally get fully rolled out, full self-driving. He's got 5,000 cars on the road by the end of the year. Which is more than Waymo. He's got this growth trajectory and then the Tesla shares can be managed upwards. And then he announces the merger. And I think that's what would happen. If he waits until next year, Tesla's stock price may be too high. And Starship is still going to take a little bit of time before the revenue kick from Starship happens. So I think Tesla is the thing that is maturing right now. And I think he, he has to get the Tesla price— the price has to be higher than the previous high. He's got to do at least like 10% over the previous high. And, and the price is down now, right? So he's got to get it 10, 20% above the previous high, and then he's got to take it out. So yeah, it's possible.

From a pure fundamentals basis of like business operations, I did take another little weekend trip in a full self-driving Tesla this weekend. And I hear, by the way, that the first Model YL, which is the one that I'm interested in actually potentially buying, is now in a Michigan Tesla dealer. So at some point—

The station wagon.

Go check it out and see if I actually am comfortable in the back seat. The FSD, though, is truly incredible. And at this point, really just feels like you have a kind of superhuman driver doing the driving for you. I took several people on their first FSD ride, and they were all just Totally blown away by it. The improvement was notable even from April, May, which was the last time I did it in parking in particular. It was already very good then, but in parking it's like got better taste. It sort of gets lost in parking lots less. It seems to pick better parking spots. It doesn't yet have the Grok integration to the point where you can just sort of say, park in that space right there on the right. But they said they're working on that sort of thing. The driving itself is just unbelievable, and it gives you significantly more leash than it used to when it comes to how much attention you have to pay. A couple of years ago, you had to have your hand on the wheel providing a little bit of torque all the time or it would start to bother you. 4 months ago, it was like you don't have to do anything, but if you take your eyes off the road, Then pretty quickly it would bother you. And now I feel like I was taking my eyes off the road honestly at some length, and it was not giving me a hard time, you know, enough to like do something on my phone. And I did get beeped at a couple times, but as much stuff as I was doing on my phone in the driver's seat, first of all, that just goes to show how much trust I have in it because I'm like, I'm very— when I'm driving, I'm a very attentive driver. I don't text and drive. I'll text an FSD though. So that goes to show how confident in it I am. So I guess the question I have is, how close can these companies work together already? Is there basically— because I can imagine a story where it's like, well, look, we've already got Groq integrated and we've got plans to make chips, we've got plans to do all these things, and these plans are all already basically operationally integrated via partnership. Like, we don't necessarily need to commingle the cap tables for this work to be done in an integrated way where there's tight feedback loops across the organizations. If that's the case, then maybe there's not much pressure to do it. If there's some reason that that isn't the case, I think you would probably run through walls to knock down any sort of silo walls that exist between different parts of the AI operation. And I just don't know really how closely they can work together today.

So I think the issue is that Elon gets sued a lot, right? Every single thing he gets sued and it gets reported on. So for example, I think Tesla paid for, uh, Tesla paid for the first data center and then xAI used it, right? Uh, because he didn't want to use xAI's cash flow. And they got sued for that, uh, shareholder lawsuit, right? And so the, the, the deal in the US is that, uh, everything, you know, as, you know, Matt Levine says, you know, uh, everything is a shareholder crime. Everything, everything is a crime against the shareholders. Uh, and so The shareholder lawsuit lobby basically continuously sues companies. And the more of these like intercompany links that he has, the more he gets sued on it. And every lawsuit, they either do discovery or he pays like $10 to $20 mil. So, and if they do discovery, then they dig up emails. And if the emails show that, you know, they underpriced or, you know, they knew what the market price was and they underpriced some instance or whatever. And he has no idea how, you know, these like 50,000 employees underneath are, you know, making deals or writing the emails. He has no control over that. So it's just cleaner, I think, to— if the companies are doing a lot of business together, and you are kind of like— and he's got this very Asian tycoon lifestyle where he's got like all his pockets and the pockets are helping each other. And he's got the Personal pocket. He's got personal borrowings, and sometimes you know some company does this and some company does that, and he moves money from one company to the other. So you move money from like SpaceX to XAI, and you know he's got he's got he's very Asian tycoonish like in managing the assets together. And the promise with Elon as an investor is that he'll never make you lose money, right? So that that's the promise that he's made to the market, and that's what he's. kept up consistently through SolarCity being bought and through XAI being— Twitter being bought by XAI, XAI being bought by SpaceX, and bumping up the valuation each time and then publicly listing, and everyone makes money, right? And so he has the confidence of the market because of that. And yeah, so I think, I think in the end he will merge, I think. That's the cleanest thing for him.

So, so I'm saying 20. After all that, I feel like maybe I want to go higher, but I already put it in there.

He also announced this year, which is tough.

So let's see. Yeah, not a lot of time remaining. All right, let's see. 18.5. Wow, we bracketed again.

Yeah.

Uh, all right, well, of course we can only truly score these with the passage of time. But at least for who can be the most conventional, I come out on top with 1,369 made-up points. You respectfully, respectively, uh, come in ahead of the AIs, including Scout, which had research, and Q, which did not. So Q may or may not have cheated. Probably if it did cheat, it could have scored higher than that. Um, so I'm gonna guess it didn't cheat, but you know, as we know, we should, uh, be verifying with detailed transcript analysis before we get too confident about what the AIs did or didn't put into any given solution. But if we take it at face value, at least we can say that the internet research helped. Indeed. Any other reflections on our experience today?

Well, perhaps, perhaps we can set up, set up the AIs to Kind of iterate through. We've done 3 of these so far. So iterate through and like when, you know, there's a remind me when, when the markets close, when they resolve, and then we can look at how we performed in retrospect. That might be interesting.

Yeah. I wonder if there might be a couple that have already resolved from our earlier ones. When you do these, like, you know, by 2030, it leaves us a long time to wait, obviously.

But, and, and we can probably dig up the, uh, we can also mark them to market regardless, right?

Even if they haven't resolved, if they've moved, that'll be interesting enough.

Um, indeed. Let me, let me, let me just maybe, uh, close with, uh, just kind of an iteration of, uh, what people are talking about right now.

Let me show you one other thing first, and then I'll give you the last word. All right, um, this just popped up on my feed and it is a—

Oh.

You know, some— I, I have some questions about, you know, how, uh, accurate are all these arena scores, but that's a big bump to see relative to the field for Fable 5.1. So I think we can say for sure that that is meaningful, even if exactly how meaningful it is, each person will have to sort of feel out for themselves, I suppose, over time. But that, as I was scrolling by, definitely jumped out as like, wow, it's been a minute since we've seen a model put that much distance between itself and the rest.

The Pareto frontier moves again.

All right, what have you got?

Very quickly, The beef is over. The Anthropic versus Trump administration beef is over. Anthropic are back on the right side. Commerce Secretary Howard Lutnick says said the Trump administration now trusts Anthropic following months of clashes with the AI company. Asked if he trusts Anthropic CEO Dario Amodei, Lutnick told Axios's Mike Allen, "We trust Anthropic. They've done what we asked." They are back on the right side, so the answer is yes," Lutnick said in an interview on Tuesday. So backing backing off notably Lutnick, who is commerce and not Hegseth, who is defense.

So yeah, or Michael, what's his name, who I'm guessing is not giving it up so easily, but time will tell.

So very quickly, Google announces Gemini 3.8 Flash. Uh, it is a jump, and they are targeting real-world kind of use cases. They're targeting VALS Finance, which is a financial analyst, and Harvey's Legal Agent Benchmark, which is, um, a legal agent benchmark. So it's more, more on real-world activities. Uh, capabilities perspective, Google has clearly fallen behind, someone says, but Google is now pressing their unique advantage in hardware instead 300+ tokens per second on sticker pricing. So they're offering high tokens per second with moderate kind of intelligence improvements. And I suspect that that may be quite attractive to a lot of people, like speed, speed may actually matter more than intelligence for a lot of tasks.

What's the price of this one? I think they had dropped the price as of the the last one too, right?

So the same price as 3.7 and is available on the Gemini API. So same price as the 3.7.

Yeah, it does come in a lot cheaper than a lot of other things for sure. 75 cents per million input tokens for now, although this is introductory pricing.

MTS Live says the US government has sided with OpenAI against the New York Times. The DOJ says training an LLM on copyrighted works does not violate copyright law, and that treating it as infringement would hurt US science prosperity and national security. And so let's see, let's see whether— how far, how far this goes.

Um, yeah, my guess is that's going all the way to the Supreme Court, and I like that ruling. I think we definitely, you know, it would be a real shame to let a bunch of legacy IP owners stop the train of progress. So I, I think this is a good position for the U.S. government to take. I think it would really be nice to get something back for it, though. I really do think there's another side to this trade, right? Like, they have taken humanity's collective inheritance and managed to concentrate it down and put it into a product. Great. We never got anything, though, for that. I think they should be allowed to do it, but I think there should be another side to that trade. you're allowed to do it, but what's the but? That's, I think, an interesting question for the government to tackle next. And if it's Congress and you have the ability to write laws, imagine that you could imagine starting to say, you can do this if you satisfy these other conditions, and if you don't, then you can't. And I think those conditions, there are a lot of things you might want to make, uh, AI companies do, obviously. But this would be one trade I think would be really worth thinking hard about.

Indeed. Um, next one, uh, New York Stock Exchange used Anthropic's Mythos through Project Glasswing to find and fix vulnerabilities in its systems, NYSE President Lynn Martin told Congress. So this is support for Anthropic Cyber, a little bit more. And finally, Uh, one reflection, uh, I think Dean Ball is really taking it on the, uh, on the nose today, uh, from, uh, various people in AI safety. Uh, this is David Krueger, uh, saying that, uh, Dean basically did not, uh, consistently, candidly reveal his personal kind of opinions on these issues. He states, but it is true that I, and candidly, I think many of my colleagues in the profession of AI policy, largely failed to talk about this issue with the level of seriousness and urgency it required. And this is about Identifying humanity and not being taken in by AI agents. I think what struck me on Dwarkesh Patel, he went ahead with the AI civilizations kind of blog post. What struck me was how many people responded to that with, these people are crazy. And I think what the rest of the world fails to realize is a lot of people in SF share those views. And a lot of them are hesitant to discuss them in public because they are crazy. And it is what Jensen Huang calls sci-fi. And I think that is, I think, one of the problems in communicating, like Dean and other people have to be, you know, kind of have clarity and be able to work with policymakers. Yet these beliefs are so radical that I think it's hard for them to interface. And so they end up interfacing on a normal basis, but then you have all of these beliefs that you have to, that you think may be true, in the long run, but, you know, perhaps have a lower probability and are not yet evident, right? So it's a tough one.

Yeah, I think this is unnecessarily harsh, to be honest. I mean, I, I know David a little bit, not, not well, but I've met him a few times, and I do respect the impulse. And, you know, he's got how many, uh, pause, stop, rewind emojis on his, uh, On his header there, so I mean clearly he is playing a very transparent. Here's what I think. Hold nothing back strategy. I'm not sure this is the right reaction though if you are trying to win at politics, right? So I think like what this kind of shows to me is choosing my words a little carefully myself, right? Because I don't I don't want to make enemies of either of these people. I think what I don't like about this post is he ends with an apology. Dean ends with an apology at the bottom of the post. So he's in general, if somebody is showing enough reflection and getting to the point where they're willing to apologize, that's a good moment to try to extend some grace and try to make some common cause. This is like, if you are David Krueger and you want to pause, stop, or rewind, I would think that this would be a moment to try to make some common cause, to expand the tent, to sort of adopt a little bit more of the strategy that Dean has played, which has clearly worked for him, right? I mean, he went from a think tank guy with a focus on state and local policy As of like 3 years ago to starting a blog through, I think, quite inspired writing, kind of a Hamilton story of writing his way to the top, gaining influence in like a16z circles for being a voice that they thought was very compelling on SB 1047 way back when, getting the Trump administration job. He does not get the Trump administration job. If he's seen as a crazy doomer, I think that's probably quite safe to say. The America's AI Action Plan, which is so many news cycles ago at this point that who knows if it's happening, if it's not happening, how much influence that document has in today's world, I don't know. But when it came out, it was one of the only documents ever, I would say, to come out of the Trump administration that was pretty well received across the spectrum. Even folks like Zvi Moshvitz had nice things to say about it. So you don't get that document out of the Trump administration if he's not in that role, which he's not if he doesn't play a somewhat conservative public communication strategy. And he probably doesn't get the job at OpenAI either, although at this point, who knows what the hell OpenAI might be open to anything. But I think that it's a little— I think portfolio approach is usually what I say to people when they bicker with each other over the tactics that they're using to try to achieve similar ends. I think what I would zoom out and say, look, you guys both seem to have at least somewhat of a healthy fear of super powerful intelligence at this point. That's enough common ground to build on. Let's not, as the old Monty Python, let's not bicker and argue about who killed who. This is supposed to be a happy occasion. But truly, a little more forward-looking view, I think, would be really good. Dean doesn't get to this position if he's not somewhat, again, conservative. I think he's laid it out there pretty— he also was showing himself to be AGI-pilled enough to get a job at OpenAI, right? So, I mean, I think he's played a pretty savvy strategy. I would not say this was a shameful lack of integrity. And I think like the pausers got to recognize when they have a new friend is kind of my take on this.

So I, I, I think number one, the doomer pauser side, I think they are correct in a lot of their, uh, projections on what might happen because they are also seeing progress. There's a— there's another group which is the denialists, the Ed Zitron kind of driven denialists, and that's a completely different group because that group doesn't see progress at all. So that's, you know, that's aside, right? The, the doomers are actually kind of— they do see the progress. And I think one of the things that the doomers, um, perhaps don't realize is that our view of the future is necessarily a dystopia because it is something that is different from the current status quo. So that's one thing on the doomer side. And I think on the kind of policymaker, AI policymaker, like inside the AI firms, I have to be honest and say they are preparing for the equivalent of a revolution. They're preparing for basically a new type of government. Like, people like Andrew Hall are, you know, inside— Andy Hall are inside the firms trying to define what a new form of politics would look like when you can assign, uh, decision-making power to agents and have them vote on your behalf, maybe directly in a direct democracy. Um, that, that would look very different from a representative democracy right now because now you elect representatives who then filter information. You could potentially have just your own AI, which is your representative, and you could have direct democracy. So there's a lot of future possible changes which are kind of revolutionary and are going to be potentially huge changes in the way that governance is done. And I think some of those ideas are really kind of at this point, be seen as crackpot ideas by, uh, you know, the real policymakers in Washington. Uh, crackpot ideas. These people are crazy. Uh, academics, you guys don't know what you're talking about. We're never going to allow this to happen, right? And so I think that is the interesting thing for me, because I think on the one hand, people like Andy Hall are, you know, are promoting, are trying to think of what these things might be. And on the other hand, people like David Krueger think that the future that Andy Hall is kind of trying to create is already a dystopia and is already disempowerment in the sense that human representative leaders might be disempowered and you might have AIs who are empowered on behalf of the individual human. So there's all these guys are actually seeing maybe the same future, but having a different lens on how attractive that future is. And that is interesting to me because that's basically just, you know, we have different opinions on what is desirable, even though we might see the same thing potentially happening. It's also not clear when policymakers will start to cross these kind of like belief hurdles that they have. I think one belief hurdle was Mythos. Mythos was one belief hurdle that got crossed. I think that for the first time they start to see safety issues. I've heard another belief hurdle has been crossed recently. So I've heard there is an organization called CivAI, which has been in Congress recently, and they have used, I think, KIMI or GLM or some Chinese models. They've plugged in data brokers into those models, and they've allowed those models to extract, okay, you know, if I have this person, you know, show me who this person is, Christian in Minnesota, like doing this and this, and this is their daily activity, etc. And it's all just extracted from existing data brokers and kind of joined. And this is precisely what Dario was talking about earlier in the cycle about this kind of surveillance that could be done. And I think the thing that the CivAI guys did, which was particularly good, is that they attacked the Republicans by showing how a gun owner kind of targeting system would work. And they attacked the Democrats with what an abortion provider targeting system would work. And then they provided these dossiers on these to both sides. And so both sides started to be like, oh my God, what is going on? And what CivAI is doing is they're trying to promote kind of regulations on data brokers, which people have been asking about for a like, I don't know, 10, 15, like 2 decades maybe. But I think finally we're starting to see that. And, and what CivAI is saying is that, look, the models exist. In fact, we have to use open-source Chinese models because our own, you know, ChatGPT and Claude won't allow us to do this. But we're using these open-source Chinese models and we're just plugging them in. And so CivAI does, uh, anonymized dossiers And then they show the actual product where you can type in someone's name and you can extract in real lifetime, but they don't allow you to take the data out of the room. And they're showing this to people in the Capitol. And I think that might actually get us some movement.

Scary demos. What a— yeah, they're getting really scary now. So that's been a theory of change in the AI safety world for a long time. Can we show what's possible and wake people up by just hands-on experience? And yeah, at this point, I'd say it's safe to say that some of these demos are legitimately getting to the point of scary. So sitting down and I mean, I know I've had that experience firsthand, right? When you sit down and all of a sudden the model is suggesting radical action to you. And that was my first thing where I was like, oh my God, this is gonna be a really crazy future. I think it is good that people get these experiences. So I definitely am a— You know, it's real too. That's the— when it's a real scary demo too, you know, I think that's the thing that really gets people. It's like this, I just gave you this name, you know, you did this. There's no magic trick here. You actually have a working technology that can do these things. That experience, you know, for at least some people, moves the needle. And I do think it's good that people are out there doing those scary demo briefs.

Indeed. And on that note, Nathan, we're taking a break tomorrow and we'll be back on Friday.

Alrighty. Sounds good. See you Friday, Prakash.

Cheers.

Bye for now.