[00:00:00] Jacob Haimes: This episode was recorded on September 8th, 2026. Welcome to Muckrakers, where we dig through the latest happenings around so-called AI. In each episode, we highlight recent events, contextualize the most important ones, and try to separate muck from meaning. As always, I'm your host, Jacob Haimes, and joining me is my co-host, Igor Krawczuk

[00:00:23] Igor: Jacob. This week, Ja- Jacob is bored,

[00:00:25] Jacob Haimes: Yeah

[00:00:26] Igor: to talk about Astra, Fable, and just, like, give a small update because we haven't had, like, noticeable model drops in a while

[00:00:35] Jacob Haimes: I mean, y- yeah, noticeable being the operative word there, right? Like I

[00:00:42] Igor: I mean, like, w- what we mean by that is that, like, um, failure was kind of a doubt, at least in our bubbles, where like There were rumors online that basically Anthropic also didn't get much traction with Fable, and I've read different things, whether it was the, the guardrails or just that it wasn't that good. Um, it wasn't previous jumps, and we had already been talking about like it being like a saturation line. And now OpenAI dropped GPT-6 Astra, I think relatively fast after their last round of models Yeah, I think the last one was in, was in July. So this is like a two or three-month turnaround.

[00:01:31] Jacob Haimes: Okay

[00:01:32] Igor: And this was after the Hugging Face, um, boost and in between Anthropic had also dropped like a 5.1 on Fable where they showed like a bunch of new capabilities. They sh- actually showed some jump in a new benchmark, but, uh, the benchmarks remain saturated basically, and reports are still mixed.

[00:01:54] Igor: And I think at least for me, it is worth kind of just like checking in on where we are versus like where the storyline is going or where we were supposed to be i-i-in the storyline. So

[00:02:11] Igor: So like I think maybe so it doesn't become a pure monologue, can you elaborate on why you are bored?

[00:02:19] Jacob Haimes: I don't know. I'm just like

[00:02:22] Jacob Haimes: Fable came out and, you know, there was the whole cyber thing and, uh, that they, they did this, this whole thing about, you know, we're doing Project Glasswing. And so then people, you know, reacted to that. Uh, and then like, I, I don't know, it d- it just doesn't feel like it, it actually shifted much practically, um, in, in terms of things, especially not in terms of usage.

[00:02:50] Jacob Haimes: Like, uh, I understand that it costs a bunch of money to, like more money to, to use the, uh, model, uh, over, you know, some of the other ones. But like that was enough friction for people to not want to spend the money on it, is just like charging what it actually costs. Um, and so

[00:03:16] Igor: we don't know if, if they charge what, what it actually cost, right? We, we, we

[00:03:18] Jacob Haimes: charging closer to what it actually costs

[00:03:22] Igor: We just know people weren't willing to switch

[00:03:26] Jacob Haimes: Right. But like, I mean, that goes for, for, for me as well. Like I, I don't... W- what's the, uh, I don't know, it just doesn't seem like a big, big deal. Like the The use cases, uh, but may- maybe th- this is also because I'm partially, uh, l- like not the

[00:03:46] Jacob Haimes: mean user, I guess. But, like, I, I don't know. I- I'm more interested in more discrete uses, uh, I think, in general. And so, uh, I am not wowed by I can turn over this entire task so that I don't have to look at it at all anymore, because, like, that's not interesting or valuable to me. So I... Be- because, like, then the quality goes down and I have no way of, of knowing if I'm actually deferring entirely to the system, right?

[00:04:21] Jacob Haimes: Like, uh, so I'm not gonna use it for that reason. Um, and

[00:04:30] Jacob Haimes: It just isn't, uh, worth it. 5.1, uh, you know, came out for Fable and, like, is completely overshadowed by the, you know, uh, containment, uh, or, like, the, the,

[00:04:45] Igor: Ohm, Collusion, Wiki, Breakout Stuff

[00:04:49] Jacob Haimes: Yeah. Uh, and like, regardless of what your, uh, take is on, on those things, like, that's not, probably not what the Frontier Model developers were wanting from those releases, uh, is that they, like, were completely ignored, um, or that they were sitting, you know, second seat to the other story.

[00:05:14] Jacob Haimes: So really, it was just an excuse to bring the other story back into circulation. Um, l- which, like, is what... Th- that is the only sort of, uh, ar- the only sort of articles that I've seen about those seem to be like, "Hey, they did this, so let's talk about this thing that just happened," which, like, is using a different system.

[00:05:38] Jacob Haimes: Uh, so I don't know. It just feels they're gonna keep making stuff, and I am I- like, I'm not saying that they're not useful. I'm not, I'm just saying, like, 

[00:05:53] Igor: I, so I feel-- I, I think there's something about like stepping outside of our bubble, right? Like for a large amount of people, like the, like the ability to like hand over more stuff completely is actually like what they want and what like the kind of like mundane utility of the AI hype looks like And

[00:06:15] Jacob Haimes: think they want that, but I don't think they actually want that. I guess that would be my claim

[00:06:20] Igor: Sure, but like, um, I think The fact that some people and like really like only some people seem to be really into Fable and like it's mainly people who set up like these like long loops on like well-defined problems and just like have it like crunch, crunch, crunch, crunch, crunch. And there's really a lot more of them who just don't get anywhere.

[00:06:45] Igor: But then we get these like little nuggets of like, oh, this math proof, uh, worked out, this math proof worked out. And usually they have some form of connection to the, uh, the big labs. But sometimes it's also just like randos, like some mathematicians have, have, uh, have done this. So I think that's interesting that like We are already seeing kind of like, again, like more like of a niching and that like become-- don't become more useful uniformly.

[00:07:13] Igor: They, they bec- they become useful for what they're being probably like trained on or what is currently like on the scaling frontier

[00:07:22] Jacob Haimes: Right. But like, that's just, that's just the jagged frontier sort of argument, which is, uh, like the specific thing that I'm referencing is, uh, Ethan Mollick wrote some article about how he-- like a while ago, about how, um, model capabilities are very, uh, unpredictable. And so, you know, task A could be very similar to task B, but it's still, you know, meaningfully different.

[00:07:50] Jacob Haimes: Uh, and like, it could just be that the, the kinds of text you're sorting is different, right? So you're, you're sorting it for the same thing, but it's different kind of text is just like an example, and the models perform drastically worse on task B than they do on task A. Uh, and so it's like a jagged frontier because it's uncertain how well the model will actually perform.

[00:08:12] Jacob Haimes: Um, and that-- like that's, that, that's not new, right?

[00:08:16] Igor: Yeah, but so like there's a thing also like true things need to be said multiple times because there's many false things but only a few true things.

[00:08:24] Jacob Haimes: Okay, sure

[00:08:26] Igor: I think it is worth just noting that like the theory of there is no single general intelligence thing that you push up, but basically like the models get better at the stuff you, you train them on and you, you do need to train on all of them, uh, uh, to make progress. Like

[00:08:43] Jacob Haimes: Well, you don't need to train on all of them to make progress. You need to train on all of them to make progress on all of them

[00:08:49] Igor: Yeah, um, like in, in a sense kind of like if what we're after is, is like AGI and, uh, and like the stuff that the big corps are like selling,

[00:08:57] Jacob Haimes: Oh, I thought we made... I, I thought they said we made it to AGI. I thought they, they stuck their, uh, flag in the sand. I thought we, we did it

[00:09:07] Igor: I, I, I maybe we've, we're at AGI v2 now. Like, uh, I, I don't think they've made a trillion dollar, uh, profit as of right now, which was like, I think one definition.

[00:09:17] Jacob Haimes: Oh, no. I just think I, uh-- Well, I heard that Altman and, uh, I think Jensen Huang both said something about, uh, Astra being AGI. Um, I don't remember exactly. I'd like... Like I said, I'm, I'm bored of CEO said a thing journalism. I'm bored of we released a new model, pat us on the back, um

[00:09:45] Jacob Haimes: I, yeah. So for- forgive my lackluster, uh, uh, or like lack of attention to detail here

[00:09:58] Igor: It, it was Brockman who said we're in the AGI era, which is probably like distinct from claiming GPT-6 as a, as a, as AGI for, for contractual purposes

[00:10:10] Jacob Haimes: Sure. Well, I mean, you know, we're also in the, uh, what is it? The w- whatever that era is, uh, the current era, right? Is it...

[00:10:23] Igor: 7

[00:10:24] Jacob Haimes: Yeah, we're, we're in the Tropocene era as well, right? Like, that doesn't really tell you that much, so

[00:10:31] Igor: Yeah,

[00:10:31] Jacob Haimes: even know if that's the right one. I just repeated what Igor said. I probably shouldn't do that, but here we are

[00:10:38] Igor: the era of humans b- because we're now causing geologically noticeable, uh, changes to Earth, uh, to the Earth, li- like landfills and shit like this

[00:10:47] Jacob Haimes: Cool

[00:10:48] Igor: We, we, we're, we're gonna put the definition in the show notes f- uh, for the, uh I don't even know what, uh, what, what the field is, but like, uh, I, I guess it's like g- geology? Geologic epoch. Yeah, it's technically it's the g- geology nerds, um, or aspiring ones. Okay. But For me, one thing that I found interesting basically was that Where's this divergence now where like, you know, like you and I, we're kind of bored. A lot of people who can afford it are getting very hyped and they seem to get some utility. But also if you closely read the press releases, they, they keep moving the benchmarks because like they're not getting better at the things that they are better.

[00:11:38] Igor: They're just claiming, "Oh, this is good enough now." And then they kind of like move on to new things. Whereas An gram, that's maybe something worth noting, like the, the AI detector now have like really, really good F1 scores in detecting things like very low false positive, very low f- false negative. Cantrill did a piece about this and I, I, and I can tell why because like it's a meme at this point, right?

[00:12:09] Igor: Like L- uh, LLM speak of like you are absolutely right, this is the load-bearing seam. Uh, like Claude direct is a meme. is less bad, but you can still kind of like spot it if you use it a lot And I don't know about you, but like I, I thought we, we were supposed to at least pass the Tu- uh, the Turing test a few years ago. You know, kind of like can't tell the, uh, the L- the AI LLM from a human and

[00:12:42] Jacob Haimes: Sure. Well, I mean, l- like, I guess exactly how you... Like, 'cause if, uh, if you were, uh, opposed to your position here, then you would say, "Oh, well, like we did," but it's just that's the style that they are using now, right? So, like, if you actually did the prompting right, then you could get the, to, to do the... Like I, I'm not tr- I'm, I don't agree with this.

[00:13:06] Jacob Haimes: Like I understand what you're saying, but I'm, I'm just trying to say like, uh, yeah. Trying to play devil's advocate a bit here. Like the performance on ARC-AGI 3 or whatever did increase, and that was one of the, uh, tasks which you have previously said is like better. Um, now they did obviously like train on exactly those kinds of puzzles, but like, I don't know, maybe that means something

[00:13:37] Igor: yes, that's, uh, that's a thing, but I think it's still just worth like trying to keep the, the reality in, uh, in, uh, in mind that like the current bubble, the current valuations, it really hinges on like this is AGI, it's superhuman in so many things now already. Uh, it can replace all of these knowledge workers and so on and so on. And we might be talking about this in like a fu- future episode about the replace knowledge workers. But

[00:14:12] Jacob Haimes: I mean, like, fucking do it then. I don't know. Like

[00:14:16] Igor: Like, like, like, like it's not there. like I think being precise about why it's not there and there is being like a, like a hater is still im- uh, important to keep track of this. But like, if it was actually as superhuman as good, they would be showing off like advances on the old benchmarks, not on the n- uh, on the new benchmarks. But they can't do that because like what they would have to do is like make more public, "Hey, the old benchmark was bad actually, and we were saturated like, like a long time ago and it's not really meaningfully measuring human performance." Instead, they make the new benchmark and they say, "Oh, this is now even more challenging." And at this point, who even cares anymore? They just make a new benchmark, a new number, and it's like whatever they train on or whatever, like they're, they're targeting right now 'cause they want to show progress and Noting that is important because when people say, "Oh, but it's so good now," you need to be able to say precise, "Okay, but like what is the, the bait and switch? What is the, very cleverly done, like re- rhetorical lie?" And like the, the rhetorical lies, like you can only hand over to the top-tier model this like thing that you put most things on if your task is like in this very narrow field that is already kind of like highly automated, and only because they have already trained on it, and only if what you care about is just hitting your, your benchmark number and not any of the details, and it can be, it can be formally verified.

[00:15:56] Igor: All of these things we have been saying already. This is stable of la- of the last couple of years. And now also from my previous job, I, I switched jobs, uh, uh, by the way. Yay. Um, like there's a lot of engineering effort right now in making more stuff look like that. You know, to like try and squeeze out utility. But this is not the story that was originally being sold. The, the story that is being sold to investors is not, we will make a tool that if you put enormous amounts of engineering effort into it, you, you, you can, uh, squeeze out continued productivity gains of up to 7% or what- whatever set space. I think like we don't have good numbers on this yet. The story was this will replace humans. This is superhuman. And the thing is, even in fields where it has devastated, like it, it being like generative AI employment opportunities, like copywriting for example, where it is kind of like massively polarized things and like a lot of kind of like pay the rent gigs have dried up. You can still tell, like it's not like, uh, like it replaces humans and it's like better now. It's like people, we hate it, we just don't care, and like press releases being LLM'd is a meme now. Like... And I think it's just worth noting that as like one aspect. Um, the other interesting aspect I did find is kind of like As someone who for professional reasons uses LLMs a lot Astra and overall Codex, the last two Codex versions For what they are with all of the, um, ethical caveats and all of the correct criticisms They're kind of in-incredible in the field where like Anthropic still has like the gold bar, but if you put some kind of like workflow into it and you, you treat it as a tool, man, like Luna like the, the dumbest model of the previous generation, but on max effort is just as good, if not better for my workflows as Fable, like anecdotally, and I've heard this a lot, and the artificial analysis benchmark backs it up as well.

[00:18:11] Igor: There's like a, a Pareto point basically of like it's the best cheap model at that check

[00:18:18] Jacob Haimes: Okay

[00:18:19] Igor: And DeepSeek and Kimi and obvious ch- ch- ch- training models, they're also catching up, which means that like The tightening of a band between the cheapest and the most expensive model And the failure to like spark joy basically is also kind of like something worth tracking that like if even if now things keep getting better and all of the utility arrives and so on, the The investment story of capturing this value is still waiting to materialize, and I can't wait for the IPOs that are rumored to be coming, and also for SpaceX quarterly, uh, reports because then we will ac-actually be able to assess these stories that you bought by now and look at CEO S-SX talks against, like, hard numbers

[00:19:16] Jacob Haimes: Well, I mean, uh

[00:19:18] Jacob Haimes: That would mean that I'd have to believe that the free market does a good job at assessing the value of something, which like

[00:19:30] Igor: I don't care about a free, free market, but like, um, quarterly reports that the CEO can go to jail on if they lie and, uh, uh, accepted accountings, uh, uh, standards breakdowns, those

[00:19:44] Jacob Haimes: I see what you're saying. You're not saying... I, I thought you just meant like the IPO, like when, when it, you know, happens, but you're not talking about when it happens, you're talking about a month or, or whatever, the, the quarter after it happens

[00:19:59] Igor: Yeah, like we, we will ha- we'll have information. We actually have, um, information that is not just like annualized rate of return, blah, blah, blah. They will make up all these numbers as well, but then there will be the actual numbers that they have to report, and like they will try to play tricks with that.

[00:20:17] Igor: Like right now, a lot of, uh, criticism isn't that like they are hiding obligations by basically contracts that require them to, uh, to be, to pay on, on delivery, on self-delivery, and there's like minimum usage requirements. But th- this does not show up as liabilities yet on, on the balance sheets. So the invested and the cost of capital and all, all of this stuff looks better it is right now than it should, even though it doesn't look that good even

[00:20:50] Jacob Haimes: Well, they're just, they're just doing that to, to make themselves look better for the initial offering, right?

[00:20:58] Igor: No, this is, um, also the current hyperscalers. So kind of like the people that are doing reports right now, they are not reporting like we, we have committed to paying NVIDIA or whatever data center provider has processed, compute this many bill- billions, then this shows up as a, as a liability because it has-- it's not yet a liability it becomes a liability when data center is up and basically it t- it converts into a contract And then This is basically some accounting sleight of hand where like I d- I didn't know this. I am following my nerdy edutainment YouTubers and that dig into this stuff and point it out as shady. But it's something to like, we can find out and f- like get like corrected numbers. Whereas right now all we can rely on is like leaks, and leaks are not trustworthy And a lot of, uh, story and a lot of like, you know, like of nuanced criticism lies on, okay, what is the actual cost to providing these things?

[00:22:08] Igor: What is the actual numbers and if we see that, like, the real cost to provide these models implies 10X the price, then we'll know. If we see that actually they are profitable on inference as Dario claimed in a bunch of appearances, then

[00:22:28] Jacob Haimes: Well, so I guess one thing that I would wanna flag about that and, uh, you know, I'll pre-register my opinion here, Based on the, the one, uh, study, or not study, but like release from Google that we have that does that kind of thing, I believe it was like looking at, oh, what does, what does inference actually cost in terms of like energy usage and water and et cetera.

[00:22:52] Jacob Haimes: Um, and not like theoretical, but like actual, you know, what does this look like? Um, and if you dig into the, the values that they report there, um, you know, they say, "Oh, it, it only takes like, what, uh, five drops of water, um, per individual query," uh, right? And the thing is that isn't characteristic of the, all of the, like, queries and, and questions that are being handed, because as we said in, I think the, the water, the, the data center and, and water uses episode, like, one query is not typically how people are using this, and it does, you know, rack up additional cost.

[00:23:45] Jacob Haimes: Um, especially if you're doing something like an, uh, a coding session or using, uh, some or other sort of agentic workflow. Uh, but then also on top of that, uh, it's only reporting the water that's used on site, uh, and energy that's used on site, and so it isn't taking into account all of the, uh, uh, water and, and, and other resources that were used externally to, uh, do things.

[00:24:10] Jacob Haimes: And it's also not taking into account the, um, like, training cost and, and how long, um, you know, e-even if you normalize that to, like, training cost over the course of, uh, the time during which that model is actually being used, um, it, it, like, it just becomes very misrepresentative of the actual cost that's in there.

[00:24:36] Jacob Haimes: So I'm, I'm just saying that I'll pre-register that they're gonna be doing that.

[00:24:40] Igor: But right now we can't do any of those checks, whereas if they have to actually report their total spending and broken down in a, um, reasonable way, like things like cost of goods sold or, or, or like the R&D required to de- to develop a, a, a good and the amortization of it, like that's gonna be something v-very valuable.

[00:25:06] Igor: And the last thing I wanted to, to, to also kind of like highlight what is relevant is Astra is pretty good. Like I'm actually like, uh, happily surprised. It's kind of like what Fable promised to be. Um,

[00:25:27] Jacob Haimes: What does that, what does that mean? 'Cause like I, I, I don't know. I haven't used it

[00:25:30] Igor: I just, you know, like bump, noticeable bump in goodness. Kind of like, at least for my workflows, kind of like finishes faster, um, does the thing, if I have something that Luna or the other models I use can't figure out, then Astra is now good. Usage-wise, like ROI seems good always with the, um, like $200, uh, thing obviously. So I, I wouldn't-- I don't think it would be worth paying for, for, for my use cases, and I don't know any like startup who pay the API prices. Um, but, um, OpenAI is really trying to, to like push this with like, um, math results basically. Like they did a bunch of blog posts already about like preliminary math results where like always of this pattern of like it either finds a counterexample via lean exploration, or it finds like ideas from literature and pieces them together in kind of like a reasonably close way.

[00:26:41] Igor: Like no mathematician I've seen so far was really like surprised by results and there was never like a, uh, somewhere there wasn't a precedent in literature already that was close to what the, uh, LLM did. And now they have claimed like a Navier-Stokes solution and the same day or the day after mathematician, uh, basically said, "Hey, I've been using ChatGPT for the last weeks with colleagues to try to formalize this in lean and we were almost done and asked, uh, we were in touch with, with OpenAI about, um, whether they could sponsor us and then this dried up and they're being cagey about whether or not they trained on our chats and basically leaked our results through like last minute, uh, RLHF fine-tuning." And OpenAI so far has kind of like responded, "Oh, this is slander. This is, um, uh, unfortunate misunderstanding we will clarify soon." So this is like a story that is still developing, but with the context of IPO looming, we will see now Anthropic and OpenAI kind of intensify their attempts to show the value for like investors again.

[00:28:03] Igor: So for Anthropic, they did this big thing biology and like- How it can, you know, design drugs I think we talked about this maybe even in a previous episode. And OpenAI, what I've seen on the ground is basically people are happier with the reasoning, and I think they might be trying to lean into that because they also have a lot of like investment into compute and scaling compute. So they have like, I think they have a board or got exclusive licenses with Cerebras, which was like a wafer scale accelerator company. They made their own chip, Jalapeño And they're going all in on scaling out inference basically, which if you want to basically, um, brute force auto-formalize is what you would be doing 

[00:28:56] Jacob Haimes: Okay

[00:28:58] Igor: like If you're trying to understand, okay, is there actually like a big jump in capabilities or like what does this spell out for world economy? I think tracking this and keep, keeping track of like when is there ever something that does not fit this pattern of like it was in the, in the close to frontier region of math. Even if it, if it now was like a millennium, uh, uh, prize, like how much of a leap did they do and can this be explained by scale up the attempts you will hit something eventually? Without, without, without saying anything about the utility, right? It's about kind of like what you can do with these models and what types of problems are actually gonna be affected and how this is gonna like affect the economics, uh, around and have like downstream effects

[00:29:51] Jacob Haimes: Okay

[00:29:53] Igor: I have not captured your interest. You still seem bored

[00:29:56] Jacob Haimes: I Just trying to think why, why is it that I'm not OpenAI defrauds, uh, intellectuals that use their platform. Like, yeah, what's new? Um, OpenAI says, "Oh, this is slander. This is, this is bullshit. Uh, w- don't worry, we'll tell you later, uh, what our response is." And, like, they probably will, but, you know, who knows?

[00:30:25] Jacob Haimes: Uh, they'll probably get someone that they know who's a friend of theirs to do analysis on it, uh, and then they, that will, uh, corroborate their story. Um, like, it

[00:30:38] Jacob Haimes: It's turned from

[00:30:39] Jacob Haimes: Something that is

[00:30:41] Jacob Haimes: at least in, in my opinion, like genuinely more interesting into like... And I, I mean, this has been the case for a while, it's just sort of making the point of like, uh, scale compute and, uh, do whatever it takes to, to try to, to make a buck and, uh, you have enough so you won't actually be held accountable.

[00:31:07] Jacob Haimes: And like, cool

[00:31:09] Jacob Haimes: I don't know

[00:31:11] Igor: Okay. Uh, then I guess let's start with you. Like, like, uh, I think you have your registered pre- preregistered. For me, um I think we are seeing su- subtle shifts and The winter pr- is promising to be one way or another like, uh Change in the, in dynamics, I would say

[00:31:34] Jacob Haimes: Why?

[00:31:35] Igor: Because of the IPO rumors for once

[00:31:41] Jacob Haimes: But like, like I, I, I get that, but at the same time, like there were IPO rumors last year at this time, and there were IPO rumors the year before that at this time

[00:31:52] Igor: I don't remember IPO rumors, like concrete ones and like OpenAI also kind of like kind of flat out they filed confidentially. Like, like now there's like a concrete timeline, there's a concrete valuation, there's like discussions around it and like also at a scale at this point and we had one IPO already. So

[00:32:19] Jacob Haimes: Well SpaceX IPO, I, that, that doesn't really... I don't know if that counts

[00:32:24] Igor: You don't go public if, uh, if you don't think it's, like, necessary for the, for the sur- survival of the company, unless you want to cash out, obviously But both these things are signs, and then if they don't go public, then either they need to become profitable or they need to raise more money privately And again, both these things would be science The data center build out staggering/will need to deliver at some point this winter as well. At least like the first ones. So one way or another, it's either gonna like kick into the next mode it's gonna start to deflate. But like the current storyline can't really continue without some Some adjustment, I would say

[00:33:21] Jacob Haimes: Okay.

[00:33:23] Igor: Would

[00:33:23] Jacob Haimes: What about, what about that, uh... What was the thing that I saw something about recently that OpenAI is doing, like at inference time optimization or something like that?

[00:33:42] Igor: Best Time Training,

[00:33:44] Jacob Haimes: Yeah. W- i-is that enough?

[00:33:47] Igor: Um, the big thing with Astra is that they're doing looped transformers. Um, so

[00:33:54] Jacob Haimes: are back?

[00:33:55] Igor: kinda, but, um

[00:33:57] Jacob Haimes: Let's go. Sorry, I just think that's funny

[00:34:01] Igor: Like also technol- technol- technology-wise, like this actually is some interesting shift because like actually makes some people worried about like, because like if you do the loop transformer, like multiple passes in the latent state and less what they call chain of thought. So like the idea of monitorability is decayed

[00:34:22] Jacob Haimes: Right. Yeah. It's, it's essentially like all of the, all of the things that, uh, people get worried about, about like, uh, internal optimization that isn't, uh, monitorable. It, like, makes all those worse basically. Um, and like, yeah, I think it does

[00:34:41] Igor: The flip side of this is that basically that, um Monitorability is what you need for debugging, not for safety. Safety is, uh, is something you design into a system

[00:34:52] Jacob Haimes: Y- y- yeah. I agree. It's just l- I don't... Like, the, the idea that

[00:35:00] Jacob Haimes: People are going to design safety into these systems is so far away from where I'm at right now. Uh, like

[00:35:10] Igor: What I mean is that if, if they are letting go of this easy debugging hook, it means they're really squeezing also how to keep going

[00:35:18] Igor: And it is at the same time interesting how you can actually, you know, like use these things. Like loo transformers have been I think, since the first year of a transformer basically. Like Universal Transformer was a paper, I think. And they are interesting because like they are basically really chunky RNNs and, and they, they are Turing complete, uh, other than transformers. They're just a bitch to train. So How far they can push it is actually interesting. And also like if the big, big guys show that it's feasible, smaller LLMs imitating them are gonna be interesting as well. So even if like the Hater narrative, like my-- the, the IPO and the cost being sh- uh, being thr- uh, being like the change breaks down, technologically there's gonna be like diversification.

[00:36:23] Igor: We're go- we're, we're gonna see like more and more kind of like commodity-oriented things coming out trying to like find their niche And we are still in a high, uh, interest rate regime, right? So

[00:36:42] Jacob Haimes: I know I should know that, but like I just... There's so much. I can't keep it all in my head

[00:36:47] Igor: Like, you asked me why do I think there's gonna be like a regime shift, like that's one of them.

[00:36:53] Jacob Haimes: Okay

[00:36:54] Igor: Also, of re- a regime shift, um, midterms

[00:37:05] Jacob Haimes: I mean, yeah, that's sort of a... Yeah. It su- su- sure would be good if, uh

[00:37:12] Jacob Haimes: The, the US, uh, legislature, uh, you know, was not in its current state, but, um

[00:37:20] Igor: I mean, what I mean is also that like right now, I think amongst other things like the d- the data center thing has become a bit of a, like rallying cry where like the Texas governor was using it to grandstand to score some points because people just really hate data centers. And There's a potential conflict between, like, Congress and the AI crowd there. And the question is, like, on which side will Trump fall in

[00:37:52] Jacob Haimes: He's already established he's gonna fall in on the AI side

[00:37:56] Igor: Would you bet money on predicting what Trump will do?

[00:37:59] Jacob Haimes: Uh, in th- in this case, I wouldn't bet a lot of money, but I, I would bet some money on it. Uh, now, I think the only way to operationalize that would be to p- like, put money in stock, like AI stocks, which I also don't wanna do. Uh, so I'm not going to . But, like, I, I do actually think that, uh, at least my model for, uh, like, h- how to understand and how to predict what Trump does is, like, uh He is self-centered.

[00:38:31] Jacob Haimes: He is, um, very much interested in, in how he is seen and portrayed. I think that if AI does not succeed, that is very bad for his, uh, how he, how he is perceived long term because it basically all happened while he was still in or while he was in office. Uh, and at this point, it's impossible-- Well, not impossible because, of course, he could still do it, but it's impossible to do, um, uh, like that is in any way, uh, worth putting any amount of stake in to say that it was Biden's fault.

[00:39:14] Jacob Haimes: Um, right, like, like that's, that's sort of collapsed. And so at this point, he needs to make sure that AI succeeds at least as long as he is still in office. Um, and so I think that that is how he will behave. That's my opinion

[00:39:35] Igor: Let's say you're, you're right. Then he gets in, into conflict with Congress, which

[00:39:41] Jacob Haimes: And they, and they, and they fold to him because they always do

[00:39:44] Igor: But it will be conflict and it will be friction and, uh, even just delays can ha- have like an effect. So like let's wrap it up. Um,

[00:39:57] Jacob Haimes: Sure

[00:39:58] Igor: I was pushing for this, uh, episode. Uh, I think it is worth this, uh, this stuff now, even if people are tired, if only to kind of like them at the lies and make predictions like, uh,

[00:40:12] Jacob Haimes: Yeah, no, I, I agree with that

[00:40:14] Jacob Haimes: Well

[00:40:15] Igor: one, folks. I hope you're less tired, uh, and, and Jacob of

[00:40:19] Jacob Haimes: Yeah, I might also just be in a bad mood today. I don't know. But, uh, thanks, uh, thanks for listening. Uh, and that is all the muck that we have, uh, for this week. Uh, yeah, if, if you liked this, uh, let us know, share it around, uh, do whatever. Thanks.