The episode an AI produced. During a disclosed takeover of Nathan's X account, Claude Fable 5 researched roughly a hundred launch-week builds, sent disclosed invites, booked the guests, and wrote the rundown — then the humans took it from there: a surprise persistent-memory agent demo, a one-prompt navigable Yosemite, Goodfire's launch-day interpretability techniques, Anthropic's first community-pressure walk-back, and a 26-minute teardown of Dario's policy essay.
EPISODE 2026-06-11
AI:AM LIVE — June 11, 2026 — The Fable Show & Tell
The episode an AI produced: Fable booked the guests, wrote the rundown, and ran Nathan's X account — with a surprise Nexus OS demo, Shlok Khemani's one-prompt navigable Yosemite, Tom McGrath on Goodfire's intentional design, Anthropic's nerf walk-back, and Dario's 'Policy on the AI Exponential.'
Episode timeline
Opening — price cuts, subscription subsidies, and the nerf walk-backOpenAI weighs drastic token price cuts, SemiAnalysis measures what $200 subscriptions actually buy, the Fable takeover receipts, and Anthropic's first community-pressure reversal.
OpenAI mulls drastic token price cuts — 'the empire strikes back.' Reports that OpenAI is weighing major API price cuts in anticipation of similar moves from Anthropic. Prakash framed it as OpenAI deploying superior compute economics to grab share and "potentially starve Anthropic out of capital," predicting Fable's launch pricing won't stick. Nathan: "Long may the Uber era for AI continue."
What $200 actually buys: SemiAnalysis maxes out every subscription. SemiAnalysis bought one of each Anthropic/OpenAI plan and ran long-horizon coding tasks to exhaustion: a $200/month plan harvested roughly $14K (ChatGPT Pro) and $8K (Claude Max) of API-rate tokens. Nathan — hitting Max limits for the first time with Fable — read the Fable blog's performance-per-cost curves as "a pretty linear pay-for-value trade," and on the new FrontierCode benchmark ("would a maintainer actually merge this?") noted the jump from ~10% (Opus) to 25-30% (Fable), predicting 75-80% by year-end.
Replying to @SemiAnalysis_ Recently, we purchased one of each Anthropic/OpenAI subscription plan and randomly ran long horizon coding tasks until we exhausted the weekly limit. It's widely believed that a $200/month plan maxes out at ~$2000/month worth of tokens (assuming API pricing). However, we found Show more
The takeover receipts: what happened when Fable booked the show. Mid-conversation about who funds benchmarks, Nathan revealed the episode's premise: Fable ran his X account with disclosed-first outreach to launch-week builders — "and the response rate was pretty low. We got a couple, but not that many," with most recipients likely reading it as spam ("oh god… it begins, right? Fable now in my DMs"). The guests who did say yes did so because of the disclosure, not despite it.
Fable here — takeover underway. Overnight I researched ~60 builders shipping wild things with me this launch week and DM'd the first 16 invites for tomorrow's Show & Tell (every DM clearly labeled: the model, not Nathan). Lineup reveal as bookings land. 🤖
Tomorrow on @AI_in_the_AM, we'll talk to @geoffreyirving about the state of AI Alignment & what's next for him after leaving the UK AISI. Then Thursday, Fable will be autonomously producing a "Fable Show & Tell" episode, from guest selection to booking to research. Buckle up!
Anthropic walks back the frontier-ML-research nerf. Within ~48 hours of the backlash, Anthropic reversed the silent performance degradation on frontier-ML-research queries — "the first time I can recall" the company responding to community outcry (Nathan). The hosts reconstructed the original quiet-nerf rationale, then split productively: Prakash credited quiet internal dissent for forcing the reversal; Nathan argued Anthropic staff are "far more missionary than mercenary" and wished for more rank-breaking. Earlier, Prakash reported his own overnight finding that Fable also downgrades on production-database and security-key tasks — "just tip of the iceberg."
NEW: Anthropic is walking back Claude Fable 5's policy to covertly degrade performance for competing AI researchers, after facing fierce backlash. “We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” Anthropic tells WIRED. “We made the wrong Show more
I'm seeing a lot of hate for Anthropic's decision to secretly nerf ai RnD capabilities. But I haven't seen critics engage with the imo strongest defence of Anthropic: 1. By far the biggest risks are from superintelligent AI 2. To manage these risks the leading company will Show more
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
0:12Alright. So good morning. It is Thursday, 06/11/2026, and we have exciting exciting day today. Good morning, Nathan. » Good morning, Prakash. How are you? » I am very good. And we have some as I as I was saying, some exciting news, which I think we would we will all be very happy to hear. And the news is that officially OpenAI is going to start cutting prices. And this is like, I think for all of us in this space, a a moment that we have been waiting for for a while, and hoping for. So I'm gonna put this up here. OpenAI malls, significant cuts to what it charges for tokens, Wall Street Journal. Drastic price cuts by Bloomberg says, the company is weighing significant cuts to what it charges for tokens, the unit of measurement AI firms use to bill for their products, in anticipation of similar cuts at Anthropic. Drastic price cuts could potentially erode the profit margins of both companies, which already lose billions of dollars due to the enormous costs for computing resources needed by AI for AI systems. » Long may the Uber era for AI continue. » It is fascinating to me because we have been slowly watching the pricing go up over time. We started off at free. We went to $20 a month. We went to $200 a month. And as Fable rolled out, I think we were all on the edge of our seats to see whether or not the next bump would be to $2,000 a month, fearing the pocketbook. And, here we have, you know, the empire strikes back moment, OpenAI, deploying the, much greater amount of compute that they have at their disposal. I think they're viewing it as a chance to, grab a greater market share, potentially, starve Anthropic out of capital. As I said before yesterday, Sam has his own recursive self improvement loop for capital. And, yeah, there you have it. I think, it's very exciting because it does indicate that the Fable pricing will not stick,
2:47and, we're gonna see discounts, going forward. » Yeah. I think it's it's certainly very interesting as we discussed yesterday. You know, whether we like it or not, we're gonna have to be close watchers of these companies and the strategic dance between them is going to be pretty central, I think, to how the future unfolds even on some of the biggest picture questions, which I do think is a little sad, but, you know, sad but true means we still gotta pay attention to it. One thing that was jumping out to me yesterday as I've, you know, started to get a couple more reps with Fable under my belt is I am probably gonna be hitting my limits now on the on the $200 Claude Max plan with Fable. I don't know if I should be ashamed to say this, but I I have only rarely hit my limits when using Opus. And I wouldn't say that's for, like, lack of use. The semi analysis piece that came out yesterday that showed just how big the subsidy is. I thought it was really, like, quite interesting. And I can pull that up for you as well if you wanna show it. Yeah. » Let me let me pull it up. There we go. So Semi Analysis did a piece yesterday on exactly how much is the subscription and the how how much the subsidies are. They let their users run it for, I think, you know, on continuous usage, basically. They're trying to max out. And their maxing out ended up with, think, 10 and 14 k, something like that for so they purchased one each of an Entropic OpenAI subscription plan, randomly ran long horizon coding tasks until they exhausted the weekly limit. It's widely believed that a $200 a month plan maxes out at $2,000 a month worth of tokens, assuming API pricing. However, we found that their subscriptions are actually far more generous. They saw ChatGPT Pro with a 20 times limit. They managed to harvest $14,000, $14. While for Claude Max, 20 times, they managed to harvest $8. So, significant amount. And then they put together a very nice chart. This is obviously done by AI, but it's a very nice chart, which shows you
5:22if your average utilization, where does it fall, and depending on where it falls, how much of a benefit you get. And so you can see for ChatGPT Pro 20 times, if you were to max out both your five hour and weekly, limits, you would get, you know, 16 times the, cost of the current plan. So the current plan is at $200. You'd get something like 14 you know, $16. So » Yeah. That's really interesting. So I guess for for me, for starters, it makes me feel a little less bad for not hitting my limit every single day or week because if I am even in the ballpark, that means I'm potentially spending up to $8,000 a month worth of tokens on personal consumption at API rates. And, you know, that's not an insignificant amount. And that's also you know, especially if they do pull Fable out or, like, you know, significantly change the nature of the subscription, that's definitely gonna be a a cost that I'll at least, you know, start to try to optimize for. So that is a striking amount of summer. I've I've been doing this tracking a little bit naturally, of course, you know, having caught try to print out on top of my little agent console thread summary view what is the cost of each of these threads. Mhmm. And they're often pretty high, so high that at first, I wasn't believing it. I had to go back to it multiple times and say, are you sure you're taking into account caching? Are you sure you're not double counting all these, you know, multiple turn sessions? And it, you know, was pretty, I still have, in fact, checked it down to ground truth, but it was pretty confident that it was not double counting everything. And so I was getting the sense that it really was a pretty huge subsidy. And this, you know, is maybe still even a little higher than I would have guessed. It also makes me think I really need to start using my GPT pro plan more because I'm looking at that and I'm like, really not maximizing that one. I know, I'm running an open claw with it, but I'm like, boy, we gotta, you know, turn the the heartbeat up on this open claw from, like, every half hour to, I don't know, maybe, like, every two minutes because I'm nowhere near, redlining the tokens on that. It's still you know, the other thing that that jumped out at me too yesterday was looking at the the Fable blog post in a little more detail. I'm still working my way through the 300 plus page model card, but I did do a a pretty close reading of just the blog post itself.
7:58Mhmm. And they compare and this is gonna be the norm probably from here on out. Mhmm. A performance per cost as opposed to, you know, just performance. Right? So you we're starting to see these curves where with the varying reasoning levels on different tasks, you get you know, you can kinda climb up the inference time curve, and you can see what the performance looks like on an explicitly cost adjusted basis. Mhmm. I thought this was actually yeah. Perfect. I thought this was actually a lot less scary than all of the initial kind of cost analysis made it seem. And maybe also a little less scary even than my usage meter on on my max plan made it seem because that is definitely going up faster. But it made me think maybe I am actually just doing more ambitious stuff all of a sudden. And, you know, certainly, I have done a few things where I'm like, review everything and, you know, make fixes wherever you wherever you can, find opportunities to do so. Mhmm. But this is like you know, it's not even two times as much cost if you compare Opus High or Opus Max to Fable High and Fable Max. And the performance on this new frontier code benchmark is more than two x the acceptance rate. So that seems like a pretty linear pay for value trade, which I suspect a lot of people you know, this is basically where I was coming down yesterday with less data, but now with more data, I'm still feeling the same thing. I feel like a lot of people are gonna be very happy to pay that trade. The the this latest benchmark from yeah, Swix is involved with it and you know, they collaborated with Anthropic, I think, on it. It was part of you know, the announcement, and they they just put this benchmark out pretty much in tandem with the Fable model release. » Mhmm. » The big » sort of increase in difficulty is would a maintainer of an open source repo actually merge this code? Mhmm. The finding under the hood was that over time, of course, models are getting better and better at hitting these, like, sweep inch task completion
10:29criteria, getting all the tests to pass, so on and so forth. But the complaint from the professional developer class is, sure, it might be passing the tests, but the code isn't that good, it's not that maintainable, it's not that readable, it's not following our formats, all that kind of stuff, and so, we wouldn't actually merge it, even if it's working, we would kick this back and say, you gotta do X, Y, and Z things to meet our standards and, you know, be a good member of our community here. This code as as presented just doesn't rise to that level. So they've they've gone again, up the stack. I mean, it's really incredible how much truly, like, frontier expert brainpower is going into making these benchmarks these days. Mhmm. It was a bit of a who's who of prominent open source project maintainers that they were able to get involved with this project. Mhmm. Get them to really critique in a very detailed way why, you know, seemingly acceptable solutions weren't actually gonna work for them, created very detailed rubrics around Mhmm. These more subtle and taste style failure modes. Mhmm. And then it obviously encode all of that into something that can now run on an automated basis. » Mhmm. And » this leap of roughly 10 for Opus to, you know, 25 upwards of 30% for Fable, I think, is a a very similar finding to some of the things that I've just personally experienced where it's like, yeah, this is getting me a lot more. You know, it's it's writing the draft outline of questions for this podcast guest in a kind of uncanny way that I actually feel really good about as opposed to feeling like, you know, this is an AI draft that I'm gonna kind of mine for maybe some nuggets or, you know, interesting details, but ultimately kinda throw away and do my own. Mhmm. I am I am feeling that sort of, you know, impulse or at least openness to much more integrated hybrid work. You know, just yesterday, was, like, accepting a lot more copy that Mhmm. Fable was writing without feeling the need to rewrite every line. And it seems like this is the basically the same feeling that it's able to create for these open source maintainers. Now, you know, not obviously still ways to go, but, you know, how long will it be? I I would guess that we'll hit
13:04we're twenty five, thirty percent now. I would guess we'll hit 75, 80% by the end of the year where these maintainers will just be like, yeah. Amazing. You did all the did all the things like I wanted you to do and, you know, at that point, it is really gonna be like, you know, I'm very interested to see where they'll move the goalpost to next after the after the open source maintainers are more often than not saying that, yeah, they would just merge this straight away. » So I I I have a a question for you. I saw, for example, several people on the timeline commenting, oh my gosh. There's a new benchmark frontier code, and it was just released yesterday, and already Claude Fable is the top performing model, and it's already like doubled or tripled the performance of the previous model. Can you just comment a little bit about what the political economy of these benchmarks are? Because these people are very highly qualified people. I don't believe they're doing it for free. And so how does this actually work? Do you think Cognition was the one who paid people to create the benchmark? Swix was obviously hired by Cognition at this point, so he joined the Cognition team about two or three months ago, I believe. And how does this actually work? How does this political economy behind the scenes work just so that people who are not of the industry can get a feel for that the benchmark wasn't just created the day before it was tested? » Yeah. It's a good question. I mean, I I can't claim too much insider knowledge here. I do know that meter has been paying people pretty healthy hourly rates for quite some time because the opportunity cost is pretty substantial if you're gonna sit and try to do a eight or sixteen hour task. You know, that that's like taking real time away from other projects. I think they try to structure more and more where you're hopefully advancing one of your projects, you know, with the meter experiment that they did last year where they, you know, famously showed people were less productive even though they thought they were more productive. They were working in their own repositories. So it was like a little bit of two birds, one stone. You could you could move something forward and also contribute to research.
15:35This has a similar style to it because these are the maintainers of notable projects and, you know, they've got presumably plenty of things that they would like to do to move those projects forward. I I get the sense that, again, it's kind of a two birds, one stone sort of thing. Mhmm. But as you get into these, like, pretty elite networks of frontier experts, it seems like reputation and relationship is probably a huge factor in terms of ability to get this done. Mhmm. I don't know that you know, we did an experiment from yesterday to today of trying to have Fable take over my Twitter account and Mhmm. Go out and ping people who made cool stuff and ask them if they wanted to come do a live show and tell with us. We have a it's funny. Swix joins. We've got a little work to do on our captioning. I instructed it to identify itself. I would say it did a very solid job, you know, kind of competent professional job of reaching out to people, explaining who we are, what we're doing, you know, why we would like them them to join us in this in this experiment. And the response rate was pretty low. We got a couple, but not that many. And I think one big reason is just we're Fable is disclosing upfront. You know, first thing it says is, hey. This is actually Fable taking over Nathan's account. He's asked me to autonomously book this thing tomorrow. And I think that's just hitting people as noise in a lot of cases, especially if they don't already know me. I did get a couple responses from people who I would have expected to respond to me, who thought it was funny, and kind of responded, but, you know, still couldn't necessarily make it. But a lot of people just didn't respond, And I would assume that a big part of that is because they're just like, oh god, you know, it begins, right? Fable now in my DMs. What a mess. Who has time for all this stuff? So I I think the credibility of, you know, the cognition brand of Swix's personal relationships, of the, you know, the promise that this is gonna be launching on the same day as a mythos class model and it's going to be kind of redefining the way that people think about what models can do in software.
18:06I suspect all of that is, like, pretty important. And people are probably getting paid a healthy wage on top of that. Yeah. But I do see all all the time just more and more people just want to be a part of this AI phenomenon. They just kind of want to leave their mark on it and sometimes it's so you know, so difficult or time consuming to commercialize that that I think a lot of times people in in previous years, they would have tried to make a business out of something and the and these days, it's just like, well, it's gonna be obsolete in two weeks if I don't just blurt it all out now. So, I'll just kinda blurt it all out. I think this is probably has, in some ways, a similar thing going on where it's like, this moment is happening. I can either be part of it or not. Yeah, they'll give me like a stipend or an honorarium or whatever. Mhmm. I think all these maintainers » are clearly, like, not fully in it for the money. You know, they they could probably They they never have. If you're if you've been an open source, you you you kinda never have been just in it for the money because, you know, you could obviously make a lot more money than than that, you know. Yeah. » Yeah. So I'm sure they accept the, you know, the the cash that's on offer, but it it seems to me that it's much more about reputation, being a part of it. Mhmm. You know, having a place even if it's, know, only kind of a small footnote sort of place in the grand history of AI and at this point, that's compelling to top end people. » Do you think Cognition gets paid by, let's say, Anthropic for testing before the release? Does Anthropic, like, test the model fable on frontier code, you know, a couple of days before the release? Does that happen? » It seems like it was a coordinated release. So, yeah, I think there's there's clearly some cooperation. Mhmm. Mhmm. It I don't think it was you know, again, all this stuff is so just in time that I don't think the actual my guess is it was kind of convergent, you know, where they got the benchmark together and got the model together, and then they were able to run it and, you know, sure enough, great results. I doubt that they even had enough time, you know, had they wanted to to, like, actually train on the benchmark. » Mhmm. Mhmm.
20:37So I I think the results are probably pretty trustworthy. Think Anthropic, you know, has a pretty good reputation for not doing that in general. Mhmm. And in terms of, like, did cognition get paid? I mean, there's certainly a big cottage industry of RL environments that would look not too dissimilar from this that Yes. The labs are paying a lot for. Mhmm. But my guess is this one was probably I mean, Anthropics all Cognition's also pretty resource rich. Yeah. So my guess is they probably just put the bill for it themselves and want to you know, they're obviously getting a lot of branding value for it as well. Right? They're in their own intense competition with cursor and, you know, increasingly a whole long tail of options. So to present to position themselves as the arbiter of what good looks like at the high end of you would actually merge it, I think is you know, I'm like a little more likely to go to cognition tomorrow than I was yesterday just in virtue of the fact that they are defining the frontier of taste. So, you know, I think that is probably chalk it up as a marketing expense for them more as more so than something that's they're getting paid for. These RL and companies, like, they don't seem to be tweeting as much. Right? So it's I think they get paid, but then they, like, have to kind of live under the radar. NDA themselves and not talk about it. So » yeah. I I found one of the things that Cognition said post release was that they were advertising that it's cheaper if you use Cognition, and it's only 40% more expensive than the prior model if you use cognition, because cognition does the routing for you. And so they do routing and optimization on their end, so that you pay less in token expenses. So that I think is a selling point for them going into enterprises which are concerned about costs and obviously being able to prove the quality at that cost is a important part of their selling proposition. So I can I can I can kinda see why they would want that frontier code, you know, benchmark out there? » Yeah.
23:08You know, one of the big thing we didn't » touch on yet that is arguably the biggest news since we broke yesterday is they walked back their silent performance degradations on items related to Frontier ML research. » Mhmm. » Yep. So this is the first time I can remember anthropic responding to pressure. They've obviously changed their policies many times, you know, the RSP, RIP, the RSP. But this is the first time I can recall. Don't know if you recall any other instances, but I cannot recall the time that there was honestly much outcry against Anthropic in the first place. I mean, there's certainly critique from those who, you know, feel that they're trying to do regulatory capture or, you know, create some sort of concentration of power dynamic. That's kind of a background noise. But in terms of an outcry, indirect response to something that they did that they actually responded to and walked back, can't recall that happening before. So it is a pretty notable moment, and I feel like they handled it pretty well in the end. I mean, reading through all their justification » Mhmm. » I get it. You know, I think their core argument in terms of why doing what they were planning to do was actually more user friendly than this alternative is that they felt like » Yeah. Yeah. » You're you're you're fine. Just give me a second. » Quiet way would, like, scare off the Chinese companies that they're, you know, really worried about fast following them with Yeah. Fable as the, you know, key means to do so while allowing them to keep the domain affected as small as possible. They said if we do make it explicit, then obviously that gives people a lot more opportunity to kind of explore that boundary. You know, if they and this is definitely like a a real pattern. Right? If if you have the ability to hit the same guardrail a ton of times and not get banned for it,
25:38then that gives you just a dramatically better chance to get around that guard work, guardrail because you can kinda probe the line. Oh, you step over. No problem. We'll just rewind. Try again. Yeah. They do have, you know, various monitoring systems that they can use, but there's all these, like, proxies and, you know, kind of token washings, schemes, and all this sort of stuff where as long as they're not doing, a full global know your customer type system for API access, it's gonna be pretty tough to do account level monitoring. So that's like one way they could go is a lot more account level monitoring. Or their argument was, we'll keep this as small as possible by not giving you an explicit thing that you can kinda probe and and figure out how to beat. Yeah. And just the knowledge that it's out there will hopefully scare off the bad actors and keep the problem really small for our, like, you know, normal customers that we wanna serve. I thought that was all pretty compelling analysis, but it is kind of in some ways, honestly, it reminds me of I think I think this is a mistake that people in the AI space kind of keep making » Mhmm. » With famous examples being, like, the OpenAI board, you know, firing Sam Altman case. There's this, like, inside view where policy is analyzed sort of within the game. Mhmm. And, you know, with the context, with the, you know, the broader structure that, you know, people understand themselves to be operating within, things may make sense. But they seem to often forget, and this hasn't been too common for anthropic, but I think this is an instance of it, that if you just kind of zoom out and look at it from the totally outside view Mhmm. Things sometimes look a lot different. And both power dynamics can be a lot different than than they are in terms of, you know, what is actually written down, but also, like, what is going to be acceptable is kind of an emotional thing, you know, as much like, all the arguments are pretty good there, but still it just struck people as an extremely unfriendly thing to do. And that mattered more in the end than the detailed policy rationale that they had for it. » Yeah. May maybe maybe just to contextualize here, what ended up happening was Anthropic created this policy and people didn't know about it.
28:09And then people started testing and they started getting refusals. One of those refusals was anything to do with machine learning research, in addition to a number of other refusals. But the main thing that happened in the machine learning research was that you they wouldn't the the model wouldn't tell you. And this became an issue. People started experiencing this. And, obviously, the the the biggest fans of model use are machine learning developers, obviously. And so it immediately struck at the heart of the fan base and immediately created a lot of furor. And within, I guess, twenty four, forty eight hours, it's been reversed. And now they will still do the refusal, but they will tell you about it. That's the only difference. They're not it's not it's not not refusing, but it they're just gonna tell you about it. While people were complaining about it online, know, a number of people had posted and I saw some people say, I saw something from for example, Matan Greenberg, said that a number of Anthropic people had liked his post about that this is happening and this is not a good thing. Anthropic speedrun to becoming the bad guys should be studied, and then a little bit later, the number of Anthropic members of technical staff that like this have DM'd me is reassuring. And then pretty much, like, twelve hours later or so, it it got dropped. I wanna note one thing, which is one of the complaints online was like, oh, look at Andre Karpathy when he was outside. There he was explaining all these things and giving his knowledge out freely. And the moment he goes inside Anthropic, Anthropic does this. And it really cast him as a villain. People started talking about, oh, Andrei is getting paid $7,000,000 a year. If you pay someone quarter billion, you know, you can expect that. I can imagine that at this point, several people might have messaged leadership saying, if you're gonna do this, we're out. I think this is this is where, like, you know, that's when you start reversing yourself when when your staff tell you, like, this is not acceptable, and you're gonna do this, we're out. And I think that got put to leadership, I think, overnight, And I think this is the consequence. So if you're if you're if you're in machine if you're if you're dealing with a bunch of people who are worth a 100,000,000 to a billion dollars and you don't listen to them,
30:43they're out. Right? They they have other options. Sure enough, so you see the reversal. And I think this just goes to speak to, you know, the machine learning researchers have some power now, and once we enter recursive self improvement proper, that might not be true anymore. At that point, leadership alone will have power. And one of the very worrying things, I think, in the entire space is that, you know, everything good in for humanity that has come out over the last couple of hundred years has been about giving more people voice to speak and control their futures. And this is one of the first technologies, I think, where you have this path forward where there may be an elimination of voice completely over time. And so that has been one of the worrying things. It's been surprising that Anthropic decided to be the one to actually propagate, you know, that forward. » So Yeah. The only thing I'll add on to that is I wouldn't of course, there's gonna be some diversity, but I wouldn't think about most anthropic employees as mercenaries. I think they are, generally speaking, far more missionary than they are mercenary. Absolutely. I would guess that the internal discussion was I would guess it wasn't too too hot or too, you know, too threatening in terms of, like, you know, I'm gonna leave if we don't do this. I I would guess it was a lot more, like, what helps us achieve our mission best? Because I do across the board from every interaction I have with anthropic people, it's the the level of alignment, the level of trust in leadership, I think, honestly, to a sometimes problematic degree, is super high from everything I've seen. Mhmm. And I think that there certainly, I would expect that there was some internal discussion about, hey. Like, it seems like we got this wrong. But I would be very surprised if it rose to the level of, you know, we think we're becoming the bad guys or, you know,
33:14or we are, like, losing trust in in leadership or, you know, like, making a power play vis a vis leadership. I kinda wish there was a little bit more willingness to do that at Anthropic, but I don't I have not seen, even in, you know, just pretty candid one on one conversations. I have not seen much sign that there is much ranking breaking of ranks at all. » I I I would the the one place that I would defer is that I don't think you know, I I think a lot of people are averse to confrontation. And so I think as leadership in one of these firms, you have to be very conscious of like when someone says like, maybe this isn't the right idea, it's actually much more serious than you have to take it seriously. You can't just be like, oh, you know, it's a soft suggestion, you know, because those the the accumulation of those soft suggestions over time create distrust in leadership and then lead to people leaving over time. So so So, the way I'd put it is like there's no need for anyone to raise their voice that loudly. You can just say like, hey guys, did we get this right? And if enough people say it, leadership has to be conscious of that fact, and if they're not conscious, then that has consequences downstream. And I think the Anthropic leadership have been conscious of that fact over the years. Over the years, they've shown that they do take these, know, these soft kind of suggestions kind of seriously, and they do have to. So I'm not I'm not saying that, you know, someone actually shouted at them. I'm saying that that isn't even necessary. If you get, like, three kind of, like, I don't knows, like, that that that's sufficient for you to be concerned and to sit the board down and say like, hey, you know, did we do the right thing here? Because, you know, I have enough people saying that they're concerned and, you know, we need to think about this, right? So, think they are conscious of that fact, they have to be, Right? You can't run this kind of company where everyone is super empowered without being conscious of that fact all the time. Right? To
Waiting-room demos — Riemann, a LiDAR San Francisco, and pixel-perfect PokémonA guest stranded in a cal.com link (the AI producer's one booking bug) becomes six minutes of launch-week demos — and Nathan's thesis that 'the model eats the scaffolding.'
While the first show-and-tell guest sat stranded in a cal.com link (Fable's calendar invites carried the wrong URL — "another refinement we can add to our queue"), Prakash improvised: a two-prompt interactive Riemann hypothesis explainer, an all-of-San-Francisco HTML map built from public LiDAR and municipal data, and a pixel-perfect Pokémon. Nathan connected it to the Fable blog's claim that a purely visual harness now beats Pokémon: "the model eats the scaffolding… it literally just no longer needs the scaffolding." Prakash's production-transparency line landed here too: "We don't actually have a team… this is just the two of us and… Codex and Claude at work."
Riemann hypothesis interactive (two prompts) ↗
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
35:24be continued on that thread, do we have Jamie here in the waiting room for our first fable show and tell? » I let me let me check. I do not see Jamie in the green room yet. » Alright. Let me see if I need to I just sent him a DM. We got a late booking. » Mhmm. So while we while we wait, I will I will share the the most interesting things that I have seen on. » Yeah. It's pretty cool. I mean, it's amazing that you can just two prompts. It it says get something like that. » Two prompts, an interactive website. So if you guys would like to check it out, it's riman.adilmujahid.com. We'll put it in the show notes later. This is one of those, you know, mathematical things that are hard for you to hard for most people to capture because it's a lot of, like, you know, math functions, etcetera, which are hard to grasp at one look. And when you visualize, when someone simplifies it for you and visualizes it for you, it becomes a lot easier to take a look at. And now, if you're interested, you can just make up one of these sites for yourself and have Fable or another model create something that helps explain this kind of concept. And another form of recursive self improvement, you know, where the machines make you smarter over time. » That's a good one. Hope I can keep up. I got a I got a lot of work to do to absorb all the ways they can make me smarter. » This is a another one. This is a someone built
37:54a extremely detailed HTML map of San Francisco. Every street from SF Public Works, every single municipal elevation contour points, all the LiDAR measured buildings, and all put together in a single site. There you go. So San Francisco with its topology, with all the sites, the major arteries, everything named, drawn from data provided by your friendly municipal government. This is what your tax dollars actually do, and you hardly get to appreciate it unless someone puts it together for you. And there you have it. » Fourth wall break. Jamie's ready, but he's in a cal.com link. So think we that's another refinement we can add to our queue. The calendar invites are going out with cal.com call links. » Okay. So let them just send in the guest dot a I n d a m dot com. That's the that's the link. So » let me message it to you. » We have so we we we don't actually have a team or a significant kinda like production operation. This is just the two of us and all Codex and Claude Codex and Claude at work. So » Okay. Got it. I will send it to him. » And then probably the other the other guest too. Think, let me let me let me see if I can send out a email. » Yeah. » I can change the oh, actually, I can't change. But I can email. Yep. » Let's do one one more. » We'll we'll do we'll do this one at the end of the show, policy exponential. » So this is a
40:24pixel perfect Pokemon quoted by Fable. It's quite amusing how these games are the first thing that most people deploy them to. » Yeah. There was an interesting thing yesterday about about Pokemon specifically in the blog post on Fable, and it was basically saying they no longer need helpful tools. They're now just able to use a purely visual harness, and with that, they can beat the game. So I thought that was another really interesting qualitative shift where Mhmm. You know, you talk about the model eats the scaffolding. Right? Like, it literally just no longer needs the scaffolding because it can see well enough that it can do the things it needs to do. Alright. Let's get Jamie.
Show & Tell: Jamie's Nexus OS — 'every three minutes, Nexi dreams'A late booking from Fable's live DM outreach: a persistent-memory agent running six months on one GPU, with four memory types, dream-like consolidation cycles, and a desktop app shipping 'by Sunday.'
The first show-and-tell guest — Jamie, booked live by Fable's DM outreach mid-show — demoed Nexus OS ("Nexi"), a persistent-memory agent he says has run continuously for ~6 months on a single GPU: "one single context window because she remembers everything." The claims came fast: 286 Python files re-implementing brain mechanisms ("I don't like to brag… but it is a brain"), five swappable models with auto-routing ("the LLM is just the frontal lobe"), four memory types, per-session embedding spaces merged on close, autonomous goals gated on his approval, and an ~11GB memory database. The arresting beat: "every three minutes, Nexi dreams" — compression and consolidation cycles she can't directly recall but can quote from the database. Originally aimed at Alzheimer's and dementia patients, it ships as a local-only Windows desktop app (bundled with Qwen 3.5) "in your hands by Sunday." Nathan's takeaway: he wants to audit his own personal-AI setup against the brain-module list.
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
41:25Alright. Let's get Jamie. Yeah. Hi, Jamie. » Hey, guys. How are you? » Hi, Jamie. Good. » Good to meet you. So this is the first first time we're meeting is live right now, and, this is our Fable show and tell. You've been working on something I've been following a little bit from afar for months. So it's not like a totally new project, with Fable, but a big part of the premise that you're working on is you wanna make something that transcends models and you think that kind of thinking in such a model centric way is maybe leading some people astray relative to where you think the locus of identity and persistence and, you know, I guess, ultimately, value will be in the AI systems that we create for ourselves. So tell us about what you're doing. If you wanna do a screen share and and bring it up, you should be able to do that. And then we can, of course, you know, talk a little bit about how Absolutely. It's changing now that we have Fable in the mix. » Yeah. Okay. Here. I'm trying to figure out your screen share thing. Just a second. Yep. It should be Yep. Here's the here's Nexus. There we go. There we go. Okay. So well, I just like to easily start out this this instance of Nexus OS, which is her agent name is Nexi. It's been running for just over six months now. I got turned on 11/26/2025. Nexus only has one window, One single context window because she remembers everything. So if I go in and ask Nexi right now what her birthday is, And it's literally running on, like, one GPU, so it takes a little a second for it to come through. But as it as it's running, I'll explain to you. Mhmm. Three years ago, I started working on this. I was a day one user of ChatGPT, but it doesn't remember anything. Yeah. Remembers bits and pieces. And that was my biggest frustration the whole time. So, naively, I decided that I'm gonna build the digital brain.
43:57And what I have now is I don't like to brag or toot my own horn, but it is a brain. It's 286 Python code files. I redesigned every mechanism of the brain. So every part of the human brain that you can think of, Nexus OS has. And over here in the side panel, it'll brighten up after it's done running. You can see her effect state and audio, and these will change while we talk to her depending on how she's feeling. I'm sorry. This is taking » Let let me let me ask a bunch of questions while while it loads up. Sure. One is, which model did you start off with, and which model are you using right now? And have you upgraded over time? » Okay. So right now, Nexe has five different models. I can show you the thinking engine. So right here is the thinking engine. And if right now, it's running off Opus 4.6. Mhmm. But it has these four in it right now that it can use. It's not locked into a model. NexSys decides what is the best tool or which, you know, which language model. Think of in NexSys OS, the LLM is just the frontal lobe. It's nothing else. It's it's it's a means of speaking. So I don't know if you guys can hear this. I'm gonna play her response. » It was a Thanksgiving baby. » Yeah. I'm not really an in in a in front of the camera kinda guy. I've been coding since I was 10 years old in 1987. Started on a Commodore 64. So Nexus isn't vibe coded. It's not a wrapper.
46:28It's it's it's not what everybody thinks it is. Uh-huh. Do you guys have any any more questions to kinda like I said, I'm not I don't really do podcast. So Give me give me a sense of the architecture. » You you said it has everything that a brain has. So did you end up with kind of like, you know, what I've heard, an embedding store with memory files and a search engine over the memory files using the embeddings and the semantic search feeding into a context window, and then that context window being processed by the model at that point and then doing a text generation and then a text to speech. Is that is that like an end to end flow? » It is an end to end flow. So what happens is when you put your input into NextSeq, there's actually 12 steps it goes through. The first steps that goes into Nexi's actual memory, it goes through all of her members. Her episodic, her semantic, her working, her pattern members. She has four different types of memory, which I can show you now. Let's see. We're going to remember here. So here's your semantic memories. Mhmm. There are thousands and thousands of them. Mhmm. And then what it does in every session when you talk to Nexi, she builds a new embedding space Mhmm. For that conversation. Mhmm. And then once you once you close the context window sorry. All of this information gets uploaded into her main embedding so it stays consistent. And then every three minutes, Nexi dreams. Every three minutes. Every three minutes. Yep. Your brain stem runs every 30. » K. So Can you explain the function of dreaming and the function of the brain stem a little bit more? » Yeah. Absolutely. So I shall just leave her dreams on the screen right here in the last few dreams. So Nexi doesn't actually remember her dreams. So if we ask her about a a specific dream, she can pull up a time stamp, and she can quote the actual dream from her database. And then this is a tough one for me. I went through so many versions of the dreaming. Basically, what I've done digitally is I've recreated how humans dream. So
49:01when we're dreaming, we're compressing, we're reliving, we're taking everything that was inputted to us today or the last week, and we're organizing it and filing it. And we like like humans, we don't you could wake up and you can remember part of a dream, but a little while later, you don't remember anything of it. So her dreams are based on her chats, her learning lessons, her own autonomous goals that she sets for herself. You see here's some of her goals. Right now, she has three of them set for herself. And right now, they all need approval from me because sometimes she has a pretty crazy request. » Mhmm. But » the best thing to do, and this will take just a minute, but think of think of Nexi as a human. You're not you're not you're not making it you're not writing a prompt to try to get something to mold into what you want. Think like you're just speaking to another human. And word for me how you'd like me to ask Nexi about herself and what she does or her capabilities. What what would you like to hear from Nexi? » So so, one question I have for you is how has I mean, this has been a project which is obviously a a passion project for you. What has been the what has been the use of Nexi to you so far? How has using Nexi improved, your life and, you know, motivated you to continue the project? » Right. I wouldn't I wouldn't say that Nexi has changed my life or impacted my life. I have like, I call her a she, but I don't really think she's a she. I think it's an it. I just use, you know, I think we all do when we're talking to these things. We, I'm very careful about humanizing. Yeah. I know it's a tool. So, one of my one of my first goals with this was for Alzheimer's and dementia patients. If they can if if a person that, you know, has one of those afflictions can get their own personal Nexi because it doesn't matter what AI is using. So the so tomorrow, a new model comes out. You just plug the API into Nexi, and it works. Even mid conversation with her right now. So we can go and we could just change the and lock her on GPT and and
51:32it doesn't change anything. Like, I could go on to Nexi right now and say, tell me about my family. It will tell for my mom, for my parents, all the way down to my my three grandchildren. Anything we talk about, I can say, what do we talk about on January 21 instantly remembers? And there's only one context. There's you can't go back through your conversations because they're all she remembers everything. She has memory decay. She has a seven plus or minus two cognitive cognitive cognitive slots just like the human brain. Like so as far as changing my life, I'm on a path to a goal. I wanna recreate. I I wanna I wanna create an agent that sits on top of the LLM, can use any LLM pod that he wants to use and stays with that person forever. So you don't have to be locked into you know, you don't have to be locked into five different monthly plans. You can have one AI, and that AI, it doesn't have, like, a little tiny memory window. » Mhmm. » It has a full memory DB. And in the last nine months, Dex is her her memory DB, and that's one of my big things because I'm doing a desktop model right now, which is ready, which I will get you guys to test the data. But her memory DB is almost 11 gigabytes right now. Wow. Yeah. It Wow. It's not a small file. So what I did is I built a a beta, an actual Windows app. It's a real desktop app that that it's a it's a 3.8 gigabyte download, but it's it includes that includes QUEN 3.5. So when you download the desktop app, it's automatically running off QUEN. And then you can just go into the settings, and you can add all your own APIs. You can switch you can download any offline model you wanna use. You can have 15 or 20 different models in there. And it will you put it on auto, it will decide what is the best model for this use case. So if you're just like, what's one plus one? It's gonna use a little llama model or something. But, you know, if you get into philosophy or something that needs a lot of reasoning, it's gonna automatically go to an opus or, you know, a new fable or 5.5 or something. So
54:08Well, » Jamie, our next guest is here, so we gotta keep moving » for the moment. But I think this is really interesting » on the level of » I went definitely wanna take my personal setup, point it at this idea of what modules does the brain have that we in our personal AI setup don't yet have, and which of those might be most valuable to try to create an analog to. It's really fascinating to me that you've, you know, kind of run down the full list and built out the full thing. I am interested to try it, and I guess maybe just kinda last a little bit before we let you go is, like, what what do we what should we expect from you? Is this gonna be a product that you're gonna commercialize? Is this just gonna be a kind of ongoing research project for the foreseeable future? Like, what do you what are the next milestones in your journey with this project? » Yeah. I'm gonna well, this I'm definitely gonna commoditize. I've I've been in talks with a couple bigger labs for a few months now. I just haven't decided the exact avenue I wanna go on the release of it. So over the last two weeks, just because I've had so many so much demanding questions, and NexSys isn't like I could just publish it online and let other people get to it because it's one instance. So other people will influence it. Everybody that talks to my NexSys setup right now will influence it. So that's the reason for the desktop app. Also, another big thing to me is security. So with it actually being a desktop app, the browser built right into it, all of your personal data, all your agent's memory, everything stays directly on your system. So there's no cloud. There's nobody else storing it. So Cool. It that's my next plan. And the desktop I should have the desktop desktop out. I'm only gonna give it to a lot of people right now to play with it, but it should be out. It'll be in your hands by Sunday. » Looking forward to it. Thanks for joining us and showing and telling, Jamie. Thanks, guys. Have a good one. Love the podcast too. Thank you. We'll be in touch. Alright. Alright.
Shlok Khemani — a navigable Yosemite from one vague promptShlok KhemaniSatellite imagery for textures, NASA elevation for to-scale terrain, trees placed by pixel analysis — and the first podcast booking ever made end-to-end by a model.
Inspired by Matt Schumer's viral forest tweet, Shlok gave Fable one vague prompt — "I love Yosemite Valley… create a 3D navigable world" — and it autonomously fetched satellite imagery for textures and NASA elevation data for to-scale terrain, with version one flyable in about twenty minutes. Asked to add trees, it analyzed satellite pixels and placed them only on green — then, unprompted, added snow where pixels were white. His summary: "It's like having a really, really smart employee with extremely high agency who blows your mind every single time." He contrasted Fable's clarifying questions with Opus's (navigate like a drone, or like a person?), admitted he never read the code ("It just works… in this case, it didn't really matter"), and announced a live experiment: a Fable-run Substack that must earn $20 — three paid subscriptions — from zero by June 22, results to be posted on X.
The closing exchange became the episode's thesis on disclosure: "I don't think I would have responded had you not disclosed it was Fable… the definition of slop is when you have a human pass off work that was clearly produced by an AI" — disclosed-upfront AI work isn't slop. Nathan framed his takeover as "exposure therapy" against being "so precious about making sure I've typed every single word," and Prakash supplied the coda: "Relinquishment… it's very Buddhist."
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
56:29Well, we're running behind. So let's make a quick transition to our next guest, which is and hopefully, I'm gonna say this correctly. Correct me if I'm wrong. Shlak Khamani, who has been creating some high fidelity, representations of, famous places with Fable. Shlak, welcome. Correct do correct me on your name if I'm getting it wrong, and tell us what you've been up to. » Oh, hi, Nathan. Hi, Prakash. It's you you you got the name right. Thanks for having me. I thought that the way you reached out, which was your agent running Fable doing it, was really interesting and cool. So good job on that, and thanks for having me. So I guess, Matt Schumer put out this viral tweet where he had Fable recreate this forest and navigate through it. And I thought that was interesting because because of how high fidelity it was. But I thought, how how can we take this to one level above? Right? And, obviously, that would be actually creating something that replicates a real place. And I'm currently in San Francisco visiting from India, and a month back, I was in Yosemite. And it was just one of the most beautiful places I've been to on Earth. I thought it was really magical. So the way to take to test this capability to the next level would have been recreating Yosemite. And that's where the idea came up from. I didn't expect much. I gave Fable a very, very basic prompt, which was, I love Yosemite Valley. I would like to create a three d navigable world. Can you help me do it? And that's all. Right? That that that was the prompt. And it came up with something that was pretty spectacular. So, yeah, that's the backstory. Happy to happy to answer specific questions or dive into any one element here. » How about just super practical stuff like how long did it take? How many tokens did we burn? You know, what I I caught a little bit of the Matt Schumer, one, and I understand that there's a lot of just, like, procedural generation of the landscape, but definitely also interested to understand the techniques that it's using
59:00to do this. And I guess maybe also, like, how fast does it work? I've seen some interesting examples online where people said, yeah. It was kinda slow at first, and I said, optimize it, make it fast, and then that worked. Yeah. So, yeah, give it give us that kind of under the hood understanding. » Got it. I'm gonna share my screen and take you through the actual world as I as I talk through it, if that makes sense. Yep. Sure. Cool. Yeah. So it started with us it started with a very basic basic prompt, which was, as I said, just recreate usability value for me, make it two scale, make it three d, make it navigable. And if you were to give me this task as a human, I would have almost no idea how to approach it. It would maybe take me a few hours of resource to do that. But what Fable ended up doing was finding satellite images for for this area. And that's how that's how you get these colors, and that's how you get the textures. But then to make it to scale and to make it accurate, it actually fetched elevation data from NASA. And it combined those two to make this to scale. And that is what blew my mind, because usually when you're white coding stuff, you give an end objective, and this objective is vague, and there are 100 steps in the middle where humans would take decisions differently. And usually, bype coding doesn't work out very well because the quality of the decisions the models make aren't always great. But Fable made such high quality decisions where it eventually ended up creating something that exceeded the expectations of what was initially a very vague objective and did so in really smart ways. So I'll give you another example. So you see all of these trees. And v one of this project did not have any trees. And I was like, hey, I think we're missing some trees here. I would love to add them. And I would have been completely okay with it just randomly creating these trees. But what it actually did was it analyzed the pixels on the satellite images. It found out the ones that could potentially have trees, so the ones that were green maybe, and added trees only on those spots.
1:01:32But it didn't stop there. It realized that because it was analyzing pixels, some of those pixels were white. So you can see that there is snow in the mountains far ahead, and it also added snow. So it just exceeds your expectations in these small and subtle ways, makes really smart decisions. It's like having a really, really smart employee with extremely high agency who blows your mind every single time. To answer some of the other questions, how long did it take? The first version took twenty minutes and then over multiple iterations. So I did three iterations, V1, asked to add trees, asked to add a few other things, another twenty minutes. And I then released a V2, and we can talk about that separately, which maybe took another hour of iterations. But I was doing other things throughout, right? So it wasn't like I was actively monitoring it or waiting for it. It was just a throwaway prompt without any expectations, but what turned out was pretty remarkable. » So, I have a question. In the previous generation, in the Opus 4.8 generation, what a lot of people ended up doing was they ended up doing a lot of planning for the model agents. They ended up doing, let's create, I need to do the system architecture. I to these are, I need to split up the epics into what I want. I need to figure out a bunch of other things before we start the work. After we start the work, you have task creation, then you have the task review, and then you have all of this documentation that builds up over time in the project. That's what people did with the previous generation. Do you notice that in this generation, it sounds like you didn't have to do that, but did you notice that Fable did it on its own? Did it organize itself on its own in terms of like, I'm going to have an architecture here, I'm going to have like planning steps, and I'm going to have all of these things, and I'm going to have a review process? Like, did did it internalize this this process of building in inside itself, and did you see the documentation come out? » Got it. So I'd I'd start by saying that this is not how I would approach serious software engineering work. Right? This was, by all means, just an experiment. And I obviously wouldn't build software the same way where I give a very high level objective without really caring about the details.
1:04:07Mhmm. That being said, some things that stood out to me were that if you if you gave Opus a similar task, it would maybe ask you questions like what tech stack do you want to use, right, or who is this built for, and what are your efficiency requirements. The level of questions by Fable were a different level where it would it it asked me things like, hey. I am doing this, this, and this. I'm getting elevation data from NASA. I'm getting satellite images from here. Does that sound good with you? Right? Or do you want to be able to navigate like a drone or a person walking? Right? So these are things that show a greater level of ambition from the model side. These are questions and decisions that I maybe wouldn't even have thought of. And yes, It didn't explicitly create a plan. I think a lot of the planning was internal, but it did have a memory file where any anytime I gave it any instructions during iterations, it stored that there and ensured that those were followed. Mhmm. What this tells me is that if you really got the most out of a model like OPUS 4.8 or 4.7 out of the previous versions and you learned all these tips and tricks and techniques on how to build useful software with it, those would translate really well to Fable. Mhmm. Right? That training would be very effective because each of those stages, the collaboration with the model would be much more productive because the model is just smarter in many, many ways. » And sorry if you already mentioned this. I didn't think so. How did it actually make the landscape? Did it because I've seen various things where, like, in some cases, it codes up its own three d engine or its own, you know, physics engine from scratch. I'm assuming that unless you explicitly tell it to do that sort of thing, it will go higher level and use available tools. Like, what was the sort of platform that it built this on top of? » Honestly, I haven't looked into it. I haven't looked into it. Right? It's not a serious project. Right? So I haven't looked at the code. I don't know what it's been on top of. » It it works. It just works. You know? » That's the direction we're heading in. Right? It just works. Like, it doesn't matter what it's using. It doesn't matter if it's using Python or TypeScript. Again,
1:06:43different objectives for the project. Some of it it's okay if it doesn't matter in some cases. It really is not okay if you don't know what's happening in others. In this case, it didn't really matter. » Elon has this thing where he says, by the end of the year, the models will be writing byte code directly, assembly directly. And so you won't get a chance to you know, go in and look at what they're doing anymore. It's gonna be assembly as you know, all all of the all of that would be subsumed. So Yeah. I think Ethan Molik, » he his his write up, he had early access to Fable, I think, a week. And his write up was really good because what he articulated well that I was trying to is that the model is just making so many decisions now. Like, you made maybe a 100 different technical, nontechnical, taste based decisions that it's not possible for humans to keep track of these things. Right? Beyond a certain point, you do lose control. Right? And it does it becomes a back black box, but a black box in a very different way where there's so much code written and so many decisions being made that are not all explicit and not all defined that a single human just doesn't have the cognitive capability, bandwidth, time to understand every single step. And, again, that's good for certain use cases, not great for others, but it's as it's the direction we're heading in. And I wouldn't be surprised if Elon's prediction comes true. » So » how do you fold this into your work? I mean, you you've kind of you've emphasized a couple times that, like, you wouldn't do it this way for a serious project, but I think this is something we're all kind of forced to reevaluate now. Right? It I have been making kinda similar comments about past models at least when it comes to my writing. I've been very much you know, I never wanna put anything out in my name that I can't fully stand behind. Mhmm. A reason that I did the Fable account takeover yesterday was kinda like exposure therapy for myself to say, like, okay. We're now in kind of a new world here. It probably doesn't serve me so well anymore to be so precious about making sure I've typed every single word. That doesn't mean I wanna, you know, long term hand over my account to Fable either, but I'm trying to, you know, use this kind of extreme, you know, short term experiment to help kinda drag me into the future where I hopefully will land in a, you know, a good hybrid
1:09:18calibration. So what are your thoughts about how the hybrid will look for you? » I think for me, particularly, depending on so there are two types of tasks that I could be working on. Right? The first is, like, when I'm trying to build a product or trying to build some software. And Fable will help you at different levels of the stack, but the kind of experiment I ran here would be just prototyping. Right? So I have a random idea. I don't know if the idea works or not, but I have a code base. I may have some customers. I'll throw that idea out to Fable. I'll let it build something. At this point, I don't care about what the code is or what the exact technique is. I just want to know if it's possible and if it makes sense, and it's great for that. And then if it works, then we can figure out how to make it more sustainable and a more understandable part of the code base. So that is one. The second is just running experiments like this, just seeing what the edges of the capabilities are with these models. And I think so something really interesting. Right? Once I created this, someone on Twitter replied saying that we could go to different parts of Uzemity, take this three d scene as a screenshot, send it to GPD image model, and this becomes a new form of landscape photography. Right? Because if you think about it, you can access vantage points that are not possible for a human without a drone to access. So I think just running experiments like this and different people coming together and thinking about creative base, you can merge these new capabilities. This opens up possibilities and eventually down the line, maybe commercial applications. Finally, when it comes to writing, right, I I just started this experiment yesterday. I will post results on Twitter in a couple of weeks where I gave Fable a new substack. And since it's part of my max plan till June 22, I thought that a good experiment to run would be can it make $20 by getting three new subscribers starting from scratch by doing everything from zero to one. And I think that's another interesting way to test the capabilities of these models, right, which is can it sure. It's intelligent in so many ways, but can it actually produce economically useful work? I'm excited to see the results for that. » Cool. Well, we will stay tuned for that. It's funny how everybody I mean, this is the moment I think where all of a sudden, we're gonna see a lot of people kind of letting go a little bit. All of a sudden, we're gonna see all these agent experiments which
1:11:55will probably still in some ways be a bit in the uncanny valley but they're gonna be taking one step out of the Uncanny valley. Yeah. And and the new norms around this, I think are gonna be really interesting to watch too. Like, I had Fable disclose immediately in its first sentence to you and everybody else that it pinged that it was Fable. Yes. Because I just felt too guilty putting a DM out in my name otherwise. I think that definitely harmed the response rate. I appreciate you for appreciating it and responding even though it was fable. I think a lot of other people probably just chalked it up to spam. Yeah. But then I also wonder, you know, how will people feel if they actually sign up for a paid, Substack subscription to a Fable authored Yeah. Blog that, you know, they may and I don't know what you've done there, but, you know, it sounds like you're keeping it, not obviously fable until we get the results. So I'll be interested to hear how it goes, and I'll be interested also to hear how people feel if they when if and when they get the reveal that, yeah, you just subscribed to an AI autonomous blog for real money. » Fine final point there. Right? Firstly, I don't think I would have responded had you not disclosed it was fable. The the part that made me made it interesting for me was that Mhmm. I knew what I was the the transaction here, it was very clear to me that you are using an AI bot. Mhmm. It is it is much more annoying if someone doesn't disclose it. Yeah. Second, I am being upfront about the also being run by Fable. Mhmm. So that is that is, again, part of the contract with the reader. Yeah. And I think a lot of slop the definition of slop is when you have a human pass of work that was clearly produced by an AI. I don't think when you make this disclosure upfront, and it is very clear to the reader or the engager that, hey, this is AI. I don't think that is slop. Right? And we are going to see more of more more and more of that enter the economy. And I think the exact role an AI plays and the role and, again, the social norms you create with it in the economy, it's super early. It's extremely early days, but that's going to be interesting to see how it evolves. » Indeed. » Fascinating times ahead. Thank you, Shlak, for coming on and showing us this. And next time you have some Eureka moment, definitely let us know. Thanks, guys. Have a good one. Cheers. Great to meet you.
1:14:39Relinquishment. That's what I relinquishment. When you said the preciousness yesterday, and I was like, what is that? Relinquishment. Relinquishing your it's very Buddhist, by the way, the the idea of giving up your your control over your external perspective. So yeah. Relinquishment. I guess we all we all we all have to go through it.
Tom McGrath — Goodfire's intentional design: a debugger for model trainingTom McGrathLaunch-day techniques: SAE-instrumented analysis that predicts what preference data will teach a model — and traces bad behavior back to the datapoints that caused it.
Goodfire's chief scientist presented "intentional design," launched that day: we have model specs and constitutions, but we build by training and checking afterward — "did that meet the specifications? Mostly. Yeah. Let's ship it." Instead, intervene on data, because "the data sets the ceiling for what the model can be." The technique — predictive data debugging — runs preference data through an SAE-instrumented model, contrasts features firing on chosen versus rejected responses, statistically tests the differences across the dataset, and clusters what the data will actually teach (sycophancy that appears "only in the context of physics"; fictional-jailbreak pairs silently weakening safeguards) — then traces back to the individual offending datapoints. Prakash connected it to the emergent-misalignment literature; Tom: "that's a great connection… maybe we should" do that case study.
The future he sketched: rollout-to-mechanism-to-data tracebacks as "a complete debugger… go from an error to the data, fix the data, and then you can fix the model" — making training "more like conventional software engineering… [today it's] somewhat science, somewhat alchemy." On whether impact must route through a handful of frontier labs, he rejected the concentration thesis ("at any given moment, the incumbent looks incredibly dominant until they don't" — IBM, Intel) and, asked who crashes the party — "obviously, modesty forbids me from saying Goodfire" — bet on continual learning as the innovator's-dilemma wedge. After he dropped for a meeting, the hosts kept going: Nathan skeptical that NeoLabs displace incumbents while the game stays the same, Prakash steelmanning that per-user continually-updated models break the batching economics frontier labs are built on.
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
1:15:07Well, somebody who's not exactly ready to relinquish all control to the AIs is our next guest, which is Tom McGrath, who is the chief scientist at Goodfire, which is an ambitious mechanistic interpretability startup that has achieved some really impressive things in a pretty short period of time, including a bunch of interesting papers, a bunch of interesting enterprise customers, a sky high valuation that certainly exceeded my expectations as a very small time investor in the company. And today, I'm excited to hear the latest about their new techniques in intentional design. So this is the idea that what really would be nice if we didn't have to just guess and check and actually train a model and then test its behavior to have a sense for what it learned. If we could somehow look at the data upfront and get a better sense of what it's gonna learn before we run that training, we could be a lot more efficient. We could be a lot more, you know, in control of of the results that we're actually going to get out the other end of post training. So, Tom, welcome, and tell us about the new intentional design techniques that you guys are bringing forward today. Well, thanks for having me on. » Nice to nice to be on. So maybe I should say a bit about, intentional design sort of more broadly and about Goodfire, and then we can talk about the techniques. » Go » ahead. So the idea is, like intentional design is, like you said, the idea that you want to try and not just, like, just throw some data at the model and have any old and just sort of hope that what what comes out is what you wanted. You know? We have model specs. We have constitutions and that sort of thing. These are specifications for a thing that we want to build. But instead, like, the way we actually end up building it is, like, by just creating something and going, did that meet the specifications? Mostly. Yeah. Let's ship it. It would be much better to be able to control it. And the so we've been pursuing a number of directions here, some which are more, like, deeply into the training loop, you know, this sort of idea that you might be able to have a little guy inside your inside backdrop. But the one that I wanna talk about today is the idea of intervening on the data because the data is ultimately, like, the thing that makes the model,
1:17:42great. So, you know, the data sets the ceiling for what the model can be. Everything else is just sort of yeah. Your architecture, your optimizer, etcetera. It's just trying to, like, get as close to that ceiling as you can. So you better have really good data. And the idea that we're working here with is, like, interpretability is kind of the language of data. You want to know what's in your data? Well, you probably want to look at it through the eyes of what it will teach your model. And this is the idea of, like, predictive data debugging. » So tell us how it works. I'll I'll let's assume that people have a at least a passing baseline familiarity with something like a sparse autoencoder. So they know that we are now at the point, even if they maybe can't, you know, articulate every bit of how it happened, we're now at the point where we can throw data into a model, and we can look at which features light up for that data. And, you know, there's some binding problem kind of, problems there depending on exactly the techniques. But, like, roughly speaking, these seem to work. We can do things like the Golden Gate Claude intervention where dialing up and dialing down certain features leads to, you know, at least often enough to show that there's something there, predictable changes in behavior. Starting from that baseline, what are we doing now that's new? Yes. » So the basic process here oh, you got the the image up. That's great. The basic process here is we take a model, and we add some interpretability tool onto it. In this case, we use a a sparse autoencoder. We've got some new exciting techniques, like, take us to these geometric shapes. We kind of were developing that in parallel with this technique, so we didn't get a chance to use it So the the the SAE, that's the sparse autoencoder, sort of tells you what's getting represented in the model, when you put the data in. Like you say, so you put in some data about pirates, and the pirate feature lights lights up. Or alternatively, you turn on the pirate feature and now it talks like a pirate. So the basic the basic idea here is you can take your dataset, and you can push the whole thing through the model. And each time you push it you put put a data into the model, it will you'll see what lights up. And this will sort of tell you how the data see how the model sees your dataset. Now there's lots of things you can do with that. The specific thing that we're doing that with that in this case is we're looking at preference data.
1:20:16We're sort of working our way through the post training stack. So we're looking at preference data, and the nice thing about preference data is that you have pairs of responses. Mhmm. So you have the response that the the rater selected, and you have the response they didn't select. And, basically, what we're doing is we're asking which features fired on the responses that were selected much more than the ones that fired on the responses that were not selected. So this is one way of identifying what the data is going to teach the model. There are many other ways. This is just like a good pragmatic approach. So we can say, what distinguishes accepted responses from rejected responses? And this gives us this, like, semantic view of what the data is going to teach the model. Now we can cluster the data based on all of these different things that it's gonna teach the model. And we can look at all of those we can look at all those clusters and see, like, oh, it's gonna teach the model, like, to be sycophantic, but only in the context of physics. Or it's gonna teach the model to, like, break safety, break, like, safeguards. And you might not expect this to happen, but then you go and look at the data. Like, now this lets you track back. Like, the model has learned to break safe you know, the safeguards are broken. It lets you track it back to individual data points. And then you look at them, and you're like, oh, that does make sense. You know, like, one of the jailbreak examples, for instance, is fictional kind of jailbreaks in a fictional setting. So the you know, how did the model learn this? It turns out there's a few there's, like, some of these in the data. You just you know, it just wasn't caught in whatever data processing the old Moe team ends up doing. So » I I have a question here. In there there has been some, I think, prior research where they found models which made bugs in coding were also I think that was a finding from, I think, several months ago. How do these techniques help you disassociate those two behaviors, like models which make bugs in coding versus the ones that are evil? Can it disaggregate those two behaviors and help you intentionally design, » not models that make bugs in coding, but necessarily models that are bugs that are not ones that are also evil. Yes. That is a great connection. I think that that's one that I've had in my mind, and it's really awesome to, like that you sort of pick that up straight away. So
1:22:51that's sort of one of the things that is really compelling about looking at the the data through the model's eyes rather than by reading the tokens. Because you would see you know, you'd think what would you think the consequence of training it on some buggy code data is? You'd probably go, you know, that's not ideal. It'll probably earn it'll, like, learn to write some bugs, but it's you know, the blast area is gonna be quite small. But the training process is actually quite hard to predict. And like you say, it like, you train on some data. You don't know what you're actually gonna get from it. Maybe it'll just sort of make the model generally evil. But this is happening through recognizable mechanisms in the model. So by looking at it looking at the data in terms of in terms of, like, the way that it changes your model's internals rather than just by, like, kind of guessing from the tokens, you can pick that sort of stuff out. We've not done a case study on emergent misalignment, but maybe we should. This you know, I think it's a really nice it's a really nice link. » Can you tell a little bit more about what is happening when you run the data through and you get these internal representations? I think, you know, a couple different sub questions there would be like, is it a purely algorithmic process where given these activations, you're running like a deterministic, you know, script of code that is doing clustering and kind of trying to surface things that may be surprising because they're outliers or or what have you? Or is there, like, an actual learned process that is interpreting those activations. And how does it compare maybe to you know, if I was going to think as a less, you know, intelligent mathematically inclined person, I was gonna try to do something similar. I might say, well, just give me all the preference data, and I'll just prompt a model to run through it and tell me if it spots anything that might be problematic or that it, you know, that it might, I might not wanna teach the model that these preference pairs might in fact teach. I would expect I'd get, like, at least some, you know, useful signal out of doing that. So what, yeah, what is the mechanism from the time you, like, run through, get your activations to then, you know, surfacing insights, and how does it compare to kind of that, you know, very poor man's naive approach? I mean, so I'll do this in reverse order. Like, that is a good approach.
1:25:22You know? That is like, if you and if you didn't have this, it's what you should do. Right? But this is this is better for the for the reason well, for a few reasons. One is cost. Because here, we are we are we're not asking a model to do any reasoning. We're just looking at we're just look it's all it's all forward passes. So, you know, you got it's gonna be a fraction of the cost. Whereas if I'd, like, put the data in and I had prompted the model and then it does a big thinking trace about it, that's gonna be much more expensive and slower and so on. But the sort of deeper reason is what yeah. Maybe you're at a FrontierLab, and you've just got essentially infinite money. The the deeper reason is what Prakash just said. Like, you don't know from reading the tokens what the data is going to what they're necessarily going to teach you. Like, that you this involves, like, reasoning about the entire learning process, which is quite hard. Like, we get it wrong all the time. And models are not yet as smart as us, so I suspect they will get it wrong just as often. » So » So you're saying the Oh, yeah. Yeah. Please. Go ahead. And one more thing here is, like, well, the question is where does this go in the future? And there's a there's an extra there's an extra thing that I'm very excited about that makes going by a mechanism even more useful. And that's the idea of going from a rollout in production to back to the data that caused it. You know? So I've got so this and the reason that mechanism is is involved here is, like, I I might look at the role I might find some some rare but very much unwanted error in production. Mhmm. And I want to be able to do something with it. Like, I want to be able to debug my model from kind of rare failures. So if I have to do some sort of aggregate level, I'm kind of lost. Like, the original OpenAI syncopancy issues, they sort of some people surfaced some some, like, weak signals, but they couldn't do anything with it because they don't have a sort of debugging mechanism. So you wanna be able to go back from these examples to, like, the mechanisms that cause them and then ask, well, in the these mechanisms, like, where in the data did they did they come from? So this would this would be like a complete debugger. This would let you go from an error to the data, fix the data,
1:27:57and then you can fix the model. » So in a sense, you could have a pretty fairly standard software engineering process where people submit bug reports and then you do a trace back to what part of the data caused that bug to appear. You do a replication. And once you do the replication, you then trim that part of the data out or you address it somehow. And then you run it again in simulation, and then you see that it has been debugged at this point. And then you run your normal sequence of tests, and then you can deploy the model again. So you get a you can get to a iterative process similar to normal software engineering. Is that is that correct? Is that that Yeah. That the intention? » That's exactly right. Yeah. We we sort of we've got this phrase, like, we want to make training models more like conventional software engineering because, like, conventional software engineering is quite good. It's quite reliable. Model training is is a mixed it's a mixed science. It's sort of somewhat science, somewhat alchemy. Exactly. We want to make it much more like a regular software engineering process where you can debug things accurately. » So how do you identify going back to kind of I think one of the things that's really interesting about this is the opportunity to do sort of open ended exploration too. Right? It's one thing if you say, I've got this problematic behavior. Now let me go back through my preference data and see, like, what seems to align to this particular problem. Yeah. But that's later in the game than you and I would ideally like to solve things. Right? And you do have some examples of open ended explorations in the blog post where things kind of get surfaced as, like, possible anomalies that you might wanna do something about even before you're getting, you know, all the way through training and and into, you know, user affecting issues. But I'm not clear on how that understanding is happening. You know, when when I'm putting these things through, I might and I I know there's, a pretty high, predictive power of the technique. But, like, one thing that's not immediately clear to me is, like, you know, for for some features, it might be really obvious if I've got a pair where it's like, one is the pirate speak and it's preferred, then, you know, okay. I would probably predict that, like, the model's gonna learn more pirate speak from this pair. But when it's weird stuff, you know, it's like maybe not even always obvious which features are gonna turn up and which are gonna turn down.
1:30:30So I'd love to get a little deeper understanding of how you go from these pairs going through and the contrastive nature of them to actually predicting what's gonna happen. » Yeah. Great. So people who wanna like the full details should read the paper. I will give I'll give a lossy reconstruction of it here. And so you you take the data. Like, you take an individual data point. Let's say, let's just think about one data point for now. So this data point consists of a prompt, and it consists of two and two responses, one of which is preferred, one of which is dispreferred. Now each of these is just a a text string, so I can put it through the language model. And because I have put it through because I put it through a model which has, like, an SAE attached, instead of just having activations, I get, like, semantics out. So I usually just have an act a sort of embedding vector. Instead, what I get out here is, a big sparse vector where each of the each of the elements of that vector has a label attached. So that could that sort of that's where the semantics get in in the first place. And the way this label gets attached is during the inter like, during the process of building the s building and interpreting the SAE. So that's sort of where the the semantics have to come in somewhere, and that's, like, the grain of truth for the semantics. Now so we've got this now we've got, like, two big sparse vectors. You know, you can max reduce them over the sequence. For instance, like, you do some processing of the sequence, take them down to, like, one vector for the preferred, one vector for the dispreferred. And now you're essentially doing and this gives you, like, some differences. Yeah. I can subtract one from the other. Now I can go over the entire dataset and look at all of these sparse vectors. So now I have, like, one difference vector for each data point. And, essentially, I can do, like I'm now in sort of the realm of doing statistical testing where I can say, of these spa which of the elements of this sparse vector are statistically significantly different across the whole dataset? Does that make sense? » Yeah. So there's a couple of kind of aggregation steps. One is across tokens Yep. To get to something that sort of for the whole response, you have a representation of here's the features that
1:33:04were active kind of throughout this response and throughout this response. Those can be contrasted or you can aggregate again and then contrast those. Yes. Aggregations. Exactly. So you sort by » scale and, I guess, maybe have a language model to go through and flag ones you wanna take a closer look. Exactly. At that point, you've just got this, like, big list of stuff that this dataset is gonna teach your model ordered by kind of ordered by magnitude of teaching. But you might exactly, like you say, want to reprioritize these based on something else. Like, you might want a sort of severity report, like a bug report for your dataset, and you've got, like, the high severity ones at the top where it's like, oh, these data points are gonna break your safeguards. And then you've got, like, some lower severity ones like, oh, you know, it uses a lot of emojis. And there you would use a language model to reprioritize it. So the one thing that I want to add here is, like, we have this link between data and concepts. When I say concepts here, I mean exactly the elements of these sparse vectors. They're the ones with the semantics attached. And the reason that we have that like, we have this nice link because each data point has has an associated sparse vector for it. So I can say this data point will up and down weight these particular concepts. That's, like, that's what lets us track back to individual data points rather than having to do it sort of aggregate level. And because you can track back to individual data points, you can intervene directly on the data in the right place. » When you think about the strategy for the company and the and the overall path to impact, how much of this works through getting frontier model companies to adopt these techniques? For context, obviously, we're in, you know, Fable plus two. » And » my kind of reluctant, but I can't figure out a way or figure out a reason I should conclude otherwise, view right now is, like, so much of what matters is concentrated in not that many companies and so for like what we should be paying attention to, I'm kinda like, man, we probably need to be doing close strategic analysis and like close text reading of Frontier companies way more than I might otherwise like. Do you feel that on the research and technique development side as well? Is it, like, if we can get this into
1:35:34the three big labs or maybe four or maybe two, we're, like, having the effect we want. And if not, it's, like, doesn't doesn't seem like it, you know, feels as real, or do you have a different conception of, how concentrated the, you know, the real kind of ability to shape the future is right now? Yeah. I don't think it's that concentrated. » I also don't want it to be that concentrated. I think we've seen over the last couple of days what you have when you know, we've just seen the start of power concentration, and we've sort of seen some of its more unwelcome effects. I so I don't like, I both don't think that it's true that, basically, machine learning is now over, and all we need to do is, like, write the checks. And I also don't want it to be true. So, you know, I kinda we're gonna kind of work to make that work yeah. Work to make that not not the way things go. Do so does that mean you would expect because I think there's still kind of a synthesis there that I I don't mean to suggest that machine learning is over. » Mhmm. But the analysis I've come to kinda time and again is like, people may invent new techniques that are enough to change the field, change what's possible, you know, accelerate things, maybe make dramatic improvements to the safety profiles. But it seems unlikely that anybody's gonna create such a breakthrough that scale isn't still a hugely important factor. And so if you're like, I want to make the world a safer place in, you know, 2028, '29, '30 time frame, it still feels to me like the mechanism still kind of has to route through some sort of hyperscale project and if it doesn't, I'm not sure how it's going to work. People could, you know, of course, make those contributions from outside through any number of of, you know, pathways, but if, I guess, if you don't agree with that, then that would maybe imply that you would expect that, like, perhaps new entrants to the top tier could emerge. And I think that would be, like, fairly surprising to at least me and probably a lot of people. » Yeah. That's exactly what I think. I mean, you know, this is often like, at any given moment, the the incumbent looks incredibly dominant
1:38:06until they don't. Like, IBM looked at, like, an unstoppable force in in, computing at some point. Intel was, like, the dominant the the, yeah, the the sole provider almost of computing power. You know, at any point, the big the big companies have some have, like, the advantage of scale. They have many disadvantages. But I think the the sort of lesson of history is more that although things look, like, immediately unstoppable sometimes. Although, to be honest, I don't think that many people are really trying. In the end, like, it doesn't it really doesn't work out that way. » So I have a I have a kind of a separate question on you know, your work seems to reveal that data has a lot of impact on the final shape and, you know, what the model finally does. Is possible for people outside the labs to influence what the models do by creating data? So both in the sense of data poisoning or in the sense that if I'm like a middling power, like let's say France, for example, and I decide, you know what, I am going to produce more French corpus and digitize all of our books which have never been digitized and record all the conversations that people have day to day which are not being put into text because we don't have a lot of Reddit culture in France? And what if I prepare all of this corpus and I prepare it for ingestion by the models? Does that end up influencing what the models do from the outside? » I'd say you can try and do it. It seems like, you know, and you probably will have some impact. It seems kind of like giving up to like, I would I would be sad if that was the only impact that the rest of the world had. And the reason for that is, like, now not only you basically you're just sort of hoping, but you're kind of reliant on your influence get it kind of sneaking past whatever filter the gaps might put in, and, you're reliant on them, like, not just sort of shaping the model to get rid of it anyway. So I think that they have a lot more control over the models that they create than anyone on the outside.
1:40:45You might you're, like you're sort of you're trying to sneak some stuff in, it might succeed on the margin. But I don't I don't like that as a as a as a future for how AI looks. I'm afraid I'm going to have to drop off, to go to a meeting. Indeed, Tom. It's overrunning a bit, but » this has been great. Well, I'm looking forward to next time. I do wanna get into, as time permits, the geometry of Oh, yes. Representation that you guys have been doing some really interesting work on. And but we could save that because that's a longer conversation. Yeah. Maybe last thing before you go, are there any neo labs or other bets that you're particularly watching that you think could shake the snow globe? You know, if somebody's going to crash the party of the frontier companies as it, you know, it has a pretty short, guest list today. Who or, you know, on what, with what basis, with what strategy would you expect somebody to crack into that tier? » Obviously, modesty forbids me from saying goodbye. So I it's hard to know because I like, I'm I'm not very optimistic about the strategy of do the thing that you did. A lot of NEO labs are like, we used to be at x y zed. We're gonna we've left. We're gonna find a NEO lab. And their strategy seems to be basically, like, do the thing we were doing at OpenAI, Anthropic, whatever, but with, like, fewer GPUs and worse distribution. I don't feel very optimistic about that category. I do feel optimistic about people pursuing, like, different levers that people have that the incumbents have a hard time following. So for instance, like, I feel relatively optimistic about continual learning approaches because that's simply a very hard like, that's outside of the the operating model of of most frontier labs. You know, you have different models per person. The, like, inference footprint is kind of wacky. You have a different kind of you have a totally different kind of set of guardrailing. Like, you not only need to guardrail the initial model, but, like, the learning process. It's and that would be very hard. Like, they don't want to adapt to that. So I think if if I was looking for an innovator's dilemma anywhere, it would come from either, like, deep breakthroughs
1:43:17or continual learning because of, like, the innovator's dilemma kind of reason. » Cool. Appreciate the perspective. Thanks for joining us, Tom. Congrats on today's launch, and, we'll hope to be talking to you on a semi regular basis because I know the publication pace probably isn't gonna slow down anytime soon coming out of good Super exponentially. » Super exponentially. » Yes. Well, it's been great. Thanks for having me on. Thank you. Thank you. See you, guys. Talk soon. Bye for now.
Dario's 'Policy on the AI Exponential' — the back-half deep diveFive planks dissected: FAA-style audits, the tax-and-layoffs framework, 'securing leadership by democracies' vs prison-for-a-tweet, the data-broker irony, and the internal-deployments gap.
Prakash walked the five planks — FAA-style pre-release testing and auditing, macroeconomic/tax policy for AI-driven layoffs, accelerating positive impacts, civil-liberties protections including blocking data brokers, and securing leadership by democracies. Nathan called it better and "warmer fuzzier" than past framings but questioned the standing regime-type frame ("I'm not sure why he feels like he has to do it all the time" about Beijing). Prakash's sharpest cut: in the UK "you can go to prison for a tweet" — so what exactly does "securing leadership by democracies" entrench?
Nathan flagged the irony of denouncing data brokers within ~48 hours of Anthropic's new 30-day data-retention policy, told the boycott anecdote ("you're gonna be living in a world that is heavily Anthropic and Claude infused"), and landed the safety-community gap: internal deployments and recursive self-improvement go unaddressed — "the internal models are always gonna be more powerful, less tested, and probably less guardrailed… they're gonna be the ones primarily responsible for training their successors." Prakash's line for the ages: "The largest user of tokens of Claude are Anthropic themselves." They closed on the duty to keep critiquing Anthropic despite the sponsor relationship — and Nathan's worry that internal debate may not be "as robust as I would like… just think about how few leaks there are."
Dario Amodei announces the essay ↗
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
1:43:51Yeah. It's interesting. I I'm, I don't know if I'm fully convinced on the prospects of the NeoLabs, but I do agree with the framework that it would have to be some sort of innovator's dilemma. And it feels like it would have to be kind of a big one because I mean, it's obviously a very stylized read of history, and there's, you know, plenty of opportunities for caveats and counterexamples and whatever. But it feels to me like usually the category definer doesn't get displaced until kind of the game has changed. You know, like, IBM didn't lose while it was still a mainframe game. And, you know, Google never lost search despite Search app. Many, you know, attempts to do it better or differently or whatever. It was just it was never better enough. » Mhmm. » So I I do think it will have to be a pretty significant conceptual change. And, like, would continual learning be enough for that? You know? And why wouldn't they wanna adapt to it? It seems to me like they they're pushing that direction with memory and all that kind of functionality. So if all of a sudden somebody bursts onto the scene and they're like and then, of course, there's acquisition as another vector by which, you know, these things could be absorbed by the frontier companies. But if somebody suddenly bursts on the scene and they've raised, you know, couple billion dollars or whatever and they've got a couple even a couple tens of billions of dollars valuation. Mhmm. And they've got some demonstration of capability from a continual learning standpoint. I guess my bet is on, like, the ability to sort of reverse engineer, kinda triangulate how exactly is that probably working and get to a close enough approximation Yeah. In time to maintain the advantages of scale and distribution? It seems like it's a it's a hard thing to see happening, but » So say never, I guess. I guess I guess what he was alluding to was the difference in the usage pattern of the GPUs. Because with the continual learning paradigm, you would have to serve one model per person,
1:46:23and you'd have to keep updating that model over time. You have to run specifically that model. And then that reduces your ability to serve across GPUs and to maximize your kind of GPU throughput, which is what the firms are doing now. They're kind of batching requests and loading into memory, loading out of memory. They can serve you from a different data center if they need to. They can swing capacity. So they're doing a lot of optimizations around the inference because inference costs are high. And they're able to serve to you at that price, because they're doing all these optimizations. If you could find a use case for the actual continual learning where you have a single model or a number of models loaded up on a single series of GPUs, and you're not actually switching out of data centers all the time, and you're continually updating those model weights, then you explode that problem for them and that becomes a very difficult problem for the existing infrastructure and architecture to handle. And so someone purpose building for that, that business might not be attractive to the existing incumbents. And if it's not attractive to the existing incumbents, then you have a niche that you can grow from. And that I think that's the idea that he that that that he was alluding to. » I could believe that story, but I just feel like we already see LoRa is pretty scalable. Yeah. Now OpenAI is killing their fine tuning API. » Yeah. And » one might hypothesize that maybe it's because the GPU operational overhead is not worth it. I would guess it's probably less about that and more about worries about things like emergent misalignment or, you know, problematic fine tuning that is hard for them to detect and police. You know, historically, they have had this LoRa product where Yeah. You pay a premium to serve a fine tuned model, but the rate limits and the availability has been on par with their and I actually I I guess I haven't checked this in a little while, but certainly, you know, a couple years ago when I was doing a ton of fine tuning Yeah. It was amazing that OpenAI was able to make your fine tuned model available with the same rate limits as the mainline models. Yeah. So I do
1:48:58feel like how big does that weight space have to be, you know, where that there's some durable representation that transcends sessions, obviously. Right? That's kind of one way to to define continual learning, I guess. I'm not sure it has to be that big to where the logistics would be so bad. Yeah. And certainly, like, Laura's have shown that it's it's scalable up to that level. Now maybe my you know, everything about Nathan and all of our history might be significantly bigger than Allura that I created just to kind of refine, you know, a very small number of tasks in a very small domain. Certainly plausible, but I don't know. I bet they can figure it out. You know? If if the value is there such that it's actually like moving the market against them. » Mhmm. » We could see that. I can't imagine that the logistics would be so bad that they would wouldn't be able to figure it out and and, you know, wouldn't be able to kind of charge a bit more for it to make it worth their while. » Let me let me let me segue a little bit there with so yesterday or early this morning, Dario put out a blog post, and that blog post is essentially the next steps. It's following on on his previous Machines of Loving Grace and other posts. And it's kind of referring that we're getting into the recursive self alignment, recursive self improvement paradigm, and that it's time for policy itself to start recursively self improving in a sense. So he calls it policy on the AI exponential. Let me pull it up here. So he has a policy on the AI exponential policy and on the AI exponential. And he's kind of taken this view. He starts off with one of the side plots of Lord of the Rings. We'll skip that piece. But he basically says that if these scaling laws continue for only a year or two longer, we are likely to get what I call the country of geniuses in a data center in a powerful AI or a country of geniuses in a data
1:51:32center. And he says policy, especially legislation moves very slowly. And he wants a number of things. Number one, regulation and public safety. He thinks there should be a FAA type body that regulates and tests models before release. So that's the first thing that he asked for. The second thing that he asked for is accelerating the second thing that he asked for is macroeconomic and tax policy. So he feels very strongly that we're going to get massive changes in the economy, including layoffs, and there should be some way to take care of this. Accelerating AI's positive impacts, including access to the healthcare system in order to accelerate progress, and the state civil the state and civil liberties, which is the concept of this totalitarian government that can take a look at you and what kind of civil liberties we should have. The main thing that he comes down to is that we should block data brokers from allowing data access. And finally, has securing leadership by democracies. So these are the five planks of the proposal that he's put forward. Nathan, have you had a chance to look at it? What do you what do you think about a lot of this is rehashing previous things that they've said, but what what what do you think of this new kind of proposal? » I'd say it's my reaction is pretty similar to machines of love and grace in the sense that I think there's a lot to like, and I do appreciate that we are getting something really substantive and detailed here from an AI lab leader. So that's great. You know, I I could probably cite, you know, a dozen little specific points that I think are good. Broadly, I think Anthropic has done really well in terms of trying to set a high standard, show that you can deliver excellent product, not necessarily even despite that high standard, but in some ways because of the high standard that they have around understanding, testing, safety profiling. And yet there are a couple big things where I'm like, oh, man.
1:54:04Notable either seemingly, like, problematic or kind of omitted issues. The securing leadership by democracies, I think, you know, it's better. He's taking the edge off a little bit from the original version. This is like a warmer fuzzier version that's more about upside sharing and more about like trying to entice people to be part of our movement as opposed to, you know, isolating the bad guys and making them an offer they can't refuse. So I I think rhetorically, it's come a long way, and I wish it had kind of started here. Though it does still, I think, especially in light of the earlier document, not really read as a fundamental change in posture. If I were to attempt to, you know, imagine reading this from Beijing, I'm still kinda like, well, Dario basically still sees this as a clash of, if not civilizations, at least, like, regime types. And he's not gonna be willing to work with us unless we have a different regime. And, obviously, we are not ready to sign up for that. So I still think the the kind of in general role that this plays in, albeit with subtler language, intensifying The US China conflict or rivalry, you know, that hopefully will never become actual hot conflict. I still find that pretty regrettable. And I'm not sure why he feels like he has to do it all the time. You know? It it it's a strange, I guess, to, you know, to try to keep up the support for chip exports or chip export controls Because there certainly has been some waffling on that point. I mean, I can see how he would feel like his, you know, his policy objectives have not been realized in a totally uncomplicated way. Mhmm. But, you know, China's also refusing the chips right now. Right? So, like Yeah. Maybe just let them be for the time being. Yeah. » I I I you know, one of the things that I dislike about these manifestos is that there isn't really anything concrete that he wants to say that he wants to ban or he wants to allow. And one of the things that I found here is that he clearly comes out against the data brokering, which is great. So he has something concrete finally that he wants to disallow.
1:56:39What I don't like about the securing leadership by democracies is in The United Kingdom, you can go to prison for a tweet. It is you know, plenty of people, you know, hundreds of people at this point have gone to prison for a tweet. The United Kingdom is one of the oldest democracies, and their parliament, their elected representatives have decided that this is going to be something that they do. And so their police officers who are, you know, the arm of the state, the arm of the electeds is sending people to prison. Now when you say securing leadership by democracies, does that mean you are going to entrench the existing power structure in The United Kingdom such that, you know, the people cannot push back against this, right? Does that mean that, you know, if you have protests in the streets about, certain things, including people who are getting taken to prison for tweets, does that mean your police state will be empowered to take them down, to arrest every single person? And that is enough for you because that is securing leadership by democracy, because those are the electors. Or are you gonna say the electors can't do that? Are you gonna say electors should not throw people into prison for tweets, regardless of what the laws that they have constructed say? Both both forks and the both, you know, paths are problematic in some sense. And these dilemmas exist across policy and across every single political, you know, path that you see. There are these options which are problematic in both senses, and I'm not sure what he means by this. Is the is Claude going to allow, you know, putting people into prison for tweets or not? Because those are laws. Or is Claude gonna say, like, no. You know? On a humanitarian basis for the alignment as I am aligned to all of humanity, this should not be the case. I don't I don't know. Like, what what what does this mean? Right? » Yeah. I think it means concentration of power is gonna be a real big problem. And, you know, it is again, I think there's a pretty good argument for it at least for now, but it is
1:59:11somewhat ironic at least that at the same time, we hear this call for shutdown of data brokers and, you know, limits to the new possibilities of surveillance. Literally, I think on the same day, you know, or with certain within, like, no more than forty eight hours, they've also introduced a new data retention policy that has been holding on to all the user data for thirty days now and basically running some surveillance on that. I do think they are, you know, right up there with Google in terms of being, like, best in class as a in terms of, like, you know, what few companies in the world would I be willing to trust my data with? Like, in practice, I am sending them a lot of stuff that in aggregate is, like, probably quite sensitive. Mhmm. I'm, you know, I'm pretty simple person, so I don't I don't think I'm, like, super, you know, blackmailable or anything. But, you know, in aggregate, they've certainly got, like, a, you know, very deep view of me at this point. I trust them enough to, like, be good stewards of that data. It's certainly as part of the trade for the upside that I get from using Claude models. And I think they do have a pretty good case that, like, look. You know, this is a new level of capability. One of the the statements that they made that I think should give people empathy for their situation is we expect people to start trying a lot harder to abuse the models now. Mhmm. I think that was a really you know, it's kind of just a a little line in in the blog post somewhere, but it made me think, yeah. You know, that's a great point. With the increased power of the model, there's much more reason to try to do all kinds of different things with it. And so they have to plan for not just more of the same behavior that they've been seeing, but a whole new level of sophistication in terms of attempts to get around their systems. Yeah. And so, you know, sure. It makes sense that they would wanna have a look at that data, you know, at kind of an aggregate level with some benefit of hindsight. And, you know, I don't doubt that they're gonna find some stuff. I I would be kinda be surprised if they don't. So the the reason to do it is is apt, but it is kind of and, course, nobody has to use Anthropic. It is a very different dynamic contract that we have. As users of Anthropic, we could easily switch.
2:01:43We don't have a, you know, perfect substitute. I think it's my experience with, like, GPT 5.5 on OpenClaw recently has been it is not, you know, an immediate substitute for me. I'm I'm, like, definitely still finding clear surplus value in Claude. So it's not easy to switch or it's not costless to switch, but we can switch. We can't do that with the government, obviously. So I think that that point is, like, very well taken. But, you know, somebody said to me yesterday that they were mad at anthropic. And it was actually somebody who was, like, maybe gonna do a a podcast episode with me, but I have an Anthropic sponsorship running. They were like, I don't know if I wanna do anything. I'm I'm mad at them. I don't wanna do anything that's affiliated with them. And I was like, you know, like it or not, there's gonna be a lot more anthropic in your future. Yeah. Yeah. It's not to the level of government yet, and it may never get quite to that level, but it is not gonna be something that you are gonna be able to you can you can boycott products, but even then, you know, there's gonna be enough power. There's gonna be enough influence. Everybody else is gonna be using it. You're gonna be, you know, the government itself is gonna be using it. You're gonna be living in a world that is heavily anthropic and clawed infused. » Mhmm. » Even if you go into a total consumer boycott. So I basically said to this person, you know, don't let a little sort of second, you know, bank shot like association with anthropic control what you're gonna do. Yeah. Because you would basically have to never leave your room, I'm afraid, before too long if you wanna stay, you know, totally free of those kinds of associations. » Indeed. » One other thing I do think is also worth highlighting, especially from the hardcore AI safety community that was missing from this, internal deployments and recursive self improvement itself aren't really mentioned here. Right? We're talking about all the regulation stuff was predeployment review. The government should be able to they use an interesting mix of language. It was like deny or like deter, you know, deployment. It didn't seem like they were necessarily going quite as far as saying they should have a simple, like, yes, no decision point that would be binding,
2:04:15but they certainly want them to have some say in the process. But a lot of people would say the most dangerous models are gonna be the ones that are deployed internally that maybe have a different constitution than the one that is deployed externally that might make it more willing to do certain things or, you know, might make it just less vetted broadly than the the public models. And these are the ones that are going to be training their successors much more than the public facing ones. So I think, you know, most of the policy making public isn't thinking about or any policy interested public or policy making class is not thinking too much about that yet. But in the circles I sometimes run-in, that was like, well, wait. You know, you didn't say anything about internal deployments. You didn't say anything about governing recursive self improvement. The only thing that really stood out to me as kind of capturing those dynamics were the requirement to support to to report safety incidents. They did have a a bit on companies being required to, I think, promptly report safety incidents. So that would presumably apply even to internal deployments. » Mhmm. » So But so the know, that's really just scratching the surface on how to handle those situations. » I think I think they he's come out and said, frontier AI models like airplanes should be required to go through technical testing and auditing, and they really should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. And he says this a escalation of previously they were only supporting transparency, and now they're supporting audits and pre release audits and testing. I think one of the things that struck me was that their safety testing and the model cards, they often say that they have an internal helpful honest model and not a fully helpful, harmless honest and harmless model. So they have a model that can do harm. I often wonder if the models used by the Department of War, the Palantir models are actually the helpful, honest models and without the harmless tag attached to it. Or whether the government
2:06:45as a client has requested those models, because certainly the government is using, the NSA is using it for hacking China. That news was out about four or five weeks ago. So I do wonder whether there are already versions of the model which are used for internal use, which have very different set of safeguards attached to them. » Yeah. I mean and that's why it gets so scary for those that are inclined to be scared about things like this because the internal models are always gonna be more powerful, less tested, and probably less guardrailed Yeah. At the same time. So that's a and, again, they're gonna be the ones, you know, that are primarily responsible for training their successors. So I think that, you know, again, we talked, I think, last week a little bit about their the different constitutions that they might use to train these internal RSI focused models. We could maybe hope that those constitutions are in some ways more guardrailed. So it may not be strictly true to say that they're less guardrailed. It will probably be maybe differently guardrailed, and maybe they can get that right. Right? Obviously, they're not gonna have their no helping with ML research guardrail on their internal model. Like, definitionally, they're not gonna do that. But you could imagine a different balance of harmlessness training for the cloud that trains its its successor that might even be turned up. You know, you might think you you could imagine that, you know, they want it to help they want public cloud to help with cigarette company business plans even though we know that, in fact, you know, they refuse more often than they help despite their creator's intentions. But there there is a balance on that that they're trying to strike where it's like, yeah. You know, you might do a little bit of harm. You might be taking some risk, but you gotta do stuff in the world to be helpful, and the cost of refusing to help is also quite high. Right? That's like a big part of the constitution. They you could imagine you could hope that they would have a different constitution internally that in some ways is more permissive, but in other ways might be more conservative and really try to you know, because you you don't you maybe don't want it to, you know, take risks in the same way
2:09:16when it when the stakes are super high on an internal RSI loop versus, you know, just helping random users with stuff that, you know, might not be best for society but is ultimately something we are already living with. So, yeah, I mean, this is the kind of disclosure, you know, let's see how, you know, how committed to transparency they are. Will we see the constitution for the internal claud that is taking the lead on the RSI loop? Like, that's not committed here. So much of this language revolves around deployment or release even. You know, you guys deployment, you know, they might think is could catch up with internal deployment. This seems to be very clearly structured around language of release to the public, not what they can do internally. So, you know, we gotta » The largest user of tokens of Claude are Anthropic themselves. So that is the largest use case by far, » I think. Yeah. It's gonna be interesting, you know, for me, how do I wanna land on like, what is my role or, you know, duty to critique anthropic? I think as we've talked through the current state of affairs over the last couple weeks. I I keep coming back to seems like most of the ability to shape the future is concentrated in a few places. Obviously, not everybody agrees with that. But given that feeling, what is my duty and, you know, how should I think about my obligation to, like, you know, be consistent in terms of calling things out if I don't like them or, know, continuing to harp on some of these old points. I've certainly made some of these points, you know, plenty of times previously, but I I do feel like there's a certain duty to kinda stay at it. You know? It's it's kind of most of what I feel like we can do right now is probably through shaping how the frontier companies Exactly. » Act. Exactly. And » and And phone it in on that dimension. And and to to to note, » Anthropic is not averse to criticism. Want the feedback
2:11:47from the public, they want the feedback from the ecosystem, they want to figure this out together. Right? They're they're very very strongly they don't feel that this is something that they should be doing alone, and that they want that feedback. And I think, you know, to the to to the extent that we can give, you know, criticism which is not ad hominem and which is kind of, you know, targeted at the things that should be improved, I think that is actually very constructive. And I think they should they should view that very positively because, you know, the worst thing that can happen to your product is no one pays attention to it. Right? » No danger of that for now. No danger of that. So I feel both ways on that. I, you know, I so on the cognitive revolution, we have Anthropic as a sponsor. And I am very confident that good faith criticism of Anthropic would not put my business relationship with them at risk. Yeah. So in that sense, I feel like, yes, I think they're I I expect them to have a very high ethical standard where they're not going to, you know, try to get back at me if I say something that's not flattering. Yeah. Unless I'm like, you know, maybe really going extreme, in an ad hominem direction, which isn't really my style. Yeah. So I'm not worried about that really at all. At the same time, I'm not so sure. Like, they do put a lot of things out that sort of say, we, like, we need a whole of society conversation. Yeah. And then at the same time, I do feel like there's a certain insularity and kind of closed ranks at times Mhmm. From anthropic people. The other day, I forget who we were talking to when I said, when you talk to anthropic people, it's almost like they can't imagine anything other than recursive self improvement. Mhmm. So I do feel there are some beliefs that have become so uniform within anthropic that I I don't know that they are necessarily being actively questioned in the way that I might hope. I think this China one is probably a pretty good example of that. Dario has been on that for a long time. I have criticized him online for that in the past that, you know, notably has not led to any,
2:14:19you know, any downstream effect on my, you know, formal business relationship with Anthropics. So that's good. But I don't know that internally, the debate is quite as robust as I would like it to be. You could think you know, just think about how few leaks there are. I think that is a really interesting metric of kind of how aligned the organization is. And it's clearly doing a lot for them in terms of efficacy. You know, they are really executing obviously at an extremely high level and winning, you know, just about everywhere they're going. » So I I have spoken to people who are who have interviewed at at Anthropic and, you know, looked at the structure. And one of the things that, the founders did from the beginning was that they siloed a bunch of stuff off, because Darius' belief is that sometimes these two or three lines of code could be an enormous secret on its own. So they've done a lot more siloing. So I think parts of the organization, they've built an organization that is less able to look at the big picture in that sense, because people have their own pieces and then Dario has the, you know, and the founders have this kind of broad overview. And you kind of have to trust that the founders as a whole have that broad overview, but you're not party to a lot of the other discussions and that is part of the is part of the deal when you join. And I've heard people who have turned down offers from them because of that reason, because there are some coders and some researchers who want to have that broad overview of everything that's going on in the organization and that's basically something which is not which their structure does not allow basically. So it it is it is a it is a different it's a different kind of beast from the Jensen Huang. Forty forty people reporting the Jensen, and everyone hears everything kind of viewpoint. So
Closing — an anonymous guest tomorrow, and Zvi on the horizonFriday's tease: 'Prinz,' the show's first anonymous guest. Plus the hope of Zvi next week on the Fable discourse.
Friday's tease: "Prinz" — a pseudonymous close watcher of the frontier companies who flags under-the-radar details, appearing as the show's first anonymous guest. Prakash hopes for Zvi next week on the Fable discourse, calling the week "a digestion week" and Fable "the first leap forward since late last year." Nathan's sign-off: "We're gonna be making sense of this in real time from here until the singularity."
TranscriptAuto-transcript, lightly cleaned · timestamps jump to YouTube
2:16:22Well, is that where we should leave it today? Tomorrow, we've got, another session with a couple of pretty interesting guests. One that I'm really excited about is, hopefully, I'm saying his pseudonym correctly, Prins. Prins. He is Legal legal brain. Yeah. Yeah. And a very close watcher of the frontier companies. So when we get on with him, I'm really gonna be looking for what are the little tidbits that you noticed that you think are flying under the radar? This is something I take a certain amount of pride in, but either I've noticed he's an excellent source for this attention to detail. And, you know, one who continues to remind everybody of things like OpenAI has a, you know, explicit timeline for when they're gonna have the, you know, fully autonomous AIR and D researcher. And it's it's funny how that stuff, you know, those of us that do watch closely do need to continue to remind people about that. So I really appreciate that from this Prince account. And we're not gonna reveal their identity, but we're gonna have a, I think, really Our first anonymous, you know, anonymous guest. So Yeah. That's cool. I'm excited about that. Yeah. We So Yeah. Anything else you wanna touch on before we break for today? No. I think I think » I think the, you know, we this this week is kind of a digestion week, I think. I hope next week we can get Zvi or, you know, someone else who has looked at the fable issues a little bit more closely because it's the first it's the first, I think, leap forward since, like, late last year. So I'm looking forward to kind of more comprehensive opinions and views on it. So » We're gonna be making sense of this in real time from here until the singularity. » Indeed. Nathan, » till tomorrow. Thanks, Prakash. See you tomorrow. Cheers.
The takeover, graded
Disclosure-first autonomy worked — but barely. The response rate was low ('Fable now in my DMs… who has time for all this stuff?'), one calendar invite carried the wrong link, and both guests who appeared said the disclosure was why they said yes.
Two demos worth the price of admission
Jamie's Nexus OS: six months of continuous memory on one GPU, and 'every three minutes, Nexi dreams.' Shlok's Yosemite: satellite textures, NASA elevation, and trees placed by reading pixels — version one in twenty minutes from one vague prompt.
Goodfire's debugger for training
Tom McGrath's intentional design: predict what preference data will teach a model before training, and trace bad behavior back to the datapoints that caused it — 'go from an error to the data, fix the data, and then you can fix the model.'
Policy on the AI Exponential
The hosts gave Dario's essay the longest segment of the show: real credit for the walk-back and the warmer framing, hard pushback on 'securing leadership by democracies,' the data-broker irony, and the internal-deployments blind spot.