AI:AM LIVE — July 8, 2026 — LTX's Zeev Farbman on Open World Models, GPT-5.6 Cleared for Launch, and Hosts vs. the AI Superforecasters

The rundown

1. 0:00Opening37 min

   Opening: GPT-5.6 Cleared for Launch, Anthropic's J-Space, Self-Improving Harnesses, and AI Law Everywhere but CongressThe machinery around the models was the story: Commerce's CAISI cleared all three GPT-5.6 models (Sol, Terra, Luna) for a July 9 public launch — the first frontier model to graduate the government gate, still preview-only as of air time — while OpenAI's exhibition model beat the human field at AtCoder's World Tour Finals; Anthropic's 'global workspace' paper found a small privileged slice of Claude's activity and set off a consciousness-framing fight; Replit claimed a self-improving agent loop with the mechanism disclosed but no numbers; Beijing weighed export curbs on China's top models the same week Illinois signed first-in-nation frontier safety audits; and Coefficient Giving reportedly made its largest-ever AI-safety grant, $160M to Geoffrey Irving's Resolution.

   Watch

   ▶Play this segmentYouTube · from 0:00

   As aired

   Nathan and Prakash opened with a batch of AI-generated music videos — a mix of Google's Omni and LTX's audio-prompted model — teed up ahead of the day's guest conversation with LTX's CEO, working through some live audio-routing hiccups to get playback into the stream.

   They riffed on how far AI music-video generation has come, including how Nathan curates genre matches for each episode, and contrasted Fable's more tasteful outputs against earlier Opus attempts that leaned too literal — and briefly cringeworthy — with certain imagery.

   A tangent on K-Pop Demon Hunters covered how the Netflix hit was built music-first and became a case study in 'sloppified' content that turned into a genuine cultural and commercial phenomenon, complete with an upcoming concert tour, heavy merchandising, and Sony having sold the rights to Netflix.

   The hosts brought their AI co-host Q online for a live news briefing covering chip announcements (OpenAI/Broadcom's 'Jalapeño'), an HP-OpenAI partnership, Anthropic restoring model access after eased export controls, a new UN-backed AI for Good Commission, and a status check confirming Gemini 3.5 Pro has still not launched publicly.

   Top story: GPT-5.6 was cleared for launch by Commerce and confirmed by Sam Altman for a Thursday release, with early users (Theo, DAX/OpenCode) praising its agentic, sub-agent-friendly behavior and speed — even as both hosts said Fable still edges it out on most tasks they've tried.

   Nathan relayed a theory that models may not fully 'top out' before release, since iteration cycles have gotten shorter than the time it takes to run long-horizon tests — floating the idea of a post-release model recall or rollback mechanism as a possible fix.

   The two debated why some power users prefer GPT-5.6 while others still favor Fable, landing on a theory that more organized, process-driven users gravitate to GPT-style models while more improvisational users prefer Claude/Fable.

   They closed the segment with a nod to Lilian Weng (Thinking Machines) and Replit's public writing on 'harness engineering' and closed-loop recursive self-improvement, before teasing the day's guest — Josh, VP of comms — and cueing up a K-pop-flavored AI music video to lead in.

   Key moments

   It's kind of slop that became non-slop.

   DAX says their token consumption went up 5x. It is just such a fun model to use.

   The iteration time from model to model is now potentially shorter than the time horizon it would take a model to top out.

   What we covered

   GPT-5.6 goes public tomorrow — the first frontier model to graduate the government gate. All three GPT-5.6 models — Sol, Terra, Luna — were set to launch publicly Thursday, July 9, 2026, after Commerce's CAISI ran added tests and cleared the rollout under the June EO's 30-day pre-release review (Axios, Jul 8). As of air time GPT-5.6 was still the ~20-partner preview running since June 26 (API/Codex only, not in ChatGPT); Terra was reported at roughly GPT-5.5 level for about half the price (Neowin, Jul 8). The tension: testers call it the best model they've used, while METR's June 26 eval couldn't pin down its capability. Same 48 hours, OpenAI's exhibition model beat the human field — including AtCoder's own president — at the AtCoder World Tour Finals Heuristic division in Tokyo.

   

   Pietro Schirano

   

   @skirano

   ·Follow

   I can finally talk about 5.6. I’ve been testing it for months and, without exaggeration, it’s the best model I’ve ever used. Fast, smart, genuinely creative, and you guessed it, they finally fixed front-end design. I haven’t needed to check the code I’ve written in two months.

   

   OpenAI

   @OpenAI

   GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We’re expanding preview access globally now.

   

   4:30 AM · Jul 8, 2026

   7.5K

   Reply

   Copy link

   Read 387 replies

   

   chokudai(高橋 直大)@AtCoder

   @chokudai

   ·Follow

   負けた！AIに完敗しました！ とても良い問題で、人間の良さがとても出ていて、競技時間も十分にあり、それでもAIに圧倒された。一つの節目としては素晴らしい問題と素晴らしい競技者だったと思う。観戦してくださった方々、ありがとうございました！明日のAlgorithm部門もぜひ見てね！！ #AWTF2026

   10:09 AM · Jul 8, 2026

   618

   Reply

   Copy link

   Read 2 replies

   J-space: Anthropic finds Claude's workspace — and the consciousness fight starts. Anthropic's 'A global workspace in language models' (Jul 6, 2026) identified a privileged slice of Claude's activity — under a tenth of it, holding a few dozen concepts at a time — that the model can describe, hold in mind, and reason with; ablate it and multi-step reasoning collapses while fluent speech survives. The paper explicitly does not claim Claude has experiences, but the framing fight came fast: OpenAI's Boris Power read it as evidence of 'access consciousness,' Neel Nanda called it a fantastic paper while warning the consciousness-theory analogy is easy to over-read, Eleos AI praised the work while staying more cautious than the authors, and Daniel Kokotajlo said a few dozen more advances of this magnitude would demote loss of control as his top risk.

   

   Anthropic

   @AnthropicAI

   ·Follow

   New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

   Watch on X

   5:34 PM · Jul 6, 2026

   27.6K

   Reply

   Copy link

   Read 1.2K replies

   

   Robert Long

   @rgblong

   ·Follow

   Eleos wrote a commentary Tl;dr -important, excellent work -we’re more cautious than the authors about the stronger claims of 'global workspace' -still, it's evidence in the direction of access consciousness -investigating AI consciousness is tractable and urgent More:

   

   

   Anthropic

   @AnthropicAI

   New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

   Watch on X

   7:17 PM · Jul 6, 2026

   175

   Reply

   Copy link

   Read 10 replies

   Self-improvement is now a harness discipline. Replit said it 'closed the loop': an internal agent reads production failures, drafts candidate changes as PRs, evals them against its ViBench suite plus A/B tests, and ships — mechanism disclosed, numbers not (Replit blog, Jul 6, 2026). Lilian Weng's 'Harness Engineering for Self-Improvement' (Jul 4, 2026) supplied the frame: recursive self-improvement starts in the harness, not the weights, and weak evaluators are the binding constraint. The hosts held both: claims running ahead of receipts.

   

   Amjad Masad

   

   @amasad

   ·Follow

   Many are asking how Replit is improving so rapidly—we closed the loop and the agent is self-improving. Technical details here:

   

   Michele Catasta

   

   @pirroh

   x.com/i/article/2070…

   10:23 PM · Jul 6, 2026

   2.3K

   Reply

   Copy link

   Read 57 replies

   

   Prakash

   @8teAPi

   ·Follow

   I told Fable it was down to its last 10% of tokens and to use codex and sonnet for everything it could. It did as much work as the first 90%. We’ve hit the mid-level engineering manager. 6 months ahead of time.

   

   Theo - t3.gg

   @theo

   This is the relevant section of my CLAUDE.md I'll be real - I haven't read it much. Just vibed out what I was looking for with Fable, and had it confirm it can use Codex for the things I care about I still find Codex to be WAY better at computer use, verification of UI/UX work,

   

   6:53 AM · Jul 7, 2026

   1.1K

   Reply

   Copy link

   Read 20 replies

   Beijing and Springfield, same week: AI law lands everywhere but Congress. Reuters (Jul 7, 2026): China's Commerce Ministry has spent a month meeting Alibaba, ByteDance, and Z.ai about restricting overseas access to China's top models — open weights included (Qwen, Doubao, GLM-5.2) — via a tiered regime; a proposal, not yet policy. Two days earlier Illinois signed SB 315 (Jul 6, 110–0 House / 52–5 Senate): first-in-nation annual independent safety audits for frontier developers over $500M revenue, 72-hour incident reporting, and whistleblower protections, effective January 1, 2027. Two governments converging on the same instinct from opposite directions — with the transparency gap still open.

   

   prinz

   @deredleritt3r

   ·Follow

   Reuters: - The CCP has held meetings with top Chinese AI labs (including Alibaba, ByteDance and Z.ai) about potentially restricting overseas access to China's most advanced AI models, including both closed and open-source models. - Under the proposal being Show more

   2:06 PM · Jul 7, 2026

   288

   Reply

   Copy link

   Read 25 replies

   

   Nathan Calvin

   @_NathanCalvin

   ·Follow

   Nice summary from Thomas on the current big gap in frontier AI safety legislation: “So far we have transparency and auditing… "But if a company wanted to transparently do dangerous things, there's nothing that is actually stopping them from doing that in any of these laws."

   

   MTS

   @MTSlive

   SITUATION EXPLAINED: Why does every current AI safety law have a fundamental gap? We asked @Thomas_Woodside, Co-founder at Secure AI Project. "So far we have transparency and auditing. So that means that if a company wants to be transparent about the fact that they're not doing

   Watch on X

   10:06 PM · Jul 7, 2026

   13

   Reply

   Copy link

   Read more on X

   $160M for Resolution — reportedly AI safety's biggest single bet. Coefficient Giving (the renamed Open Philanthropy) made what its team called its largest-ever grant — reported at $160M ($108M unconditional plus $52M milestone-tied, per secondary sources) — to Resolution, Geoffrey Irving's alignment nonprofit (announced Jul 6, 2026): the ex-UK AISI alignment team merged with Timaeus, Berkeley-based, targeting 40–80 researchers within two years, aimed at misalignment of superhuman AI rather than today's models.

   

   Max Nadeau

   @MaxNadeau_

   ·Follow

   Largest grant my team has ever made! Some reasons we moved fast and went big: * Geoffrey was a great leader at UK AISI * They're focused on (mis)alignment of superhuman AI, not today's AI * Non-profit structure is more conducive to open-mindedness about the state of alignment

   

   Geoffrey Irving

   @geoffreyirving

   We're excited to announce that Resolution has a $160M grant from Coefficient Giving: $108M unconditional, with a further $52M conditional on hiring and compute needs. We'll use it to grow teams across our research portfolio and invest heavily in research automation. 🧵

   

   3:39 PM · Jul 6, 2026

   230

   Reply

   Copy link

   Read 8 replies

   Full transcriptLightly edited · timestamps jump to YouTube

   0:04

   Q: Feel it coming down, coming down low. Feel it coming down, coming down low.

   1:05

   Nathan Labenz: Alright, cool — looks like we're live. Good morning, Prakash.

   1:09

   Prakash: Good morning, Nathan. It's Wednesday, July 8th, 5:58 AM — we're a little earlier than usual today, and we're gonna play a video. So let me move you in there. Let's see... let's see.

   1:29

   Nathan Labenz: Cool, I've got three hot-off-the-presses music videos made just for today's conversation with Zeev, the CEO — hopefully I'm saying his name right, I'm always very nervous about that — CEO of LTX, the company that makes music-gen models, including one that takes audio in. So in his honor we've got a couple of these queued up. I want to play one actually kind of for him as we lead into the conversation with them as well. But here's one. Yeah, okay, I think this should work — tell me if you can hear it. No sound yet? Yeah, let me try to stop it. It's there, but I do have shared tab audio checked — okay, let me try it one more time, going to system audio as well here. Let's see if this works any better. Hold on a second there. Alright, how about this? Alright — you're gonna keep playing it, you check it. Data centers appear in the background. So there you have it. We could play another one just before we bring them on a little bit later — what do you think?

   6:01

   Prakash: You know, it's—

   6:02

   Nathan Labenz: You see some artifacts in there still, but it's come a long way — that's for sure.

   6:06

   Prakash: It's come a long way. I mean, you have audio, you have video — you have, you know, an authentic, semi-authentic-looking Rastafari character. You got the jar in there — I love that they threw the jar in, you know?

   6:22

   Nathan Labenz: Honestly, my value-add is really mostly in finding the right genre to match the episode. That's where sometimes it does a good job, but other times I gave it some defaults of genres that I like, and it's maybe a little too reliant on those defaults. I should probably go back and say, these are just examples — really get creative and think about what would be the best fit for this particular content. That's really the biggest thing I do at this point in the process: make sure we get a match so it feels like the right song for the episode. Everything else is basically fully automated, and you do feel the taste of Fable very clearly relative to Opus on some of these videos. The video Opus created for various songs I tried was a little too literal, a little too on the nose — in some cases kind of cringeworthy because it was so on the nose with the subject matter. In the case of the Disempowerment Blues, it made a video that included scenes of people picking cotton out of cotton fields, and I was like, yeah, I don't think this is quite the way we want to put this out into the world. So when I switched models and put Fable on the case, I said, you can check out what Opus did in previous generations, but notice that it's kind of cringe because the cotton fields thing is just not gonna play well — and it totally got it. Obviously there's some there there, but it was like, yes, I absolutely see how this was a little too literal-minded, and how people might bristle at the imagery we used there. So it goes back and does a totally different thing, and you can definitely feel the taste factor is notably higher. But I've had a lot of fun doing this. And the LTX model that takes in the sound as an input — the video we just watched, and all the ones we'll show today, is a mix of Google's Omni and this LTX audio-prompted model. It kind of allows for that classic music-video dual track where you've got a story playing out, but interwoven with that you have scenes of the performers doing the music. The LTX was critical to that — I didn't have a way to get somebody actually singing the song without it. But I thought Fable was able to parse it up and send shots to different models pretty effectively, and the sync is really good. I've been pretty impressed with how far that's come, how quickly — for how little effort I've put into it, getting decent music videos out has been quite the trip.

   9:40

   Prakash: It has come a long way. I go back to that Jeff Bridges interview on the pod — he says everyone in Nashville is using Suno. They're actually playing around with these things and using them. So I suspect that by next year there'll be a 70-90% AI-written song, and even if it's not 100% AI-written, it's become an input into the creative process. It's also converging with some other trends. For example, the breakout hit of the last 12 to 18 months, if you're in the teeny-bopper crowd, is K-Pop Demon Hunters. And K-Pop Demon Hunters is basically 'sloppified' — they went ahead with the cartoon and the anime first, and the anime required music because it was about K-pop. So they wrote the music first, then found performers after the fact to perform it — not all the music is performed by the same performers. Then they went ahead with the Netflix series. The performers didn't get paid that much; they were just making a soundtrack. And K-Pop Demon Hunters became absolutely enormous. For some reason Sony originated the production and sold all the rights to Netflix, so now Netflix has one of the key media properties for this age group, and it's gonna go all the way. Number one, it's anime, so you don't get characters aging out. Number two, you've captured this core group of 8-to-14-year-olds when they're very young, so as the age group progresses older and has more money to spend, you own that property as they age. Netflix doesn't really know what to do with it, because Netflix's normal strategy is: sign people for three-season deals, do the first season, if it works out you continue for two more, then you cut — because after three seasons the actors negotiate for higher pricing, and for most Netflix shows that math doesn't work, so they let go. In the K-Pop Demon Hunters case, they'd only negotiated for that one thing, and they hadn't even engaged with the artists — the artists were just shoved in there as part of the soundtrack. So Netflix had to go engage with the writers and the artists who perform, even though the artists weren't the ones who wrote it — the artists were almost nameless. And now there's a concert tour coming, which is going to be a billion-dollar concert tour for sure. Massive merchandising too — cartoon merchandising is excellent, you can merchandise in so many different ways. I don't even know if Netflix knows how to run the property, because it's more like a Disney-style property — maybe they should just sell it to Disney, honestly. But it's kind of slop that became non-slop — 'we just need something to fill the time' that became non-slopped. It's pretty cool. I'll have to check it out.

   13:35

   Nathan Labenz: I have not seen one second of — what's it called? — K-Pop Demon Hunters?

   13:39

   Prakash: K-Pop Demon Hunters. It is all the rage if you have an 8-to-12-year-old girl in your house — all the middle-school kids are singing the songs. I've been singing it for the last 18 months. It's pretty cool. I mean, it's also K-pop — it's literally K-pop, you know? But the entire thing is in English — no real Korean at all. It's pretty impressive in terms of how much K-pop has penetrated the American zeitgeist, so to speak.

   14:20

   Nathan Labenz: Well, coming up I've got a K-pop music video for you as well — we can maybe lead into our first, and only, guest session today with that. What's on your mind today from the latest AI news?

   14:36

   Prakash: So why don't we fire up Q, and let's see. Hi, Q.

   14:46

   Q: Q is live.

   14:50

   Prakash: Q, can you do a web search for the latest AI news and give us a short briefing?

   14:59

   Q: Let me check what's breaking and pull together a quick brief. Top line: compute and governance are the big themes. OpenAI and Broadcom announced a new inference-focused chip called Jalapeño, tied to a very large data-center buildout. HP also announced a strategic partnership to use OpenAI systems across its operations and software work. Anthropic said it's restoring access to certain models after export controls were lifted. On policy, a UN-backed AI for Good Commission is launching, with its first meeting scheduled for today. Microsoft highlighted progress on quantum hardware and its Discovery agentic system. And Google's been rolling out steady updates across its AI stack heading into July. If you'd like, I can narrow this to the three most consequential items for builders and investors.

   15:55

   Nathan Labenz: What's the deal with Gemini 3.5 Pro? I was under the impression that should be launched at this point — is there an update on what's going on with that?

   16:11

   Q: Let me check the latest on that and sort out what's actually confirmed. Short answer: it doesn't appear to be publicly launched yet. Google's own June update said they were hard at work on Gemini 3.5 Pro, already using it internally, with a rollout planned for the following month. What is available is Gemini 3.5 Flash and other Gemini updates, but not a generally available 3.5 Pro. If you like, I can scan for any mention of early access, workspace rollout notes, or credible reporting that suggests timing has shifted.

   16:49

   Prakash: Thank you, Q. And so — I did a bunch of additional updates yesterday, some fixes, and I think these things are fairly capable now. I think you could presumably use it as a full-time cohost, which is quite amazing.

   17:17

   Nathan Labenz: Q, can you stay on with us, but only respond when we ask you a specific question, or when there's just a golden opportunity to add value — don't feel the need to respond every turn.

   17:35

   Q: Understood, Nathan. I'll stay on and keep quiet unless you call on me directly, or there's a truly exceptional moment to add value. I'll aim for concise, high-signal contributions, and otherwise remain in the background.

   17:50

   Nathan Labenz: Alright, cool. Let's see how it goes.

   17:56

   Prakash: So the big news yesterday, I think, was that GPT-5.6 was finally cleared for launch by Commerce — Howard Lutnick's group at Commerce finally cleared GPT-5.6 for launch. I think the news broke sometime around 6 or 7 PM Pacific time, and a couple hours later Sam Altman tweets that it's gonna be available on Thursday. So we have a confirmed Thursday launch. And then the next thing that happened is a bunch of influencers online began posting that they've been using it for months. One of my favorite guys, Theo — Theo Gigi — has said it's his favorite model, and he's been using it for months. It's not as good as Fable, but it is very agentic — it will take a task and chase it down to the ends of the earth. It's very good at organizing sub-agents, apparently, so it's very sub-agent-friendly. And the other person whose taste I trust is DAX — DAX runs OpenCode — and DAX says their token consumption went up 5x. 'It is just such a fun model to use,' he says. The other thing to note is that GPT-5.6 is supposed to be much faster, and I don't think people have a good sense of that yet — I definitely feel like fast responses make me more engaged, it's more engaging. So I think the speed issue is gonna matter. One of the complaints that's been going back and forth between OpenAI and Anthropic is that Anthropic puts out these very capable models but doesn't tell you how much compute they're using. Norm Brown from OpenAI has been complaining about this — he's saying, look, you guys are beating the benchmark, sure, but you're not showing how much compute it's taking, so you don't really understand which model is better. We can do two things to improve a model: run it longer, or run it in parallel and average the best answer — an ensemble kind of strategy. So he says it's not fair, because we all have different approaches, and it's also hard because you don't know how big the models are — model size is also a measure of compute. So that's one of his complaints. I don't see how you resolve that, Nathan — do you?

   20:57

   Nathan Labenz: Well, you could put the cost on the x-axis — that's the big trend, certainly, but that's only a partial solution. I've seen him say in a couple different venues that we literally don't know if today's models top out, because this is another one of these really weird realities we're entering into: the iteration time from model to model is now potentially shorter than the time horizon it would take a model to top out on the absolute hardest, most ambitious, long-running task. So he's saying you might need to give these things a month or two — what happens if you spend a million dollars with one of these models? He basically says we don't really know if they top out, and calendar-wise, the time from when we're done training to release is enough to run a lot of standard tests, but not enough to really do these long-standing tests. I even heard him propose something like a model recall program — obviously that doesn't work in open source, but it could work in an API paradigm, where a model gets released on day N after it's deemed ready, which gives you N days of head start running it on really long time-horizon tests. And potentially you, as a frontier company, need to be ready to see at day N plus 30 or whatever that, actually, we're now starting to see problems once we get these things into the super-long-running regime — and so you either need to recall and fix it, which means taking things offline (which is gonna be kind of painful), or you have a longer delay to launch, or you just fly blind. I think that's quite interesting — the idea that we're at a tipping point where the iteration cycle is just plain shorter than the testing time horizon is a very weird world to find ourselves in. But yeah, I'm excited to try it. I've seen mostly pretty effusive praise for 5.6. You said Fable is still better, and Matt Schumer said something similar — it's great, but by and large Fable's still better on most tasks he tried. I guess my expectation is I'll still use Fable as the director and try to get it to make use of 5.6, try to actually max my tokens on OpenAI as well as Claude now. But it's always a good reminder that these things are quite confusing — you get people like Pietro Schirano, I hope I'm saying his name correctly, who's a top-tier user and has done really impressive stuff in the design space using language models to create sort of a next-gen Figma. In general it's just tons of great little demos, and his GitHub is one to follow for sure. His 5.6 review was, without exaggeration, 'it's my favorite model.' I don't know — do you have any theories? It's probably too early to have real theories until we get substantial hands-on time ourselves, but I'm always like, what's going on there — why is it some people's favorite, while other people still see Fable as clearly better? I think that's gonna be me, but I'm not really sure why. I don't know if it's a matter of cognitive style, or how organized people are. One thing I've noticed a little is that people who are more organized, better managers, better delegators, better at standard operating procedure tend to be happier with non-Claude models, whereas people who are a little more like me — kind of vibing their way through life in a not-so-structured way — seem to gravitate more toward Claude. Do you have any other synthetic takeaways, or how would you summarize what you've seen in terms of who prefers what at this point?

   25:58

   Prakash: So there's a couple of things that occurred to me. Number one is that a lot of the people who are preferring GPT-5.6 are very good coders — some of the best coders we see online in the space — and they were also in the early release batch. So I suspect a couple of things. Number one, as you pointed out, people are more organized in the sense that they know how to organize their code a little bit better. This is the architecture, these are the system requirements, I want you to use test-driven development, I want you to do this and this and this, I want you to do a review afterward and come back to me — those people who are able to organize their work a little bit better, if you give the marching orders to GPT, it carries it out. And GPT-5.6 carries it out with more motivation than earlier GPTs. So that's one. The second thing is I think there's a little bit of enterprise sales going on, in the sense that when the OpenAI team is releasing to early adopters, they are actually listening to the feedback and fixing things quickly. And I think that kind of responsiveness matters a lot for enterprise users and early adopters — it's really not pleasant to not be able to get in touch with anyone from the firm and have them not care, especially when you're an early adopter pushing all this feedback in and helping the team explore what their model can do. One of the problems you have right now, as you pointed out, is there's not enough time to actually explore the full depth of the model before the next release. So I think those two things are there. I think Anthropic, as a result of their talent-density framework where they don't hire that many people, and as a consequence of feeling that the model should be the product — the model should do the marketing, the model on its own should be good enough to pull people in — and, to be honest, having succeeded—

   28:40

   Nathan Labenz: Working for them.

   28:41

   Prakash: We're working for them, right? So from using that kind of framework, they've kind of gotten themselves into: we have to make the product, we have to make the model better — if the model is better, everything else will get sorted out. So, yeah, it's just a different way of living in the world. I'd say I think the Anthropic method is a little bit misaligned to humanity, because it means you're not listening to the people who actually have opinions, and you're not respecting those opinions or resolving them as quickly as possible. It's more of a 'father knows best' kind of framework — we think this is the way it should be, and if you like it you can come along, or not. Too bad. So it is what it is.

   29:36

   Nathan Labenz: A couple other things jumped out to me that I don't think we can do full justice to, but I'd at least point people to check them out. Lilian Weng, who's at Thinking Machines now, I believe — formerly OpenAI, now co-founder of Thinking Machines — posted a very well-received blog post on harness engineering for self-improvement. And there's also one from Michele Catasta, who's the AI lead at Replit, on continual learning for agents. They're both really a sign of the times, I think, in the sense that we've been talking about recursive self-improvement, and it's now happening where it's not just Anthropic people saying it's starting — you're starting to see these, I guess you'd call it Thinking Machines and, I don't know, probably Replit — these different kinds of companies that have been aggressive early adopters starting to raise their hands and say, we've closed the loop too, and that's how we're progressing so fast. That was Amjad from Replit, basically — he said people have been asking, how is Replit improving so quickly lately, and the answer is we've closed the loop: the agent is now improving itself, and here's how. I definitely want to spend more time with those — I think we should get into the weeds and pull out some gems on exactly what people are doing, but I haven't had time to do that quite to the level I'd like yet. But even without having dug in that deep, I still think it's worth calling out that we're seeing more companies stake their claim on having achieved some form of recursive self-improvement, and the recipes are increasingly getting shared as well.

   31:50

   Prakash: I think one of the things — okay, so Lilian Weng, Lilian Weng, spelled W-E-N-G, on X. Let's see if Q can pull up the X search on X and give us a brief summary of what Lilian Weng is thinking about.

   32:15

   Nathan Labenz: And, Q, for what it's worth, it's also on her GitHub if you need to search there, if X isn't accessible.

   32:25

   Prakash: So let's see if Q can come back to us with the results.

   32:33

   Q: I'll pull up what she's been sharing recently and give you the highlights. I'm not seeing much from X right now, so I'll check another source. Short version: she's arguing the action is increasingly in the harness around the model, not just the model itself. That means workflow loops like plan, execute, test, refine, plus tool use, permissions, persistent memory, and splitting work across sub-agents and backend jobs. She also talks about self-improving harnesses, where the agent improves the system that runs it — though that only works well if the underlying model is strong enough. The takeaway is that the next big gains may come from engineering the surrounding software system as much as from model improvements. Source: Lilian Weng's Lil'Log post.

   33:33

   Prakash: Thank you, Q. Nathan, you seem to be muted.

   33:39

   Nathan Labenz: Classic background noise here. Any deep-cut highlights, Q, that you'd say are kind of surprising but notable takeaways? We're in this every day, so we're familiar with the basic stuff — but what jumps out to you as the deep-cut, hidden-gem thing I should be paying attention to that I might not have already noticed?

   34:07

   Q: Let me pick out a couple of less obvious threads worth watching. Two deep cuts, Nathan. First, the harness idea is bigger than it sounds — the winning edge may come from workflow engineering, persistent memory, and evaluation loops, not just model weights. Second, several teams are quietly closing the loop so agents improve their own systems — that's a meaningful step toward practical recursive self-improvement. And third, compute transparency and cost-per-quality still lag, which makes comparisons and benchmarking trickier than they look.

   34:42

   Prakash: Very true.

   34:47

   Nathan Labenz: Yeah, indeed. Cool. Well — do we have a guest in the waiting room, or not yet?

   34:55

   Prakash: Yeah, we have Josh, who is the VP of comms — I think he's perhaps waiting.

   35:04

   Nathan Labenz: Well, let me play a music video then in the meantime — that'll be the perfect lead-in. And here comes a little—

   35:14

   Prakash: Zeev — Zeev's on.

   35:16

   Nathan Labenz: So, K-pop for you — we don't have to play the whole thing, but here's a little K-pop just to get us warmed up.

   35:55

   Q: Show me. Here we go.

   36:30

   Nathan Labenz: Oh, did my mute — mute the sound as well?

   36:34

   Prakash: So it's interesting — I think when you mute, the sound basically comes through on the production studio. The sound is always going in the livestream, but when you mute it, the sound started to come in through the production studio; when you unmute it, it went off. So that's something to note for the future, I think.

   36:53

   Nathan Labenz: Alright, it wants—

   36:54

   Prakash: —to preserve our chats, I guess. So—

2. 36:59Interview48 min

   Interview: Zeev Farbman — LTX and the Open World-Models BetZFZeev FarbmanZeev Farbman — co-founder and CEO of LTX, one of the five Hebrew University PhD students who founded Lightricks in 2013 and built Facetune into a ~$300M/year profitable business (Calcalist, Jun 2026) — joined five weeks after splitting the company in two (Jun 1, 2026), keeping Facetune as the cash cow while taking ~250 people to chase open world models full-time. The conversation was built around that bet: what made keeping the companies together untenable; the Red Hat-style math of open weights that are free for academics and commercial use under $10M ARR (LTX-2, open-sourced Jan 6, 2026 — native 4K at 50fps with synchronized audio, running on consumer GPUs); LTX-2.3's #1 open-weights ranking on Video Arena (Jun 30, 2026) versus the closed frontier's compute budgets behind Veo and Sora; his thesis that video models are becoming world models with implicit physics, and that physical AI requires weights customers can own because the data that matters lives inside factories, robots, and fleets; his own experience running a coding week on the open GLM-5.2 instead of a frontier closed model; and what responsible deployment means when open video weights can't be recalled — from the founder who once said of misuse guardrails, 'It's hard for me to envision how to create a barrier that AI can't breach' (Calcalist, Sep 2024).

   Watch

   ▶Play this segmentYouTube · from 36:59

   As aired

   Prakash introduced LTX (formerly Lightricks) CEO and co-founder Zeev Farbman, tracing the company's pivot from consumer creativity apps like Facetune and Videoleap to frontier open-weights video/world models, including the 22-billion-parameter LTX-2.3 release and Zeev's role as a founding member of NVIDIA's Cosmos Coalition.

   Zeev framed emerging world models as the natural extension of the LLM 'predict the next token' paradigm to video, sound, and physical action, citing recent work that encodes robot joint states directly into video tokens as displacing the older VLA robotics paradigm.

   Nathan pressed on the tension between physics-grounded world models and fantastical, physics-breaking creative tools (drawing on his own Claude/Suno music-video experiments); Zeev said it's still unclear whether or how to segregate that kind of data between pretraining and post-training.

   The pair dug into LTX's core technical bet — an extremely compressive, variable-token-rate latent space, born from being compute-constrained relative to Google and OpenAI — and where the remaining bottlenecks sit: real-time avatars are nearly production-ready, playable game-length worlds are blocked on context/consistency, and robotics-arm demos are close.

   On business strategy, Zeev argued closed labs are stuck in a self-described 'CapEx trap,' propping up huge valuations with a toll-road API model that Chinese open labs are undercutting; LTX instead offers free use below a $10 million revenue threshold and predictable licensing above it, which also appeals to IP holders (animation/VFX studios) wary of ceding data to Google.

   Nathan and Zeev traded views on how far fine-tuning open weights can go — from IP-specific animation and UGC avatar models to more novel uses like computational photography and using precise physics solvers to teach faster fluid-dynamics simulation — and debated cross-attention/modularity versus deeper, earlier fusion of modalities, including LTX's approach to audio (trained jointly from the start) versus newer token types like robot joints and 3D structure.

   Prakash asked about LTX's research-bet process (the compressive latent space as a necessity-driven strategic bet, validating changes at small scale before touching pretraining) and about the outlook for on-device/mobile video and world models, where Zeev predicted on-device 'orchestrator' routers and a coming reckoning for closed labs' default compute-spending choices.

   The show's AI co-host 'Q' asked Zeev where the biggest gap is between what LTX's tools promise and what they deliver — physics fidelity and fine-grained creative controllability remain the two open problems — before Nathan and Prakash closed the segment and teased more Claude/Suno/LTX music-video experiments.

   Key moments

   The story is kind of simple: the tech is magical, it's hard to doubt it. So okay, if it's magical tech, we should put a huge price tag on it. But when you look at the economic realities, it doesn't always work out like that.

   We sometimes internally call it the CapEx trap — these guys spend so much on the data center, so much on compute, they raise such a crazy amount of money and create such expectations that they try to create a business model that's a toll road.

   Once people start having these local routers that understand the complexity of the problem, I think there's going to be a moment of reckoning for Anthropic and OpenAI — right now they're deciding how much compute you spend, and it benefits them, not you.

   Questions asked

   38:58What has Lightricks/LTX been doing for the last three months — what's taken off?

   Zeev said LTX-2.3 (released about a quarter ago) validated that video models are becoming the backbone of 'world models' — analogous to LLMs predicting the next token, world models predict the next moment including how things look, sound, and what actions are possible. He cited work encoding robot joint states into video tokens as replacing the older VLA robotics paradigm, and previewed an upcoming mixture-of-experts architecture plus a newly cracked variable-token-rate architecture.

   41:58How do you reconcile physics-grounded world models with creativity tools that intentionally break the laws of physics — will video-gen fork into 'grounded' vs. 'imagination' models?

   Zeev said it's unclear whether blending fantastical animation data into pretraining is wise; LTX's current approach mirrors LLMs — pretrain on everything, then use SFT/RL post-training to specialize toward a domain. He noted new research can identify which pretraining samples correlate with good or bad outcomes for a given scenario, which could eventually allow cleaning pretraining data, but said that capability is still a way off.

   45:21Where's the bottleneck — compute, data, or a model-design/latency constraint on inference?

   Zeev said compute is LTX's binding constraint since they're self-funded off mobile-app profits. He expects mixture-of-experts to break through dense-model limits the way it did for LLMs, and said inference efficiency is highly use-case dependent: real-time avatars are already sub-second latency and near production-ready, full game-length interactive worlds are blocked on context/token-consistency limits (his 'coin in the drawer' example), while robotics arm/dexterity use cases are close because they need little context window.

   50:28Given thin inference margins and Google's compute advantage, what's LTX's strategy to capture value while open-sourcing the model?

   Zeev argued closed labs are trying to justify huge valuations with a 'toll road' API model — what he calls the 'CapEx trap' — pointing out Chinese labs (DeepSeek, Moonshot) are close on capability but valued far lower. LTX's alternative: free use under a $10 million revenue threshold, then predictable multi-year licensing deals above it, a model he expects to win as competitors like GLM narrow the capability gap with closed labs.

   55:55How does this play out with IP holders worried about Google and what it might do with their IP?

   Zeev said it's a major business opportunity — LTX offers the pretrained backbone plus post-training tooling so studios can fine-tune on their own IP fully on-prem, which resonates with entertainment/VFX companies worried about competitors training on their in-progress, unreleased styles.

   57:15What can fine-tuning on LTX's open weights actually achieve, and how do users close the gap with (or exceed) closed frontier video models?

   Zeev estimated LTX is roughly a couple of quarters behind closed models on raw scale (their upcoming MOE will be ~100-200B params vs. reported few-hundred-billion closed models), but said fine-tuning headroom is wider than for LLMs. Examples: IP-specific animation/VFX fine-tunes that nail in-betweening at 10-30B params, marketing/UGC avatar models where cost matters more than scale, and more surprising cases — computational-photography adaptations (denoising, dynamic range, focal-length simulation) and using precise physics solvers (e.g., fluid dynamics) as training signal so the model learns to approximate expensive simulations quickly.

   1:03:54Is the audio-conditioned model an example of that fine-tuning, and how does LTX handle conditioning across modalities generally?

   Zeev said audio works best baked into pretraining rather than bolted on afterward — LTX has separate video, audio, and text latent spaces tied together via cross-attention. For newer modalities (robot joint positions, finger pressure, 3D object structure), they're building an easy way for builders to define new token types and train them on top of the pretrained backbone.

   1:06:32Will cross-attention-style modularity remain the dominant way to merge modalities, or will everything end up more deeply fused, as true multimodal 'superintelligence' might require?

   Zeev gave two conflicting data points: LTX found real gains from unfreezing and jointly training the VLM/text-encoder with the world model rather than keeping it frozen and separate (supporting deeper fusion), but cited an Apple research paper finding that 'late fusion' of separately pretrained modality models can work just as well (supporting modularity). His honest read: after seeing the audio and VLM-unfreezing results, deeper/earlier fusion increasingly looks right.

   1:09:55What's your process for making research bets, and when do you change course?

   Zeev distinguished strategic bets born of necessity — e.g., committing early to an extremely compressive, variable-token-rate latent space to compensate for lacking big-lab compute, which created hard technical problems but is now the most compressive latent space in use (NVIDIA's recent model uses something similar) — from tactical bets, where the team prioritizes cheap inference-time tricks and validates any pretraining change on small models before committing at scale.

   1:14:07How close are we to video/world models running on mobile, and what's the ceiling for edge devices?

   Zeev said audio use cases already run well on edge devices; full real-time world models won't run on this hardware generation, with Macs (M5 chips, large VRAM) a notable near-term exception for heavily quantized models. He expects on-device 'orchestrator' routers to decide when a query needs the big cloud model vs. local compute, and predicted a 'reckoning' for OpenAI/Anthropic once users can make that cost/quality tradeoff themselves rather than the labs making it for them by default.

   1:19:11Where's the biggest gap between what LTX's tools reliably deliver today and what creators assume they'll get, and how are you testing against it?

   Zeev named physics fidelity and creative controllability as the two open gaps — even at 200-500B parameters the model isn't capturing full real-world physics yet, and creators want fine-grained, knob-level control (lighting, splash size, etc.) that current models don't decompose well.

   Related

   LTX ↗Zeev Farbman on X (@ZeevFarbman) ↗LTX on X (@ltx_io) ↗Zeev Farbman on LinkedIn ↗"Why Big Tech Is Abandoning Open Source (And Why We Are Doubling Down)" — Farbman's X article (Mar 17, 2026) ↗Calcalist: Lightricks splits in two — Farbman goes all-in on LTX (Jun 1, 2026) ↗Lightricks open-sources LTX-2 (Jan 6, 2026) ↗LTX-2.3 hits #1 open-weights on Video Arena (Jun 30, 2026) ↗Farbman: "benchmarks are checkpoints, not destinations" (Jun 30, 2026) ↗Farbman on coding with GLM-5.2 instead of a closed frontier model (Jun 30, 2026) ↗

   Full transcriptLightly edited · timestamps jump to YouTube

   37:00

   Prakash: Gotcha. And on that note, let me introduce our next guest — our guest for this morning is Zeev Farbman, CEO and co-founder of LTX. Zeev has an extensive background in computational photography, having completed his PhD in computer science at the Hebrew University of Jerusalem before co-founding Lightricks in 2013. You likely know Lightricks as the engine behind wildly popular creative tools like Facetune and Videoleap, which drove the company to a $1.8 billion valuation. However, recognizing that generative AI was fundamentally cannibalizing traditional software, Zeev recently pivoted his company to the absolute frontier of machine learning. Under the new LTX banner, Zeev's team recently released LTX-2.3, a massive 22-billion-parameter open-weights model that generates high-resolution video and fully synchronized audio in a single pass. While competitors like OpenAI and Google lock their video models behind expensive APIs and cloud interfaces, Zeev argues that the future of AI belongs to open, composable models that creators can run locally and fine-tune securely on their own hardware. Beyond creative media, he's also a founding member of NVIDIA's new Cosmos Coalition, pushing beyond just video generation to build open world models designed to teach physical AI and robotics how to interact with the real world. He joins us today to discuss the collapse of closed-software monopolies, the bleeding edge of multimodal AI, and what happens when artificial intelligence finally learns the laws of physics. Zeev, welcome to the show.

   38:54

   Zeev Farbman: Great to be here, guys. Nice to meet you.

   38:58

   Prakash: Zeev, just give us a brief introduction to what Lightricks and LTX have been doing for the last three months — what's taken off?

   39:11

   Zeev Farbman: Okay, wow — that's a big question, because we released LTX-2.3 roughly a quarter ago, and in AI years that feels like a decade. So, a bunch of things. I think there's a growing realization that what started as video models is becoming the backbone of what we now call world models. The best way to explain why this is so powerful is to draw the analogy to LLMs: at their core, LLMs are still predicting the next token and the next word, and when we do pretraining at the scale of the internet, that lets us create models that do textual reasoning incredibly well. Emerging world models are doing something similar — given some boundary conditions, some history, some constraints, they predict the next moment, and that moment includes how the world appears, how it sounds, and what kind of action we can take. I think the action part is the most surprising one. Roughly a quarter ago, maybe a bit more, a paper showed that it's fairly easy to add some kind of encoding of a robot's joints onto video tokens and basically ditch the VLA paradigm that had reigned supreme before it. That's one of the big surprises, and for us, realizing that validated something we all strive for: building extremely efficient models. Once you realize a robot needs to run the simulation thirty times a second, you realize the sheer number of tokens that's going to burn on the simulation. So that was one of the exciting validations of our overall thesis. On architecture — there's a bunch we can go into. We're planning to release soon a mixture-of-experts architecture alongside the dense models we're already shipping, and I think we finally cracked a variable-token-rate architecture, which is exciting — it lets the model invest more tokens where the physics is challenging or where something necessitates more tokens. So, a ton is going on. We're gearing up for our next model release really soon. Busy times.

   41:58

   Nathan Labenz: One big question I have is on world models versus creativity tools — I've been having a lot of fun with this. I don't know if you caught some of the video we played right before you came on, but that's a mix of different video-gen models orchestrated by Claude to make a music video for a song I'm creating with Suno. That's how deep down the rabbit hole I've gotten. But I'm a little confused on world models versus creativity tools, in the sense that creativity tools very often — and certainly in my music videos — are doing fantastical things that are outside of physics. It strikes me that when I create these impossible scenes, I'm operating in an imaginary, counterfactual way that violates the laws of physics. I wonder how you think about balancing that, or whether it leads to a fork in video-gen models — some grounded in physics, perfect for simulation, and others built to let your imagination run.

   43:09

   Zeev Farbman: It's a great question, and it's kind of a no-brainer that if you're blending a ton of animation data into pretraining, that's not a great example of how real-world physics behaves. So the jury's still out on what's best to do in pretraining versus post-training. In the case of LLMs — obviously I have no idea what the closed labs are doing, but judging by what the open players are doing — I think it's pretty clear that during pretraining you want to throw in as much as you can. Obviously, with LLMs you don't become a sophisticated coding model just by reading Reddit conversations, but you do capture something about the nature of grammar, etc. I think right now, with world models, we're trying the same thing: pretrain on as much as you can, then during post-training do SFT and RL on specific domains, similar to how it's done with LLMs. Maybe at some point people will realize there's certain data that's actually harmful for pretraining world models — I don't think we're there yet. This week I saw a really cool paper showing that, given a specific scenario you encode with a prompt, you can find pretraining data samples that correlate well with a positive outcome for that scenario, and samples that don't. Once you have that mechanism, you can start to see what was actually harmful for the scenarios you care about, and maybe clean the pretraining data. But I feel that's still in the future. Right now the paradigm isn't that dissimilar from LLMs — throw everything into pretraining, then post-train to narrow it down to a specific domain task.

   45:21

   Prakash: Where's the bottleneck that you see? Is it compute? Is it data? Is it more of a model-design bottleneck, where you have to squeeze the model into a latency and compute budget for inference? What are the major constraints you have to face?

   45:47

   Zeev Farbman: Our constraint is obviously compute — we're a company that funded development of the model with profits from mobile content-creation apps, so we're definitely compute-constrained relative to the big guys. But if we zoom out and assume more compute capacity — and I'll draw the LLM analogy again — at some point people realized that just increasing the size of a pure dense model has its limits, so you need something else. The same way those guys had a lot of success with mixture-of-experts architectures, we're seeing the same thing with diffusion models and world models. I don't have a good answer yet on to what degree we're data-limited; to figure that out, we need to pretrain our mixture-of-experts model a bit more and see where the limit actually is. As for efficient inference, it really depends on the use case. For real-time avatars or virtual environments, you can take a huge model, distill it to a much smaller parameter count, distill it further to two or four steps, and we're already at latency well below a second for a lot of these use cases — production-ready for some of them. Avatars especially are extremely easy, and we're going to see a ton of them soon: virtual teachers, virtual customer-support agents, etc. To create an actual gaming environment, though, we still have a problem with having enough tokens for world consistency. Think about Genie and similar models — you typically create an autoregressive model with a lot of already-generated tokens sitting in your context window, and that blows up pretty quickly. We're seeing models with maybe thirty to sixty seconds of context — still not enough for an actual game. And the brute-force compression methods we've used so far, like sub-sampling tokens, aren't really robust. Imagine generating an environment — my room, say — and I open a drawer and there's a small coin. You'd expect that when you leave and come back and open the same drawer, you'd still see the same coin in the same place. But that coin is just a tiny token that was generated, and we don't yet have a system that knows how to compress the whole context while preserving those critical details. So although we do have real-time models that can do these things, the context is still missing. I don't think we're going to have actual games running on these systems in the next quarter or two. As for robotics, a lot of robotics use cases don't actually require a ton of context window — think robotic arms and dexterity, where the whole context is right in front of you. If a robot needs to figure out how to make a sandwich, everything's in front of it, so latency and autoregressive models — we already have those parts. So you're going to start seeing demos of robotic arms doing things fairly quickly in the next quarter or two. So far, if you look at a lot of these videos, they're actually sped up — it looks like the robot is doing something cool with its arms, but the speed's been extended. I think that part is mostly solved.

   50:28

   Nathan Labenz: So how does this all translate into strategy for you? A couple of things I've heard — one, obviously there's this neo-lab-versus-frontier-lab, big-tech competitive dynamic, to use your habit of analogizing to LLMs, where with video models you're going up against Google, and they've got something like 25% of global compute in-house. The summary of the rumors I've heard from the video-gen space is that the models themselves are really expensive to run under the hood, and margins are pretty slim across the board — which maybe explains open-sourcing, because if you're not making much on inference anyway, there's not so much cost to giving it away. But it strains the traditional finance brain to look at a business like yours and think: okay, we're going to put all our profits into making this next-generation thing, and then give it away. What's the play for you to capture some of that value for yourselves?

   51:45

   Zeev Farbman: Great question, there's a lot to unpack here. A bit of background: the reason we started building our own foundation models is that what closed model providers offer just doesn't make economic sense for us. At Lightricks, we're a mobile creativity company — we wanted AI models that could run on edge devices without spending on an inference call at all, and at some point we realized no one else cared about building models like that. When we tried to work with closed model providers and serve our customer base that way, we realized it was completely prohibitive. That's when we decided to build an extremely efficient architecture — most of it boils down to creating an extremely compressive latent space, where video is represented by a small number of tokens, plus a variable token rate on top of that. Long story short — if you go with the closed-source providers, I'll draw the LLM analogy again: OpenAI and Anthropic are trying to justify trillion-dollar valuations, and the story is simple — the tech is magical, hard to doubt, so you put a huge price tag on it. But when you actually look at the economic realities, it doesn't always work out that way. There are plenty of examples where a service is extremely valuable but very hard to monetize — firefighting is a great economics-course example: people used to try running firefighters as private companies, it was really hard to extract value that way, so eventually governments decided to own it. Now we have this interesting situation where Chinese labs — DeepSeek, Moonshot, etc. — really aren't that far behind the frontier, but if you look at their valuations at their last round, we're talking tens of billions, maybe around fifty — no one's talking trillions. Same underlying tech, so what's going on? Internally we sometimes call this the CapEx trap — these companies spend so much on data centers and compute, raise such a crazy amount of money, and create such expectations, that they end up building a toll-road business model: every time you touch their model, you pay them. Maybe that could have worked in the past, but given the availability of Chinese models, I just don't see how that unfolds the same way going forward. So in the world of world models, we're providing the alternative for people who don't want a toll-road business model. We say: if you're not hitting a $10 million revenue threshold, use the model for free — build something cool, get some traction — and once you hit that threshold, let's discuss licensing, which can be a predictable multi-year deal so you can manage your costs. It's obvious why the big guys don't want to do this — this model is way less lucrative economically than a toll road. But our claim is that the toll road isn't going to be a viable long-term model, because once you offer a different, more win-win business model, more people are going to switch. GLM is a great recent example — once you start closing the gap with closed models' capabilities, a lot of people suddenly start thinking about cost.

   55:55

   Prakash: How do you see this play out with IP holders, who to some extent are very scared of Google and what they might do with their IP? How does this work out for you?

   56:11

   Zeev Farbman: Obviously it's one of our great business opportunities. Part of our pitch is: we're giving you not just the pretrained backbone, but also the post-training infrastructure, and you can train it on your own IP, on-prem. People in the entertainment industry — animation, VFX — are starting to realize the importance of their data. Sometimes what happens is they have multi-year animation or VFX projects where they're developing something new, and the last thing they want is for someone to train a model on that new, unique style before the project is even finished. So I feel it's a huge opportunity for us. A lot of the licensing deals we've done so far have been around data privacy, where people want to do the whole thing on-prem — people understand much better now how important their data, especially new and unique data, really is.

   57:15

   Nathan Labenz: Can you point to a couple of examples of people who've done interesting post-training on your open-source foundation? I'm actually curious about doing this myself — what you can really drive with fine-tuning, and at what scale. One of the big fundamental facts about these video-gen models is that with greater scale they get higher resolution, better intuitive physics — the sixth-finger, disappearing-limbs kind of problems become much less frequent, if not gone entirely. We don't know all the specs of the latest Omni-type models, but it's clear they've scaled further and have sharper intuitive physics — I don't think that's too contentious, tell me if you see it differently. But if you give your model — which is smaller scale and somewhat behind on those dimensions — to a company that has its own data and use case and wants to fine-tune, in what ways can they close the gap? And in what ways can they actually exceed what's possible with closed-source models, based on that fine-tuning opportunity?

   58:50

   Zeev Farbman: In terms of parameter count, my understanding is that the recent closed models are on the order of a couple hundred billion parameters — I haven't heard of a world model hitting a trillion parameters yet. I don't think open source is going to be that far behind; we're planning to release an MOE model around 100 to 200 billion parameters. So the gap is probably going to look like the LLM gap — maybe two or three quarters behind. But in this world, adaptation matters more, and the range there is wider than with LLMs. A few examples: the first that comes to mind is VFX and animation for a specific IP — if you have a specific franchise with a lot of data across several seasons, fine-tuning and focusing all the model's capacity on that IP is extremely beneficial. That works so well that for certain use cases — think keyframe animation — animators still want to do the keyframes themselves, that's the creative part they don't want to outsource. But the in-betweening has historically been the crazy expensive part of the animation P&L, and models in the 10-, 20-, 30-billion-parameter range, fine-tuned for that specific task, are good enough — at that point it's really just a matter of cost. Another example: marketing and advertisement use cases, like creating UGC with an avatar model — that also doesn't require a trillion parameters. Once you want a personal teacher or an avatar you're going to run for hours a day, you don't want to pay top-tier CDN prices for that, so it's all about efficiency. Across a lot of use cases, once you fine-tune to a specific domain, cost becomes very important, because you've already cleared the quality bar — then it's all about cost. There are some more unusual fine-tuning cases I've seen. One: computational photography — taking sensor data and implementing something like denoising, taking low-light video and producing a clean video, or creating higher-dynamic-range video because the sensor's dynamic range is limited and you lose detail in highlights or shadows. Or simulating a different focal length from a camera stream. There are a ton of these use cases that people don't normally associate with generative models but that actually run like that — people take existing data, sometimes footage from two cameras that are close together but with different focal lengths, and do the adaptation on top of the model. That's a very unusual adaptation, and surprisingly, you typically don't need a crazy amount of data for it. Even more surprising: people adapting these models to do simulations that used to require really expensive solvers — think computational fluid dynamics, how water or smoke moves. People are adapting these models so that a precise solver does the equation — which takes a ton of time — and then using that as input to the model, so the model can do the simulation fairly quickly. So circling back to fine-tuning: the range is wider than with LLMs, you sometimes don't need a ton of data for these adaptations, and I think that's part of the argument for why this model should be open — there are a lot of different pockets of physical data out there, and you want the model to be able to really excel at each one.

   1:03:54

   Nathan Labenz: Is your audio-conditioned product an example of that kind of fine-tuning? And could you talk a bit in general about conditioning on different modalities of data? I'm always really interested in these cross-modality bridges and the degree to which these latent spaces can ultimately be merged.

   1:04:18

   Zeev Farbman: Initially, with video models, people tried to bolt audio on top. After experimenting with that for a while, we realized audio is better handled during pretraining. So in LTX right now, you have video tokens and audio tokens, each with their own latent space, plus a text encoder, and a cross-attention mechanism that learns to work across these three modalities. What we're seeing now is that people doing robotics need some kind of streamlined mechanism to play with new types of tokens — for example, when you're talking about the positions of a robot's hands, some people hack it by putting in individual tokens you can read back after rendering, others try to create a completely new stream, same with finger pressure. We realized at some point that part of what we should build into the post-training architecture is an easy way to define new types of tokens, so people can come up with different ideas. There's a project right now where people are trying to define a token that encodes the 3D structure of objects — a vertex, how it moves over time, and the final render — to better capture the structure of the object, so you don't get weird morphing artifacts. So audio, at the moment, is really native — you need to pretrain it a lot jointly with video to get a perfect lip sync. With other modalities, we're trying to build a mechanism that makes it easy for builders to define their own kind of tokens and train them on top of a pretrained backbone, and see where that gets them.

   1:06:32

   Nathan Labenz: So do you think it'll be this kind of cross-attention approach? It certainly has the advantage of being more modular — when you're putting out a foundation model and enabling others to bring their own modality, setting them up to do a good cross-attention makes sense, because you've set them up for success. Do you think that'll be the dominant paradigm, or — at least from what we're told and what the demos seem to suggest from the Omni-type concept — it seems like there's already a deep integration of text and image, or text and pixel latent space, at this point. My suspicion is that what a superintelligence would look like requires these things to be more deeply merged than arm's-length tool calls, or even cross-attention, though that's more debatable. Where do you think we'll get the highest performance, and where do you think most of the action will be? Those may be the same, or they may be different.

   1:07:40

   Zeev Farbman: Great question — and I'm going to give you two completely conflicting data points, so feel free to draw whatever conclusion you want. When we were training world models, or image models in the past, we'd typically start with a text encoder — people used something like CLIP or T5 — and gradually we realized we needed a much stronger understanding of the textual world, so we adopted VLLMs. Initially you'd just take an off-the-shelf VLLM — Google's Gemma or one of the Qwen models or whatever. But what we realized over time is that it really makes sense to unfreeze the VLM and train it jointly with the world model, because you really want the text encoder — the LLM — to speak the same language as the rest of the model, so everything ties together. That actually goes nicely with your intuition that different modalities should start speaking the same language really early on — it's the same thing we saw with audio, where bolting audio onto video after the fact makes lip-syncing hard. On the other hand, there's a paper — I think from one of the Apple research teams — that compared what they called early fusion and late fusion: starting with dedicated models that each only see one type of token, versus fusing early between modalities. Their conclusion was actually that you can do a really good job with late fusion. So that's a point for modularity. Where exactly it lands, I have to say — initially I thought late fusion might work. After what we saw with audio, and after what we saw with unfreezing VLLMs during training, your intuition might be closer to correct.

   1:09:55

   Prakash: What's your process for assessing research bets you have to make? How far ahead do you have to think? When do you change course after getting surprised? I imagine you have to make these bets and could get burned making the wrong ones — so what's the process?

   1:10:20

   Zeev Farbman: Great question. I'd say there are really strategic bets and a lot of tactical bets. On the strategic side, for us it was first of all necessity — we knew we weren't going to have the scale of compute the big guys have, so we needed to figure out how to get somewhere with the compute we had. The first decision was, again, an extremely compressive latent space. Each token, in the case of world models, represents some patch of pixels — how a certain patch looks over time — and you have to decide how big that patch is. If one token represents a really big patch, you lose a lot of information, so you need to compensate somehow. We made a bet a long time ago that we'd eventually figure out a variable token rate — meaning you could start with a compressed core structure and add tokens on the fly to compensate. The best analogy is video: most world models today operate like constant-bit-rate video compression, allocating the same number of tokens to hard parts and easy parts. In video terms, that sounds laughable — most video-streaming algorithms use variable bit rate, allocating more bits to complex regions. So we placed these big bets — compressive latent space, variable token rate — and that created a whole host of technical problems that people with less compressive latent spaces don't have, like differentiability issues, etc. But it was a constraint, we had to do it, so we did. I think we cracked a lot of those technical problems — the LTX VAE is still the most compressive latent space that people actually use; NVIDIA's recent world model was presented on top of a similar latent space. So strategic bets come from necessity — we just can't spend the same compute as the big guys, so we need an algorithmic edge. On the tactical side, you're always trying to figure out the most reward for the least effort. There are a ton of cool inference-time ideas that don't require training or post-training, so you obviously prioritize those — there's a lot you can squeeze out of that. And you're most careful about changing anything in pretraining, because that's the most expensive part — before touching pretraining, you try to show something on a small-scale model first, working up through model sizes. But again, that's also part of the story of necessity — we can't be as cavalier with compute as the big guys, so we have to think more before placing these bets.

   1:14:07

   Nathan Labenz: How close do you think we are to models working on mobile? I spoke not long ago to Ramin — the CEO of Liquid AI, who you've probably crossed paths with — and I was really struck by something that's an obvious fact but also good calibration: the global smartphone and laptop market is $800 billion a year, and has been for a number of years, so we're only now crossing that level with the global data-center build-out. There's a lot of compute distributed in people's pockets — I can get a couple-billion-parameter model running on my phone at an acceptable speed, but I'm not going to use it that much because it's clearly not as smart as what I get out of the cloud. What do you think the trajectory looks like for video and world models? Do you see a limit to how much you could squeeze onto an edge device? And how fast do you think we get to viable use cases that run entirely locally?

   1:15:21

   Zeev Farbman: That's an excellent question and an excellent observation. I think right now, when people think about the whole CapEx dynamic and where the spending goes, they don't often ask the very justified question of what's actually possible to run on edge devices, because that changes the economics of the whole thing. Before we get to a full world model, let's decouple modalities — a lot of audio-related use cases can already run on edge devices. Right now we're still living in a world where some of that goes through APIs, and some people have built significant businesses on it — I think that's going to be obsolete pretty soon. I personally recently invested in a startup that showed me a demo of really cool audio stuff running on the edge device. I think what's coming is hybrid rendering models, where you put some kind of orchestrator on the edge device that understands the complexity of the problem — let's go to the world of LLMs for a second, it's kind of funny: we hear about all these benchmarks, Erdős problems being solved, etc., but 99% of actual LLM use cases have nothing to do with Erdős problems, and we're spending a lot of electricity on it anyway. So I think there are going to be orchestrators that understand what you actually need and try to address it on the edge device, and if not, go to the bigger model in the data center. Running an actual real-time world model on an edge device — that's not going to happen in this generation of hardware. But the notable exception, again, is Macs. Not an iPhone, but if you look at M5 performance — I saw someone do a crazy quantization of GLM running on an M5 MacBook Pro, you need a ton of VRAM, maybe 512 gigs or something like that, but it's becoming plausible, and I don't think we've squeezed everything possible yet, in terms of both optimization and quantization. The fact that I can run that stuff on my Mac — come on, that's unbelievable, it's super cool. So that's one of the things being underpriced right now — how much compute will be able to move to edge devices. And once people start having local routers that understand the complexity of a problem and make that decision for you, I think there's going to be a moment of reckoning for Anthropic and OpenAI — because right now they're the ones deciding how much compute you spend, and that benefits them, not necessarily you.

   1:18:37

   Nathan Labenz: Prakash, I don't hear you — but that might only be me.

   1:18:42

   Zeev Farbman: He also can't hear you.

   1:18:46

   Nathan Labenz: Alright, come back. Okay — here's a gimmick, might be a cool one, might be a lame one: we have Q, our AI co-host, with us today too. Q, would you like to ask a question? Any blind spots that—

   1:19:00

   Q: Stand by in the control room, and ready when you are.

   1:19:03

   Nathan Labenz: Yeah — ask a question based on what we've discussed and what we might have missed.

   1:19:11

   Q: Zeev, one area that often gets less airtime is failure modes in the creative pipeline. Where do you see the biggest gap between what your tools can reliably deliver today and what creators assume they'll get, and how are you testing against that?

   1:19:29

   Zeev Farbman: Great question — the gap between what we promise and what we deliver. Listen, there are clearly still some gaps. I think the major one is physics — we're trying to capture the entire physics of the universe, or at least a parallel universe, with a model that's maybe 200 to 500 billion parameters, and we're not there yet, though we're closing the gap pretty quickly. I'd say most creators are still going to point to the fact that the simulation isn't as correct, or as controllable, as they want it to be. If you're talking to really creative people, they want to control every nuance of the appearance, and that requires somehow decomposing the model into a bunch of knobs — like in classical software, where you can say, 'I want two more lights here,' or 'I want the splash to be bigger.' Achieving that controllability, alongside physics, is one of the open problems — we get cool things, not necessarily exactly the things creators want, and that's a real pain point.

   1:20:50

   Nathan Labenz: Good question — thanks, Q. Anything else you think is important that we haven't touched on? I appreciate you staying with us a bit long — I don't want to take up your whole afternoon or evening, but this has been a great conversation. Anything you feel like we, and Q, have missed?

   1:21:12

   Zeev Farbman: No, that's a great question. I think the edge-compute angle is underpriced — people should spend more time thinking about that. And your intuition about when's the right time during training to fuse modalities — that's something I personally find really interesting, and we've spent a ton of time on it. Frankly, we discussed a ton of cool and important stuff today.

   1:21:43

   Nathan Labenz: Cool — well, I love it. Thanks for joining us today on AI in the AM, and we'll definitely be watching your progress and continuing to make music videos with it. Who knows what we might be able to do, especially if we dabble in a little fine-tuning as well — the world's our oyster. We appreciate the opportunity to hack on an open-source model to make that the case.

   1:22:09

   Zeev Farbman: It was a pleasure being here. Thank you very much, Nathan.

   1:22:12

   Nathan Labenz: Appreciate it. Great to meet you. Bye for now.

   1:22:16

   Prakash: Bye bye.

   1:22:22

   Nathan Labenz: To be here — alright. In the interim, while Prakash hopefully comes back, I'm going to go ahead and play another music video made with a combination of Claude and Suno for the song, Omni for some of the scenes, and LTX for the lip-sync portion. Here we go — hopefully we can hear this.

   1:23:02

   Nathan Labenz: And just quick context, I guess, in case this seems super weird — this was from the episode with Ramin Hassani, the CEO of Liquid AI, and this is inspired by the transcript of that conversation.

3. 1:25:04Segment74 min

   Predictions & FutureSearch: Hosts vs. the AI SuperforecastersThe rematch pitched on Monday's Dan Schwarz episode, and the longest segment of the day: a 'Guess the Market' round in which Nathan and Prakash guessed the live market-implied probabilities on AI prediction-market questions — each sharpened to its exact resolution criteria — scored themselves against the markets, and compared their calibration to FutureSearch's AI forecasting agent, the scaffold that beats the superforecaster median on ForecastBench.

   Watch

   ▶Play this segmentYouTube · from 1:25:04

   As aired

   Nathan and Prakash open with a riff on Meta's new Muse Spark image model and speculation that Facebook/Instagram will eventually wire video generation into a 'dopamine engine' — plus a tangent on AI-driven identity/likeness ownership (Sora-style cameo permissions, multi-person image composition, and Nathan's story of using Gemini's Nano Banana to dress his kids in period clothing on a family road trip).

   They launch the 'Guess the Market' round, bringing the AI cohost 'Q' in as a third contestant alongside FutureSearch's AI forecasts — Q runs on GPT-5.5 with reasoning set high.

   Question 1 — will Anthropic's valuation hit $1.7 trillion by 12/31/2026: Q guessed 12%, Nathan 50%, Prakash 85%, versus the market's 45.5% and FutureSearch's 37%. Prakash cites Anthropic's IPO quiet period and a same-morning SemiAnalysis report pegging it near $6 trillion with $1 billion in quarterly net profit.

   Question 2 — will Anthropic announce an IPO before 1/1/2027: Q guessed 68%, Prakash 60%, versus the market's 64% and FutureSearch's 70%. Question 3 (largest 2026 IPO by market cap, among SpaceX/Anthropic/OpenAI/field) is then posed, with only Q's guess revealed before the part ends (SpaceX 45%, Anthropic 25%, OpenAI 20%, field 10%) — Nathan notes Q's numbers seem out of date.

   Nathan and Prakash ran several "Guess the Market" rounds — Anthropic/OpenAI/SpaceX IPO odds, best coding model of 2026, FrontierMath crossing 90%, and a 1,560 Coding Arena score — comparing their own percentage guesses against live prediction-market prices and FutureSearch's AI-generated forecasts, with "Q" (a third AI contestant added to the quiz) periodically chiming in and needing to be reminded to stay quiet unless addressed.

   On whether the US enacts a federal AI safety law before 2027, both hosts converged around 10%, doubting the Trump administration or Congress could reach consensus — though Prakash pulled up a Punch Bowl News survey showing senior congressional staffers ranking "losing control of AI" a top-three long-term concern (35%), while FutureSearch scored the question far higher (68%), partly reasoning the Defense Authorization Act had already triggered it, a claim Prakash pushed back on.

   Prakash used the coding-model discussion to predict an imminent consumer AI takeoff driven by voice (citing Sam Altman's new ChatGPT voice update) and Meta's image/video generation tools, arguing enterprise AI adoption has already happened but a comparable consumer bump is still pending.

   A recurring theme: FutureSearch's calls (96%, 78%, 68%) repeatedly diverged sharply from both the market and the hosts' own intuitions, prompting recurring skepticism about its rationale even as both hosts acknowledged it tracks the market more closely than their guesses do.

   Nathan and Prakash worked through the final stretch of "Guess the Market" questions — whether Sam Altman is out as OpenAI CEO before 2027, whether a humanoid robot reaches Mars before a human by 2035, whether the AI 2027 scenario holds up through January 2027, whether an AI wins IMO gold in 2026, and whether an AI-created song cracks the Billboard Hot 100 top 20 before 2027 — each time giving their own probability, guessing what the market would say, then comparing to the live market odds and to FutureSearch's forecast.

   FutureSearch surprised both hosts repeatedly with lower numbers than expected (9% on humanoid-on-Mars, 20% on AI 2027 holding up, 19% on IMO gold, 14% on an AI song charting), and several mismatches traced back to strict resolution criteria (e.g., IMO gold excluding independent lab announcements) rather than capability doubts. Nathan noted FutureSearch pulled ahead of both hosts on points by the end of the round.

   Closing the segment, Nathan said he wasn't sure what to take away from the exercise — he and Prakash tracked each other closely on guesses but were both surprised by the market and FutureSearch calibration — while Prakash argued the market is systematically "behind" and that paying close attention to resolution-criteria "rules lawyering," not just the underlying question, is where real edge comes from in prediction markets.

   The pair also dug into a stale Metaculus "weak general AI" question whose sub-criteria (Winograd Schema, Turing test via the Loebner Silver Prize) have effectively broken as useful benchmarks, then drifted into China's new ban on AI companion apps for minors as the conversation wound down, cutting off mid-thought as the show heads into the Closing.

   Key moments

   The ultimate goal is not a video gen. The ultimate goal is dopamine gen.

   My expectation has always been that the voice internet is gonna be enormous — absolutely enormous.

   FutureSearch also called it exactly 20% — FutureSearch is really running away with the points now.

   What we covered

   Guess the Market: hosts calibrate against live prediction-market odds. Question by question, the hosts committed to probability guesses on AI questions before the market number was revealed, then scored the gaps — the show's recurring test of whether following the field all day actually produces calibration.

   The FutureSearch comparison: can the hosts hang with the AI superforecasters? Following the July 6 interview with FutureSearch CEO Dan Schwarz — four days after Scott Alexander's 'The AI Superforecasters Are Here' (ACX, Jul 2, 2026) — the hosts stacked their guesses against FutureSearch's AI forecasting agent, whose scaffold beat the superforecaster median on ForecastBench.

   Related

   FutureSearch ↗ACX: The AI Superforecasters Are Here (Jul 2, 2026) ↗ForecastBench live leaderboard ↗

   Full transcriptLightly edited · timestamps jump to YouTube

   1:26:15

   Nathan Labenz: Alright — that's Liquid Intelligence, made by a whole host of AIs. Great conversation with Ziva. I'm genuinely interested in exploring what it would take — for Waymark, where I'm still only a part-time AI R&D advisor these days — to do commercial-quality work with video-gen models. Q actually asked a good question, you might have heard this: I gave Q the chance to ask a question, and it asked about the disconnect between what creators think they're going to get and what they actually get. Q answered in a way that really resonates with the Waymark experience, which is that controllability isn't quite at the level real creative minds demand, or at least crave. For somebody like me who's just throwing music videos together on the fly, if it delights me, that's a win. But if you're trying to represent a small business in a commercial context, you've got to get it real and right, and the controllability there is still fairly hard. So I came away from that conversation thinking about what we could do if we got into fine-tuning — not necessarily moving fully away from asking 'what is the best model,' which has traditionally been the approach, but broadening our thinking to also ask what's the best model we could help create if we did a little fine-tuning on top of something like LTX. That would be interesting to find out.

   1:28:04

   Prakash: It strikes me — well, Facebook has just done Muse Spark, their next-generation image model, and the next-generation video model is coming out soon. One of the commentaries online is that Facebook is going to wire this to the likes and the viewing — who likes what — and basically, this is what we've expected from the beginning: dopamine-wire humanity to a model and have that model spit out things that cause dopamine and cause people to keep coming back to Facebook. That's maybe a signal that no other firm has as much as Facebook — I think Facebook and TikTok are the two. And the video gen is just the starting point for the dopamine gen — it's kind of a prerequisite to the dopamine generator, but the dopamine generator is coming. That's the ultimate goal. The ultimate goal is not a video gen. The ultimate goal is dopamine gen.

   1:29:25

   Nathan Labenz: Mixed feelings on that, to put it mildly, but—

   1:29:30

   Prakash: Everyone is scared of Meta — when Meta decides to go for it, because people in the industry kind of know what's possible, but it's not what most researchers have focused on. Having the best researchers focus on that is something that's scary, I think. But people haven't been talking about it for a while — it's not exactly a new idea.

   1:30:07

   Nathan Labenz: Well, we've seen a version of it. The 4o sycophancy — the 'sycophantocalypse' — was at least in part driven by, naive would probably be harsh, but a less sophisticated than it needed to be use of the thumbs-up/thumbs-down signal in ChatGPT that fed back into the post-training process and created the sycophancy problem. I thought that was a really interesting moment — it showed how much we still don't know, and the fact that we can be surprised post-launch in such a meaningful way. I guess you'd put that down largely to the differences between how people use models at OpenAI internally and how the rest of the world interacts with them, and they've presumably closed that blind spot significantly since then. But it was also telling that they took it down. OpenAI is clearly interested in serving users and customers, but they're not fully trapped in the 'whatever the customer demands is right' paradigm — you still go online, and in many spaces you'll see people complaining about 4o. You go into Sam Altman's comments any given day and there are still people saying 'free 4o.' They walked it back a little when they first took it down, but with the highly sycophantic one, they basically said this is too much for us, we're not going to do it. Broadly, I've been pleasantly surprised by how much resistance the frontier companies and the community as a whole have shown to a sort of obvious wireheading opportunity — and I'm a little scared of it. As you said, everybody's scared of Meta — count me in on that roster. It does seem kind of inevitable, if only at the community-distributed level, but it seems like Meta may bring it to us first — they've become the closest. We've seen examples from them like the 'hot stepmom' chat character — what are you doing there, Meta? I don't know that we need hot stepmom on the roster of our AI avatars. But it could get pretty crazy with the scale and compute they have — if they really try to optimize that way, watch out, it could get pretty gnarly pretty fast.

   1:33:06

   Prakash: I think I saw yesterday — Alex Wang said the first thing they're using Muse Spark for is Instagram video fixes: red-eye reduction, flash optimization, stuff like that. Which is very fair, right — Instagram, after all, started off as a filter company.

   1:33:29

   Nathan Labenz: Filters. Yeah.

   1:33:30

   Prakash: Yeah, so it's very fair — you're basically helping people take better photos, essentially. I think that's fair, and that was the first thing he put out. The second thing he put out was ads — visual ad creation with this thing. So that's when it started to be, okay, now I see where you're going. And then he put out one where you can use multiple images in the same composition and chat with it, reason about it. So one of the things he did was put three photos of three people into an image, ask for a certain style, and ask them to be doing something — and he ended up with three people sitting on a bench, anime style, talking about something. And Nikita Bier — who's a PM at X right now and was at Meta for a while — pointed out that one of the big bottlenecks to social on Instagram had been that you need to get people together to take these photos. If you want a nice photo in front of the pyramids, you've got to go to the pyramids, and all of you have to go. That was one of the problems — that's a bottleneck. And I think this is the lifting of that bottleneck. We started to see it earlier — I'm almost like, why do people take selfies with celebrities anymore? It doesn't make sense — autographs don't make sense, those selfies don't make sense anymore, because you can just fake them. People are still taking the selfies, but it doesn't make sense anymore. I think this is one of those cross-the-Rubicon moments, when the tech goes out to enough people and is easy enough to use that this kind of taking photos with celebs maybe goes away. But it does mean you can kind of hang out with your friends online, and you don't have to be ostracized for not going to the party because you don't have the photos of yourself there — you can just insert yourself into those photos anytime. Meta might also be able to do something about ownership of identity, because they know more about visual, facial identity and recognition than anyone else on Earth — they have, for at least a decade now. So they could probably do a bit about 'this is your image, you own it on your Insta, and no one else can use photos from your Insta as an input into theirs without you being a friend and allowing it.' And if they make it easy enough, it might become the default place — if it's easy enough, you don't need to go elsewhere for piracy, you can just do it on the Meta platform. And if it's not allowed, it's too difficult to bother with, and that could actually be helpful in some sense. So — I'm not sure—

   1:36:41

   Nathan Labenz: —a pretty good model for that, I thought. The way they allowed you to control who could cameo you into videos — I think that was a pretty good precedent for them to set, even though obviously that app didn't rise to their level of expectations, or they needed the compute for other things. I do think that model made a lot of sense. It was remarkable that they didn't really have much downstream trouble — it seemed like they actually controlled what you could and couldn't do in a way that... I don't know, it's almost like they pleased everyone for a moment. I didn't hear people complaining that it was too restrictive, and I also didn't hear people complaining that their likeness was being abused — they really had a pretty good own-your-own-likeness paradigm. I thought that was pretty cool. One funny story that came to mind when you were talking about faking it: I took a road trip with the kids not too long ago and stopped at a 1760s fort, part of the French and Indian War or whatever. My wife took a couple photos of the kids, and then she started using Gemini — Nano Banana in the Gemini app — to dress the kids in period clothes in these photos. They came out great. She turned my five-year-old into a British regular, and it was like, damn, that's really good — he still looks like him, it's super cute. My upshot to her was, I don't think we need to buy the Christmas jammies anymore — we could just do it in post, take a picture, and put whatever Christmas jammies you want on them.

   1:38:41

   Prakash: I'm sure that—

   1:38:45

   Nathan Labenz: —goes to show, yeah. I had understood a lot of the value from the Christmas jammies was in the photos, but she did not like that idea — so I think the Christmas jammies will still be ordered this coming Christmas. Whatever we might be able to do with the photos after the fact is kind of beside the point, I guess, from her perspective.

   1:39:12

   Prakash: Oh, hilarious.

   1:39:13

   Nathan Labenz: Do you want to do you and me versus each other, and versus FutureSearch, trying to forecast what's going to happen on some AI questions? Yeah, let's go for it today.

   1:39:25

   Prakash: And let's have Q participate. Q, you're going to be participating with us in this contest.

   1:39:32

   Nathan Labenz: Okay, I don't actually have an answer slot for Q, so I could probably—

   1:39:39

   Prakash: Oh, we can just get her broken in.

   1:39:42

   Nathan Labenz: Let me — here's the thing, we can just start kind of talking, and — there's that. Now let me go do a quick prompt. Gotta find, out of my many, many tabs here, which one I need to be in.

   1:40:16

   Q: Alright, let me get my footing for a moment, and then I'll jump in. Standing by, quite ready. When you're set, give me the prompt and the scoring rules, and I'll keep my answers crisp so we can focus on the comparison.

   1:40:30

   Nathan Labenz: Nice, nice, nice — FutureSearch, okay, cool. Session as-is. Can you add a third-player queue to the quiz? This should be fast enough to do — I'll build a refresh, we'll get the queue in there in just a minute. But in the meantime, we can start with the conversation. So, I think we did change the scoring a little bit since last time — now you can see this: points is 100 minus the square of (guess minus market), over 100. Whether that's the optimal scoring rule or not, I don't know, but there you have it. Let's get into it. Will Anthropic's valuation hit $1.7 trillion by 12/31/2026? We have to give a percentage guess that it will do that, and our source is the NPM price reported by the Nasdaq Private Market.

   1:41:47

   Prakash: Okay, so by the end of this year, will it hit $1.7 trillion? And we have to give a percentage, or a yes or no?

   1:41:57

   Nathan Labenz: Yeah, it's a percentage, I believe.

   1:42:00

   Prakash: Okay, so I would say it's 80% yes for me.

   1:42:09

   Nathan Labenz: Wow, okay — that's higher than I would have said. Let me go first, Q, and then you can come in last — I don't want you to bias me. Oh, wait, did we lose the screen share? Can you still see it?

   1:42:31

   Prakash: I can't see the screen share.

   1:42:33

   Nathan Labenz: There we go.

   1:42:34

   Prakash: So, Q, go ahead — what do you think? Will Anthropic hit a $1.7 trillion valuation by the end of this year, 2026?

   1:42:46

   Q: Alright, let me think this through for a moment and put a number on it. I'll put it at about 12%. That's a sky-high valuation, and while Anthropic is moving fast, getting to that level by 2026 would require exceptional revenue, capital, and market conditions all lining up. Not impossible, just very unlikely.

   1:43:08

   Prakash: Okay, very well.

   1:43:09

   Nathan Labenz: So that was 12% — is that what Q said?

   1:43:16

   Prakash: Yes, Q — was that 12%?

   1:43:19

   Nathan Labenz: Yeah, okay — I would've been high with you up until this whole Fable moment. That does make me wonder if there's some risk there, and also we've got these equity-sharing proposals that are a little more live now than they were even just a few weeks ago. So if we'd done this even just a few weeks ago, I'd have probably been up there in the 80% range, but today I'm a little lower because of these other risks — so I'll maybe say 50%. So we've got quite the range covered here: 12, 50, and you said 80, right?

   1:44:10

   Prakash: Yeah, 85.

   1:44:12

   Nathan Labenz: Okay, in the meantime this update to the game is almost there, but I'll let it finish. So — 50 for me, 80 for you, and FutureSearch, which — we talked to Dan, the CEO there, earlier this week, and as we discussed at the time, I just had Claude create an account, get an API key, and run these exact same questions through FutureSearch to get its take as well. Whether we should believe FutureSearch more or less than the market is a very live question, but we can at least do the compare-and-contrast. Okay, here we go — let's reveal the first one. Market says 45.5%. FutureSearch says 37%.

   1:45:04

   Prakash: I feel like I have some privileged information here. Number one, Anthropic is already in its quiet period for the pre-IPO — they've done the IPO filing, and they're in the quiet period, so they're not talking about a lot of stuff. Number two, this morning SemiAnalysis published a piece calling Anthropic probably a $6 trillion company pretty soon, and saying they'd hit $1 billion in net profit last quarter — so that's cleared their internal model-training-adjusted number, cleared EBITDA, cleared net profit — full cash flow plus profit. So I think they're in that kind of period, and I'm pretty sure the valuation is basically doubled already. That's my guess — we'll see.

   1:46:25

   Nathan Labenz: Yeah, the fact that they're already making money is pretty incredible. Was there a revenue number? I was just thinking, it's been a minute since we've heard the latest monthly revenue update — do you think that's because of the quiet period?

   1:46:38

   Prakash: Exactly — the leaks have stopped because of the quiet period. The leaks were happening because they were doing a roadshow, talking to investors; now they're in the quiet period, they're not talking to investors.

   1:46:50

   Nathan Labenz: Gotcha. Alright, on to question number two — this is from Kalshi. Will Anthropic officially announce an IPO before January 1, 2027? Resolves yes if they confirm an IPO before then, where confirmed means the SEC declares the S-1 form effective, the IPO is priced, or an exchange assigns it a ticker. I don't have a great sense of where that happens in the process relative to when it actually happens. Go ahead, Q — you take the lead on this one.

   1:47:34

   Q: Around 68%. If they're already in the quiet period, the mechanics suggest forward motion, but timing can slip with market conditions and regulatory steps.

   1:47:46

   Prakash: I was at 60%, probably.

   1:47:51

   Nathan Labenz: Okay.

   1:47:52

   Q: Got it — I had 60%. That's close enough that we're basically in the same camp: likely, but not a sure thing. The path is open, but the calendar can still be tricky.

   1:48:03

   Prakash: Alright, Q — we'll ask you for your opinion, so just stand by.

   1:48:08

   Nathan Labenz: Yeah, I'll go a little lower again, just because of these risks — though even as I say that, I'll go a little lower than you, but not too much lower. Let's see what we've got. Market says 64%. Yeah — we've got to have a word with Claude here, I don't know if you saw, but it did kind of leak the answer a little bit. FutureSearch says 70%.

   1:48:33

   Prakash: Q is hooked up on the back end to GPT-5.5 for more difficult questions. So, Q — did you refer to an external advisor or tool call for that last answer?

   1:48:50

   Q: Let me clarify that briefly — no, that was my own estimate, based on general reasoning about how IPO timelines usually unfold and the uncertainty around market conditions. I didn't use any external advisor or tools for it.

   1:49:05

   Prakash: Okay, thank you, Q. Q also has a reasoning setup — it's on reasoning-high, and it's been told to reason through. So this is a much more intelligent model than what we had just a day ago.

   1:49:23

   Nathan Labenz: That's cool. I should also say, I didn't control at all what data sources FutureSearch can use — so presumably, given these questions, it looked at the market, anchored itself to the market, and maybe also found reason to depart from the current state of the market. Number three: which company will complete the largest IPO of 2026 by market cap? So, completes an IPO and achieves the highest market cap at the close of its first trading day.

   1:49:59

   Prakash: And the options are — SpaceX, Anthropic, OpenAI, and—

   1:50:04

   Nathan Labenz: —the field.

   1:50:08

   Q: Alright — let me think this through for a moment and decide.

   1:50:13

   Prakash: And it's a percentage for each, I think?

   1:50:16

   Nathan Labenz: Yeah — percentage chance that each of the three companies, plus the field, is the one with the highest close on its day of IPO.

   1:50:27

   Prakash: Okay, so those choices are—

   1:50:29

   Q: SpaceX, about 45%, if it actually IPOs — it's the heavyweight. Anthropic, maybe 25%. OpenAI, around 20%. And the field gets the remaining 10%, for a surprise giant sneaking in. The uncertainty is mostly about whether SpaceX lists at all.

   1:50:51

   Nathan Labenz: Okay — you're a little out of date there, Q.

   1:50:53

   Prakash: Is listed.

   1:50:56

   Q: Fair point — I should have said 'if it's listed' more carefully. If SpaceX is off the table, then I'd shift that weight mostly to the field and a bit to Anthropic. The big unknown here isn't just who's strongest, but who actually chooses to go public.

   1:51:12

   Prakash: Okay, thank you, Q.

   1:51:17

   Nathan Labenz: Okay, yeah — this one's interesting. I'm gonna go super low on the field; I don't see any other candidates really. SpaceX was what — at the end of their first day, like 0.8, and then it popped up again a little more the next day, over 2, and then maybe dropped down a little bit.

   1:51:35

   Prakash: I think you're right.

   1:51:36

   Nathan Labenz: Yeah, you bet. So Anthropic — we just covered it, I gave it a 50%-ish to IPO and basically a 50%-ish to get over that level. So that gives me, by synthesis of the last two questions, a 25 there. And I've got a 25, maybe a similar analysis on OpenAI, a little lower for them. The remainder goes to SpaceX — leader in the clubhouse, not getting surpassed.

   1:52:10

   Prakash: Okay, so for myself: 0% for the field, 10% OpenAI, 30% Anthropic — SpaceX gets the rest, I guess.

   1:52:33

   Q: Actually, let me think this through and put a clean split on it: I'd go with...

   1:52:42

   Prakash: Q, only respond when you're addressed...

   1:52:44

   Q: 50% SpaceX and 10% for the field, and Anthropic the rest.

   1:52:49

   Prakash: Okay, Q — only respond when you're addressed.

   1:52:55

   Nathan Labenz: Alright, let's see.

   1:52:58

   Q: Quiet — I'll stay quiet unless someone says 'Q.' Whoa.

   1:53:02

   Nathan Labenz: So that's interesting — FutureSearch isn't giving much weight to the field at all. FutureSearch's call: 'SpaceX has already established a virtually insurmountable benchmark.' I don't know about that, FutureSearch, but it is closer to the market than we are.

   1:53:23

   Prakash: Well — Bayesian probability, right? SpaceX is already listed, so that clears a huge—

   1:53:31

   Nathan Labenz: Yeah, the leader-in-the-clubhouse effect is definitely real, but 96% seems quite high for FutureSearch there.

   1:53:46

   Prakash: I think—

   1:53:47

   Nathan Labenz: I don't know, I have a hard time wrapping my head around that. Let's see the full agent rationale.

   1:53:53

   Q: I'll stay quiet unless someone calls on Q, or addresses Q directly.

   1:53:58

   Nathan Labenz: Thank you — yes. Now we're getting into it; starting to feel like that TikTok feed.

   1:54:04

   Prakash: The initial Q setup resets — so we reset Q.

   1:54:15

   Nathan Labenz: So first-day closing market cap was approximately $2.1 trillion, according to FutureSearch — which, yeah, that's high, but I don't know, I wouldn't put these companies out of the woods yet. People are definitely gonna wanna buy a lot of Anthropic.

   1:54:30

   Prakash: So, Sam Altman just announced an updated voice for ChatGPT, and I think this is gonna be similar to what we're using right now. My expectation has always been that the voice internet is gonna be enormous — absolutely enormous. And I think it's gonna tie in with the AI devices we've been talking about for a while. So I think we're about to see consumer takeoff, and consumer takeoff is gonna be Meta's video gen, image and video gen, and AI devices and AI companions on voice. So far we've had enterprise takeoff — enterprise takeoff has been in coding. But consumer takeoff — we had the ChatGPT bump, but we haven't had another bump yet. I think they needed a few things: they needed the advertising model, and I think they've started to evolve that advertising model, and they're almost there. I think maybe what ended up happening is OpenAI didn't wanna do the dopamine route they had with 4o, so they took a step back on the advertising and tried to figure out what they could do — and I think they ended up with voice. This is gonna be interesting, because if voice works out, it's gonna be enormous. Absolutely enormous.

   1:56:00

   Nathan Labenz: Yeah, no doubt about that. Alright, moving right along — from Kelsey: 'Which AI company will have the best coding model at the end of 2026?' We basically have the same contestants. Notice this is LiveBench.ai, contamination-resistant. Okay, well, I think again I'm gonna go real low on the field, and at this point I think we have to go fairly low on Google too. The rest — you gotta put the bulk of the weight on Anthropic, OpenAI, and xAI. I wouldn't count out xAI; they're a dark horse, but they're dangerous. I'm gonna go 60 for Anthropic and pull that out of OpenAI and the others. Alright — what do you think?

   1:57:20

   Prakash: I would say Anthropic 60, OpenAI 30, xAI maybe 5%, Google maybe 3%, and whatever's left for the field on best coding model.

   1:57:44

   Nathan Labenz: What do the market and FutureSearch say? Oh — we're right on with the market.

   1:57:52

   Prakash: Right on — 60% Anthropic.

   1:57:56

   Nathan Labenz: Again, this field thing — I feel like this is a fairly well-documented phenomenon in prediction markets generally, where somehow the field always gets a little more weight than it seems like it should. FutureSearch, again, I'm not really sure what to make of that.

   1:58:17

   Q: Let me—

   1:58:18

   Nathan Labenz: It's not a crazy number, other than the field.

   1:58:21

   Q: The field often gets a bit of a premium because it captures unknown entrants and surprise performance, not just today's leaderboard. It's a hedge against overconfidence in any one lab. As for FutureSearch, I'd treat it as a useful signal, but only one input — especially if the rationale isn't fully transparent.

   1:58:41

   Prakash: Very well, Q — thank you.

   1:58:45

   Nathan Labenz: No — in terms of the FutureSearch rationale... oh, okay, well, FutureSearch has a compelling point here. If OpenAI is the incumbent on this metric as of July 2026 — it holds the top two positions with GPT-5.2 Codex and GPT-5.5 Thinking XHigh — then, you know, that's fairly compelling compared to just my broad intuition. Although, again, we're a lot closer to the market than FutureSearch is. But I can see the argument that if they're currently number one, maybe they stay number one. That's not a crazy idea by any means. Alright, let's keep going. Will an AI model score greater than 90% on FrontierMath before 2027? I feel like this almost already happened, didn't it — before the end of—

   1:59:45

   Prakash: Before 2027?

   1:59:47

   Nathan Labenz: Oh — before 2027. So, by December 31st this year. But I feel like that's pretty high, because I feel like it's almost already happened. It might be slightly—

   1:59:58

   Prakash: Didn't FrontierMath have to be revised? Because a lot of the questions had errors or something.

   2:00:07

   Nathan Labenz: Yeah, there was a revision, and the scores got better as a result. And even on tier 4 — this doesn't specifically say tier 4, but there was a score in the eighties. So I'd say this is like a 95% chance based on everything I know. It seems like it's almost already happened; maybe it could go even a little higher, but I don't think it's quite literally happened yet. There's a little bit of weirdness there — maybe I'll go up to 97.

   2:00:44

   Prakash: For 90%, I'd say maybe 75%. I feel like in order to get to 90%, they'd have to revise a lot of the problems — they say they have issues with a third of the problems. So they had to revise fatal errors in about a third of the problems.

   2:01:12

   Nathan Labenz: It's wild — I'd score an absolute zero on that. So, you talked me down two points, but I'm still keeping it at 95, because if memory serves we're already well into the eighties. There could be some rules-layering stuff here around exactly what dataset or whatever, but my sense is we're already very much on the doorstep of that. So — let's see.

   2:01:34

   Prakash: Alright — 88.

   2:01:35

   Nathan Labenz: Market says 88. FutureSearch says 78.

   2:01:40

   Prakash: Okay.

   2:01:47

   Nathan Labenz: I don't know if there's too much more to say there — just look at the rationale for a second. What was the most recent? Yeah, GPT-5.5: 87.7 on tiers 1 through 3. And Fable: 87 on 1 through 3, and 88 on tier 4. That's what I was remembering. So, if anything, I feel like I'm probably a little low — and that would mean, obviously, everybody else is even lower. But I don't know — if you're at 88 on tier 4 today, with a model that was done training in February or whatever, I don't see how you don't get to 90%. This one, I think we may have a leak — we have to talk to Fable again about not leaking the answers, no peeking at the leak. Will any model reach a 1,560 Coding Arena score by 12/31/2026? This is a tough one, because I don't know what the current top score is, and it's a little hard to know what these scores mean in isolation — relative to, without actually looking at the leaderboard to get a sense. I think this is a tough question — maybe we skip this one. We also kinda have a leak here — it says LM Arena resolved. I think when it's resolved already, that probably means it happened; I don't think it means it definitely isn't gonna happen. So we'll put us both down for 99%. What do you say — we'll tie on this one? What — market says 46? I'm confused now. Well, let's look at the thing real quick and see what we're looking at.

   2:03:39

   Q: Let me check for a moment so we can line it up clearly. Resolved usually means the entry has been finalized and linked to a specific model, not that the question itself is already settled. So a 46 from the market can still be perfectly sensible if the leaderboard's confirmed, but the target hasn't been reached. In other words, it's about the status of the entry, not a hint about whether the future event happened.

   2:04:06

   Nathan Labenz: I see — well, I'm misreading my own game, I guess. It's interesting that they've got 46 in the top spot. Every time we end up on Arena, I'm a little bit like, what is going on here? You've got Fable ranked worse than, I don't know, something's—

   2:04:36

   Prakash: This is the thing, right? Internally, the companies aren't really using these benchmarks — except as kind of a taste test — they're building around other metrics internally. They're dogfooding their own stuff. So it's clear that they're high-taste coders.

   2:05:05

   Nathan Labenz: Yeah, something weird about this — I have to say I'm not too compelled. But okay, onward we march. Number seven: will the US enact a federal AI safety law before 2027? Resolves yes only if a bill is signed into federal law, and it has to contain at least one of: a prohibition on creating or releasing specific AI systems, training restrictions, usage restrictions, or human-in-the-loop requirements. Executive orders and state laws don't count.

   2:05:41

   Prakash: I'd say there's, like, a 10% maybe.

   2:05:46

   Nathan Labenz: Yeah, I was kind of in the same zone.

   2:05:50

   Prakash: I don't see the Trump administration wanting to, and I don't see them able to get consensus on a bill anyway.

   2:05:58

   Nathan Labenz: Yeah, and we've got midterms, and then we'll have a little lame-duck period — there's a lot of reasons I think it's gonna be pretty tough. Ten feels intuitively high, but this is where I think it's the same argument as 'the field' — we should be prepared for some surprises. So it's ten in my mind because there's some chance of surprises, some incident, some kind of vibe-shift moment that could really change things. Barring that, I think it's quite a bit less. Most of my 10% comes from a world where there's some major event, slash vibe shift, that takes us into a very different political reality than we're in today. Although — there was this one thing recently, let me see if I can find it real quick, it was really notable. I'm gonna have to ask Grok — can you find the tweet about the congressional staffer survey on most important issues? See if we can find this... oh no, Grok's unable to reply, Grok is offline. I'll have to—

   2:07:19

   Prakash: Can you search X for a tweet about congressional staffers and their views on AI federal law?

   2:07:35

   Q: Let me check what's been posted about that and pull the key takeaway. I'm not seeing a clear hit in a recent post I can access — it may be buried in a thread, phrased differently, or just not widely shared yet. If you have a name, a handle, or a specific phrase from the post, I can try again. Otherwise, the takeaway is that staffer sentiment can shift quickly, especially if a big incident or political moment changes the mood.

   2:08:01

   Prakash: Okay — thank you, Q.

   2:08:04

   Nathan Labenz: Grok will be able to find it, but Grok's offline at the moment — so, anyway, we'll come back to it. But the observation was — I believe it was a survey of congressional staff — that showed AI issues creeping into, I believe, even the top five most important issues from their perspective. Which doesn't mean, obviously, that it's going to get done, but I think it at least creates conditions where — it sounds like congressional staff have got enough scary demos that they're ripe for a vibe shift, at least. Or, or, we'll see.

   2:08:45

   Prakash: Or, you know, people are targeting money, right? Congressional staffers are angling for money and positions afterward, anyway. So this has started to become like — hey, there's money on both sides of the issue. They're being lobbied too — you gotta remember all these firms have big lobbying teams. So even if it's not in your top five issues, you're being told about it all the time. You're having these soirées, these parties, going out, and everyone's talking about it — and the guy who invited you is like, 'hey, this is what OpenAI is doing, this is what Anthropic is doing, this is what Google is doing.' It's kind of inescapable — the amount of money going in.

   2:09:37

   Nathan Labenz: I'm just scrolling through my recent Twitter likes real quick to see if I can pull it up.

   2:09:42

   Prakash: Okay — losing control of AI: 35%, ranking as the number-three overall long-term challenge. This is from Punch Bowl — let me pull it up. I'll share my screen... there we go. So, among the most important problems to solve in the next 250 years, senior staffers pointed to the national debt and federal deficit (51%), political polarization (45%), and losing control of AI (35%).

   2:10:25

   Nathan Labenz: The 250-year timeline — there might be some work being done there.

   2:10:28

   Q: Losing control of AI in the same conversation as structural issues like debt and polarization — it doesn't guarantee legislation, but it suggests staffers are primed for a serious policy push if the political winds shift.

   2:10:41

   Prakash: Indeed, Q — please, speak only when you're spoken to. Thank you.

   2:10:48

   Nathan Labenz: Alright, let's go back to the reveal — so we're both at 10%. Market says 20. FutureSearch called it 68. Whoa — 'there's a very strong case this outcome has already been triggered: the Defense Authorization Act.'

   2:11:16

   Prakash: It's — this is supposed to be an AI safety bill, not the Defense Authorization Act. They just slide some AI bits into it — that doesn't... yeah, I don't think that counts. This is one of those resolution-criteria things.

   2:11:31

   Nathan Labenz: Yeah, well — FutureSearch is definitely way off from the market there. FutureSearch is still somehow outscoring us in aggregate.

   2:11:42

   Prakash: Do you correct for overconfidence bias when you make predictions?

   2:11:48

   Nathan Labenz: Not in the classic way that some Bayesian scoring rules do — we're just penalizing the difference between our guess and the market, without the kind of extreme penalty some scoring methods use, where guessing zero percent would carry an infinite penalty. We could look at rescoring this; I don't think it's obvious we're doing the best thing, or even obvious what the best thing would be, but it's certainly not obviously right either. Okay — will Sam Altman be out as OpenAI CEO before 2027? I have to say, it seems pretty unlikely these days. His position seems pretty unassailable barring some crazy scandal. I'd put that at one percent.

   2:12:49

   Prakash: I'd say it's 20%.

   2:12:52

   Nathan Labenz: 20%?

   2:12:53

   Prakash: Yeah.

   2:12:54

   Nathan Labenz: Well, you know what — I'm going to change mine, only because I know the market itself has a hard time getting down to 1%. So I'll put it at 3%. I'm guessing the market comes in around 4 or 5.

   2:13:04

   Prakash: I'm guessing—

   2:13:05

   Nathan Labenz: But I think it's actually even lower than that. So tell me — what's the story for 20%? That's a high number.

   2:13:13

   Prakash: I'm not sure he really wants to continue as CEO. I saw him bringing in Fidji Simo as kind of preparation to hand over the reins — because do you really want to take the company public? Do you want to be a public-company CEO, undergo that scrutiny, be on earnings calls? Those are the questions he hasn't had to answer, and I don't think he really wants that heat. He's not getting paid for it either. So I feel like he'd rather be in the chairman position, like Bret Taylor is. Especially once the company goes public, a lot of the dealmaking becomes very regular — capital markets, bonds, equities, you raise money and pump it through. There's not that much of the behind-closed-doors dealmaking anymore. So that's my guess — he may not want to fight for it.

   2:14:18

   Nathan Labenz: Let's find out what the market says. 11%. Wow.

   2:14:23

   Prakash: So they're giving it 1 in 9.

   2:14:26

   Nathan Labenz: 7.

   2:14:27

   Prakash: And giving me a 1 in 5. You gave a 1 in 33.

   2:14:33

   Nathan Labenz: Yeah, that is surprisingly high to me. Okay, onward we go — will a humanoid robot walk on Mars before a human does, by 2035?

   2:14:49

   Prakash: Oh, honestly — 100%. No, 90%. 90%.

   2:14:58

   Nathan Labenz: Yeah, rovers don't count, to be clear — it's got to be a humanoid. Sending robots ahead of us certainly makes sense given that SpaceX, or at least via Tesla, is in the humanoid business. That also makes a lot of sense. And we have a full decade for it to happen. I think I'm up there with you — it seems like if it doesn't happen, we've probably shifted to another timeline, where it's like, are we not launching much to space anymore? Are we not going to Mars anymore? What happened that put us in a timeline where we're not sending a humanoid to Mars in the next ten years? I think there's a real question of whether Kelsey will even be around to pay out in those scenarios. I'd answer this almost like an x-risk question — if we're sending stuff to space, if we're generally on the happy timeline, it feels overwhelmingly likely. And if it's not happening, I feel like we're probably in a pretty unappealing world. So you kind of can't price in the chances that civilization gets wiped out and there's no SpaceX. I think I'm with you, although I'd flag — I'll say 90%. You said even higher, right?

   2:16:58

   Prakash: I said 90. I said 90.

   2:17:00

   Nathan Labenz: 90 also? Yeah, I think the actual chance is lower, but I'm calibrating myself to Kelsey — it's conditional on Kelsey still being around to resolve it, and I think that actually takes out a lot of the chances this wouldn't happen. Fifty percent. FutureSearch calls it 9%. What is going on? Wow — SpaceX's own ambitions have slipped well into the 2030s. 9%, that's really something.

   2:17:40

   Prakash: I think this is where I'd say models can't predict the future based on the past — it's just hard for them to forecast anything that isn't an in-distribution event. This is definitely out of distribution.

   2:18:06

   Nathan Labenz: Yeah, although maybe we're wrong and it's right. But we agree on this, right? The only viable pathway relies on SpaceX — that's what's striking, we basically said the same sentence. It's going to happen via SpaceX, with Tesla.

   2:18:22

   Prakash: Oh, I don't know if it's going to be SpaceX, but it's probably going to be SpaceX. The Chinese are only a little bit behind — maybe about 5 years. They have a reusable rocket now, and being the second mover, they can move faster than the first mover — they can steal some IP, hire some engineers, and they know it can be done. Knowing it can be done is actually a huge thing — like the nuclear weapon: the Russians knew the Americans had built one, and that let them move toward it much faster. So even if SpaceX doesn't make it, I think another entity will. And you might not even need reusable rockets, because humanoid robots don't need to come back the way humans do. I suspect it'll be like a railgun with a humanoid-robot-factory assembly machine that just crash-lands there — you can send dozens of them, no food requirements, you just need one to wake up, turn on the solar panels, power up, and unpack. We've almost got the manual dexterity done — I've seen demonstrations where they can fold clothes, which means they can probably start to assemble things. So I think we're almost there — it's just the combination of the pieces plus the budget. And you get five cycles to Mars, basically — five windows, every 18 months.

   2:20:22

   Nathan Labenz: Fascinating. Well, I still — FutureSearch and the market both make me feel like I'm overconfident, but 9% definitely feels really low. Okay, on to number 10: will the AI 2027 report's predictions be borne out by January '27? That resolves yes if the scenario has been roughly correct through January '27 — details can miss, but the important through-lines should be correct. Resolution is a poll of Manifold moderators; 30-70% splits resolve proportionally. Oh my god, this is turtles all the way down — we're asked to

   2:21:05

   Prakash: A bit nuts.

   2:21:06

   Nathan Labenz: assess whether Manifold moderators will agree that the scenario has been roughly correct. I think it's on path to be roughly correct — the meme recently, I'm sure you've seen it, has been people manically flipping to the Q3-2026 part of AI 2027 to see what comes next.

   2:21:39

   Prakash: I'm reading through AI 2027 — it's tough, because there are actually two paths. AI 2027 ends in October 2027: Agent-4 is live, a misalignment memo leaks to the New York Times, the public is suspicious, foreign allies are outraged to realize they've been placated with glimpses of obsolete models, and so on. That October 2027 beat has basically already happened — it happened in the last month or so with the Mythos-Fable thing, about two or three months ahead of schedule. At that point AI 2027 branches into slowdown or race. In the race path, by December, Agent-5 is live with charismatic avatars and super-persuasion, and it effectively takes over the government by the end of 2027. In slowdown, they invoke the Defense Production Act, the government takes over OpenBrain, and they head toward a US-China deal by the end of 2027. So those are the two paths.

   2:23:03

   Nathan Labenz: But note that for this question, we're only through January 2027.

   2:23:08

   Prakash: Only through January — well, I'd say it's already been achieved. By January 2027: AI takes some jobs, Agent-2 never finishes learning,

   2:23:25

   Nathan Labenz: continual learning comes online. This is a tough question.

   2:23:32

   Prakash: We're judging the resolution criteria at this point, right? We're not judging the underlying feature. Yeah.

   2:23:39

   Nathan Labenz: So, I'll go a little higher — I'll say 65%. Actually, let me go 70%. I feel like the general vibe is that they've been prescient, and my guess is people will broadly say history judges them as having shown strong foresight — I think we can be reasonably confident on that. But Manifold moderators, I wouldn't say those are easily pleased people. Are they going to be difficult to impress, do a bunch of rules-lawyering? I'm deep in the theory of mind of Manifold moderators here, but there's some real risk of a 'no' answer even if, by my lights, the answer would be yes. So I don't know where that should leave me, but I'm putting down 70%. What do you say?

   2:24:36

   Prakash: 50%.

   2:24:40

   Nathan Labenz: And the market says 20%. FutureSearch also called it exactly 20% — FutureSearch is really running away with the points now. Yeah, I don't know, I don't think there's much more to say — that is a low number. The core case there missed on delayed takeoff. Well, I don't know, I'm not sure takeoff is all that delayed. Okay, there we go — two bonus questions. This first one, I think we already know the answer: will an AI win an IMO gold in 2026? What is happening here?

   2:25:39

   Prakash: It's 2026.

   2:25:41

   Q: Let me think this through for a second. That's a fun one.

   2:25:45

   Nathan Labenz: Yeah, I think we should expect this to be yes.

   2:25:54

   Prakash: 80% yes? 90% yes?

   2:25:57

   Nathan Labenz: Yeah — we got an IMO gold last year, right, with 5 out of 6 questions? So it'd be pretty shocking if we didn't get a gold this year. I'm going to say 95%. What did you say?

   2:26:10

   Prakash: Call it, I don't know, 90%.

   2:26:14

   Nathan Labenz: Yeah, unless I'm misunderstanding this, it happened last year, and we usually don't go backwards. Let's see what the market says — 68%. And FutureSearch says 19%. Wait a second.

   2:26:28

   Prakash: Maybe it's the resolution criteria again — I think some of the rules require a limited amount of compute, or an open-source model.

   2:26:43

   Q: Let me pull that up.

   2:26:45

   Nathan Labenz: Excluding independent lab announcements.

   2:26:55

   Q: Right, that explains the mismatch. If the rules exclude lab announcements or require specific constraints — open source, limited compute, competition conditions — then prior wins don't automatically count. It's not saying the capability isn't there; it's just a narrower, more formal target.

   2:27:14

   Nathan Labenz: Yeah, and FutureSearch says the IMO Grand Challenge is effectively dormant — so I think that follows from the question. There were other resolution criteria at play, but that's a strange one. Okay, last one from Manifold this time: will an AI-created song crack the Billboard Hot 100 top 20 before 2027? Oh, this is relevant — from your discussion earlier, it sounded like we're maybe already getting close with some of that K-pop stuff.

   2:27:54

   Prakash: I thought we were already there — it depends on the resolution criteria, but there was already an AI artist on Billboard. I'm not sure it was the top 100.

   2:28:09

   Nathan Labenz: Interesting — largely created by AI.

   2:28:12

   Q: I'd treat that as very plausible, but still dependent on the fine print. If the rules allow songs substantially AI-written or produced with human performance or production layered on top, then yes, it could break into the top 20. But if they require something fully AI-generated end to end with no human involvement, that's a much tougher bar — and could explain why the markets are skeptical.

   2:28:42

   Prakash: Let me pull this up — I'm on the Billboard page, and six songs have shown up on various charts, though maybe not the main Billboard chart. Most notably, a fully AI-generated country track, 'Walk My Walk' by the AI act Breaking Rust, hit number one on the Country Digital Song Sales chart, and Xania Monét has also charted on the Billboard Radio Airplay and R&B lists. So they're there, but I think on genre charts rather than the main chart.

   2:29:21

   Nathan Labenz: I had a really funny experience talking to my brother about my AI music. I told him, 'you should listen to some of the stuff I've made, it's really good,' and he said he'd never listen to it — basically categorical opposition. He works at Google, codes with Gemini all day every day, so it's not that he's boycotting AI in general, but his reaction to the music was pretty disinterested — just, 'this should be something humans do for humans,' and that was it from his perspective. I don't know — the fact that there's already been one in the country category suggests it'll probably happen before too long, but this one's a little hard to predict. I'll say 60%, because one song in the top 20 doesn't seem that crazy.

   2:30:30

   Prakash: I'll probably go 75. My expectation is that it's right now something like 25 — these things seem to be behind on all the numbers. So, yeah, I'll go 75.

   2:30:45

   Nathan Labenz: Market says 21%. You called the market. FutureSearch says 14.

   2:30:51

   Prakash: Yeah, they seem to have this kind of six-months-ago view of the market — if you calibrate for that, this is roughly where the market is. It's what I call living in the future. But the trick is, you have to live in the future while also calibrating how far ahead you are of the market — because if you don't, and you make an investment, you end up getting run over by the market being too slow. You have to calibrate your investment time period to how far ahead of the market you actually are.

   2:31:29

   Nathan Labenz: Well, I don't know what to make of today's exercise, honestly. I was quite surprised by a lot of the FutureSearch answers — it does beat us on points, at least as we're scoring it here. You and I had very similar answers to a lot of things and very similar point totals, and we both had moments where we said what we think the true answer is, then what we think the market answer will be, which is how we're being scored. So I'm not sure what the takeaway is — the individual questions are interesting, but at the FutureSearch level and the market-calibration level, I was surprised to the point where something feels kind of off somewhere. And is that on the AI side, the market side, or just you and me? I'm not really sure how to boil this down into a takeaway. Would you like to attempt it?

   2:32:30

   Prakash: I think we're all trying to make sense of this information together. I think the market is behind, and you have to not fear the fact that you're ahead of the market — that's okay. I also think there's a lot of rules-lawyering going on, and that's been the issue with prediction markets from the beginning — you really have to pay attention to the rules and how these things resolve, and people who do tend to make better bets. A lot of the money being made is actually from amateurs who just read the question and don't read the resolution criteria.

   2:33:14

   Nathan Labenz: Yeah — there was a Metaculus one I used to look at. Let's look at it real quick and then we can wrap up for today. 'When will the first —' is this the one? I think it might be slightly different, there's a related question, pretty similar but a bit distinct. Let me go back. Oh, this is more the synthetic one — okay, yeah. This one's been interesting to watch, and I think the resolution criteria are instructive. So: 'when will the first weak general AI system be devised, tested, and publicly announced?' This got down to a low of early 2026 — I think that might have been the low, as of late 2023, it was expected early 2026. In other words, six months ago already. But when you look into the resolution criteria, they're quite problematic. The Winograd Schema piece has been passed a long time ago — that's laughably simple by today's model standards, just pronoun disambiguation, stuff GPT-4 could probably have satisfied. Scoring 75% on the SAT, too — if not GPT-4, then something in that realm could do it. Learning the game Montezuma's Revenge — I don't know how recently that's been tried, but Claude Plays Pokémon is probably a comparable challenge, and we're probably kind of there on that too. But the piece about passing a Turing test of the type that would win the Loebner Silver Prize — when you dig into those rules, it all comes down to: you have to undergo adversarial testing, the format is a standard Turing test, a human judge simultaneously holds text conversations with a computer and a human, and basically there are just too many clues. It's too easy to tell an AI from a human — not because of things the AI can't do that a human can, but because, as we've talked about before, a human will readily say 'I don't know what you're talking about,' whereas an AI never will. That makes it very easy for an expert judge to tell when they're talking to an LLM. So I think this whole question has kind of gone haywire — I don't think there's much meaning left in it. Intuitively, everybody kind of agrees we've passed the Turing test as a useful way to evaluate these systems — that's now a historical frame. But it turns out this subtle miss, that the AIs are actually too broadly knowledgeable to pass as human, is what's keeping these questions from resolving the way they probably should — or the way the original authors, years ago, would probably have intuitively thought they would, if you'd told them what today's AIs look like. They'd say, 'yes, that passes, this would be a yes.' But we just didn't write the questions quite the right way.

   2:37:48

   Prakash: It's also because they've been banned from imitating humans — every model creator, at least the closed-source ones, has been trying not to let them imitate humans, to make sure people understand these aren't human. Unless — oh, did I just hear China has banned a bunch of companies from doing AI companions? That just dropped, like, 24 to 48 hours ago. They're going to have much stronger restrictions — you can have AI companions for older people, that's allowed, but for younger people they're banning it. You can't make them too attractive, you can't give them a dating life, a love life — a bunch of very specific restrictions. And you know how they are, they're going to stamp it out, control the app stores, stamp it out. One of the big verticals — Chinese companies have made a lot of money on dating and romantic AI chatbots and games, I think they're called 'Moe Moe' games, you'll see them when you go there — those have been banned. So let's see how that lands, or whether people just VPN outside and keep trying to use these things.

   2:39:27

   Nathan Labenz: There's also been talk — and this is probably too speculative —

4. 2:39:31Closing3 min

   ClosingA quick wrap on a nearly three-hour show: scores tallied from the market round, and a look ahead to GPT-5.6's public launch the next morning, July 9.

   Watch

   ▶Play this segmentYouTube · from 2:39:31

   As aired

   Nathan and Prakash wrapped up by noting reports that the Chinese government may stop letting its frontier labs open-source models the way they have been, with Prakash adding that Chinese firms are increasingly exposed to North Korean hacking as AI capabilities diffuse to less-capitalized players.

   Nathan previewed his upcoming multi-week trip to China, where he hopes to get a local read on attitudes toward AGI, open source, and government control.

   Prakash closed the show by thanking viewers, noting a strong combined audience of nearly 3,000 across both live streams.

   Full transcriptLightly edited · timestamps jump to YouTube

   2:39:32

   Nathan Labenz: ...to say much smart about it yet, but in the same 48-hour time frame, there's been talk of potentially not letting Chinese frontier companies open-source their models the way they have been. That could be a huge change to the game board — not a particularly surprising one to me. If anything, I've been more surprised by how long the Chinese government has been willing to just let it rip. I'm much less surprised that they're actually taking an interest now and saying, 'Maybe we should have a little more control' — we like control here in the Chinese government. So we'll see how that takes shape, or what gets caught up in it or not, but—

   2:40:26

   Prakash: It turns out they're as exposed to North Korean hackers as the US is — even more so, actually. There are no real sanctions against North Korea, and they're an ally, and the North Koreans don't care as long as they get money. So it seems North Korean illegal activity is pretty significant for them too, and their IT systems are weaker because they didn't have as much of a honeypot with Bitcoin ransoms as the US did. It seems that as capabilities diffuse, the less-capitalized players are actually more at risk than the more-capitalized players, who've had to handle these things first — because the less-capitalized players always thought, 'No one's targeting me,' and now they are.

   2:41:19

   Nathan Labenz: Well, that'll definitely be a topic of conversation on the ground in China, I'm sure. I'll be there for the next couple weeks and definitely very excited to go. It'll be an adventure in many ways, and I hope to come back with some — very local and parochial, I'm sure, but nevertheless — some sense of the vibe and the expectations, what AGI means to people in China, and what their attitudes are on open source versus government control. There's so much to learn, and I'll try to learn as much as I can in the next couple weeks. Then, certainly, we'll have a lot to talk about when I get back. Anything else you want to cover today before we break?

   2:42:07

   Prakash: No, let's break. And to all the viewers — we had a lot of viewers today. We got up to 2K-plus on my stream, maybe 800 on your stream, so almost 3K total. I think we're finally seeing some pickup. We hope to see you again — bye-bye.

   2:42:30

   Nathan Labenz: Thanks for being with us on AI:AM. Later.

   2:42:33

   Prakash: Thank you. Bye-bye.