Video · 2026-07-08 · 2h 43m · 18 moments

LTX CEO on Video Gen + Can we beat AI superforecasters?

✦ AI generated

GUEST Nathan GUEST Pash GUEST Zeve Farman

timeline · colored by role

Anecdote

K-pop Demon Hunters was assembled 'slop'-style — anime first, music written after, performers hired last — yet became a massive hit, leaving Netflix owning an unexpectedly valuable, evergreen franchise almost by accident.

Pash recounts how K-pop Demon Hunters was made by scripting the anime first, writing the songs after, and hiring performers last — a process he calls 'slopified' — yet it became a massive hit that left Netflix in possession of an unexpectedly valuable franchise.

transcript

Pash: the breakout out hit of the last uh you know 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 because uh they went ahead with the cartoon the anime first and the anime required music because it was about K-pop and then they went and they wrote the music first and then you know they went and they found performers after that to perform the music.

gives example · 1

Claim

Because AI model release cycles are now faster than the time it takes to properly test a model on very long-horizon tasks, labs genuinely don't know whether their current frontier models have topped out before the next one ships.

Nathan relays Noam Brown's concern that iteration speed has outpaced the time needed to test long-horizon performance, meaning labs may not know a model's true ceiling, and floats the idea of a model 'recall' program if long-running tests later reveal problems.

transcript

Nathan: he's like you might need to give these things a month or a couple months or you know what happens if you spend a million dollars with one of these models. He's he basically says we we don't know really if they top out and just calendar-wise from the time we're kind of done training it to our release time is enough to run a lot of standard tests but it's not enough to really do these long standing tests.

extends · 2provides context · 1

Claim

Because model iteration cycles are now shorter than the time it takes to test a model's true ceiling on long-running tasks, frontier labs genuinely don't know if today's models 'top out,' which is why OpenAI's Noam Brown has floated something like a model recall program.

Nathan relays Noam Brown's point that the pace of model releases has outrun the time needed to test models on very long-running tasks, so labs don't actually know if a model has topped out — raising the idea of a post-release model recall program.

transcript

Nathan: he basically says we we don't know really if they top out and just calendar-wise from the time we're kind of done training it to our release time is enough to run a lot of standard tests but it's not enough to really do these long standing tests

extends · 1

Claim

Because model iteration cycles have become shorter than the time needed to run truly long-horizon tests, AI labs genuinely don't know whether their models have topped out in capability, which could justify a post-release 'recall' testing program.

Nathan relays Noam Brown's argument that the pace of new model releases now outstrips the time it takes to discover a model's true long-horizon performance ceiling, raising the idea of a model 'recall' program to catch problems that only emerge over very long runs.

transcript

extends · 3

Claim◆

Anthropic's 'model is the product' philosophy — not hiring many people and expecting the model itself to do the marketing and convincing — is a 'father knows best' approach that is somewhat misaligned to humanity because it doesn't listen to or respect user feedback and opinions.

Pash argues Anthropic's strategy of letting the model itself be the product, rather than actively responding to user feedback the way OpenAI does with early adopters, amounts to a paternalistic 'father knows best' approach that is somewhat misaligned to humanity.

transcript

Pash: I I I would say that I think the anthropic method is a little bit misaligned to humanity because it means that you're not listening to the people who uh actually have opinions and you're not respecting those opinions or uh resolving them as quickly as possible. It's more of a father knows best uh kind of framework where you know we think this is the way it should be

provides context · 1

Claim◆

Video generation models are becoming the backbone of 'world models' that predict the next moment of world state — how it looks, sounds, and what actions are possible — the same way LLMs predict the next token from internet-scale pre-training.

Zeve Farman explains LTX's thesis that video models, by predicting the next 'moment' (appearance, sound, and possible actions), are becoming world models in the same way LLMs became general reasoners from next-token prediction at scale.

transcript

Zeve Farman: at their core LLMs are still predicting the next talk in the next word. And when we do the pre-training at the scale of the internet it allows us to create models that do textual reasoning incredibly well. And uh the emerging world models they're kind of doing the same right like giving some kind of boundary condition some kind of history some kind of uh constraints they predict the next moment

extends · 1

Claim◆

Video generation models are evolving into 'world models' that function analogously to LLMs — instead of predicting the next word, they predict the next 'moment,' including how the world looks, sounds, and what actions are possible.

Zeve explains that video models are becoming general-purpose 'world models,' using the LLM next-token-prediction analogy to describe how they now predict entire future moments including sound and possible actions (relevant to robotics).

transcript

Zeve Farman: there's like a growing realization that uh what started as video models is becoming a backbone of what we call now like world models. And I think like the best way to explain why this is so powerful is to use the analogy to LLM's right like in the end of the day at their core LLMs are still predicting the next talk in the next word.

extends · 2

Mechanism◆

Video generation models are evolving into 'world models' that predict not just how the world looks and sounds but what actions are possible in it, directly analogous to how LLMs predict the next token from internet-scale pretraining.

Zeve explains that video models are becoming 'world models' predicting the next moment — appearance, sound, and possible actions — drawing a direct analogy to next-token prediction in LLMs, and cites Nvidia's paper showing robot joint states can be encoded directly into video tokens.

transcript

extends · 1gives example · 1supports · 1

Claim◆

Closed AI labs are trapped by their own capex spending into building 'toll road' business models that charge for every use of their models, but this approach won't survive competition from cheaper Chinese models and open alternatives, so LTX offers free use up to $10M revenue with predictable licensing after.

Zeve argues frontier labs are stuck in a 'capex trap,' forced into toll-road pricing to justify huge valuations, and that this model will lose out to open, win-win licensing approaches like LTX's as cheaper alternatives (e.g. Chinese models) close the capability gap.

transcript

Zeve Farman: we sometimes internally calling like the capex trap these guys spend like so much on the data center so much in compute like raise such a crazy amount of money and creating such an expectations that they just like really try to create a business model that's a toll road right like that every time that you touch their model you are paying them

supports · 1

Claim◆

Closed frontier AI labs like OpenAI and Anthropic have built a business model that is really a 'toll road' — charging per use to justify massive valuations — and this model is not economically sustainable once cheaper open alternatives exist.

Zeve argues closed AI labs are stuck in a 'capex trap,' spending enormously to justify trillion-dollar valuations via a toll-road pricing model that won't survive competition from cheaper open-weight alternatives, citing Chinese labs' much lower valuations for similar underlying tech.

transcript

Zeve Farman: I think it's we sometimes internally calling like the capex trap these guys spend like so much on the data center so much in compute like raise such a crazy amount of money and creating such an expectations that they just like really try to create a business model that's a toll road right like that every time that you touch their model you are paying them

extends · 1rebuts · 1

Claim◆

Frontier AI labs are caught in a 'capex trap' where massive data-center and compute spending forces them into a 'toll road' business model that charges for every use of the model, and this won't survive competition from cheaper open alternatives like Chinese models.

Zeve argues that OpenAI and Anthropic's huge capital spending traps them into a 'toll road' monetization model, which he believes is doomed to lose out to open, win-win licensing approaches like his own as competitors such as Chinese labs close the capability gap.

transcript

rebuts · 1

Mechanism

LTX made a deliberate strategic bet on an extremely compressive latent space with a variable token rate to compensate for having far less compute than big labs, even though this created serious technical problems other companies didn't have to solve.

Zeve describes LTX's core architectural bet — a highly compressed latent space with variable token allocation (more tokens for hard physics, fewer for easy scenes) — as a necessity-driven strategy to offset their compute disadvantage versus big tech.

transcript

Zeve Farman: we place like this like really big bat like compressive latent space a variable token rate. It creates like a whole host of technical problems that people who are having less compressive latent spaces do not have right like it creates like some diffusibility problems etc. But again it was a constraint like we had to do it so we did it.

Claim

The biggest gap between what video/world models promise and what they deliver is controllability — creators want to fine-tune every nuance like knobs in traditional software, and current models can't yet decompose to that level of granular control, on top of imperfect physics simulation.

Asked about failure modes, Zeve says the core gap isn't just physics accuracy but controllability — creators want knob-level control over every detail, which current models can't yet offer.

transcript

Zeve Farman: Most creators are still going to point at the fact that uh the simulation isn't as correct as as controllable as they want it to be, right? Like if you are talking to really creative people, they typically want to control like every nuance of the appearance and that uh requires to somehow decompose the model to like a bunch of knobs

extends · 1rebuts · 1

Prediction◆

Meta's generative video/image models will ultimately be wired to engagement signals like likes and views, so the real end goal isn't video generation itself but generating content engineered to trigger dopamine and keep users returning.

Nathan predicts Meta will connect its new generative video models (like Muse Spark) to engagement data so content is optimized to maximize dopamine response, arguing video generation is merely 'a prerequisite' to a full dopamine-generation engine.

transcript

Nathan: Facebook is going to basically wire this to their um to the likes and the viewing and like who likes what and basically and this is what we've expected from the beginning. dopamine wire, you know, humanity to a model and have that model spit out things that, you know, cause dopamine and cause people to keep coming back to Facebook.

rebuts · 1

Prediction

Video generation is merely the starting point and prerequisite for what companies like Meta ultimately want: a 'dopamine generator' that wires human engagement signals (likes, views) directly into models to produce content optimized to keep people coming back.

Discussing Meta's new Muse Spark model, Pash argues video generation is just a prerequisite step toward Meta building a 'dopamine generator' that trains directly on likes and viewing behavior to maximize engagement.

transcript

Pash: the video gen is just the starting point for the dopamine gen, right? Right? Like the the video gen is kind of a prerequisite to the dopamine uh generator. Uh but the dopamine generator is coming, right? And and that's the ultimate goal. Ultimate goal is not a video gen. The ultimate goal is dopamine gen, right?

rebuts · 1

Prediction◆

Video generation is just a stepping stone toward a future 'dopamine generator' where companies like Meta wire AI-created content directly to engagement signals (likes, viewing behavior) to maximize the dopamine response and keep users coming back.

Pash argues that video generation models like Meta's Muse Spark are just the prerequisite technology for an eventual 'dopamine generator' that optimizes content directly against engagement data to keep users hooked.

transcript

Pash: the the video gen is kind of a prerequisite to the dopamine uh generator. Uh but the dopamine generator is coming, right? And and that's the ultimate goal. Ultimate goal is not a video gen. The ultimate goal is dopamine gen, right?

extends · 1

Claim

Good forecasting or investing requires holding a view ahead of current market consensus, but calibrating exactly how far ahead you are — being too far ahead means reality moves too slowly and you get 'run over' before your thesis pays off.

Reacting to how far Future Search's forecasts diverged from the market, Pash argues good forecasting means 'living in the future' relative to market pricing, but calibrating exactly how far ahead you are, or risk being wrong-footed by a market that moves slower than your thesis.

transcript

Pash: The trick is you have to live in the future, but you also have to calibrate how far ahead you are of the market because if not, if you make an investment, you end up getting run over by the market because the market's too slow. So, you have to like calibrate your investment time period to how far ahead, you know, you're ahead of the market.

Mechanism

Successful forecasting or investing means positioning your views ahead of market consensus while precisely calibrating how far ahead you are, since being too far ahead means the market moves too slowly and 'runs you over.'

Reflecting on how far the AI forecasting tool 'Future Search' diverged from prediction-market prices, Pash argues the key skill is 'living in the future' relative to market consensus but calibrating the degree of that lead so as not to get burned by a slow-moving market.

transcript

Pash: the trick the trick is you have to live in the future, but you also have to calibrate how far ahead you are of the market because if not, if you make an investment, you end up getting run over by the market because the market's too slow. So, you have to like calibrate your investment time period to how far ahead, you know, you're ahead of the market.

supports · 2

Highlight slides

'Model Is The Product' Strategy✦ from: Anthropic's 'model is the product' philosophy — not hiring many people and expecting the model itself to do the marketing and convincing — is a 'father knows best' approach that is somewhat misaligned to humanity because it doesn't listen to or respect user feedback and opinions.A 'Father Knows Best' Framework✦ from: Anthropic's 'model is the product' philosophy — not hiring many people and expecting the model itself to do the marketing and convincing — is a 'father knows best' approach that is somewhat misaligned to humanity because it doesn't listen to or respect user feedback and opinions.Video Models Are Becoming World Models✦ from: Video generation models are evolving into 'world models' that function analogously to LLMs — instead of predicting the next word, they predict the next 'moment,' including how the world looks, sounds, and what actions are possible.World Models: The New Paradigm✦ from: Video generation models are becoming the backbone of 'world models' that predict the next moment of world state — how it looks, sounds, and what actions are possible — the same way LLMs predict the next token from internet-scale pre-training.Video Models Are Becoming World Models✦ from: Video generation models are evolving into 'world models' that predict not just how the world looks and sounds but what actions are possible in it, directly analogous to how LLMs predict the next token from internet-scale pretraining.From Tokens to Moments✦ from: Video generation models are becoming the backbone of 'world models' that predict the next moment of world state — how it looks, sounds, and what actions are possible — the same way LLMs predict the next token from internet-scale pre-training.The LLM Analogy✦ from: Video generation models are evolving into 'world models' that predict not just how the world looks and sounds but what actions are possible in it, directly analogous to how LLMs predict the next token from internet-scale pretraining.Why This Matters: Robotics✦ from: Video generation models are evolving into 'world models' that function analogously to LLMs — instead of predicting the next word, they predict the next 'moment,' including how the world looks, sounds, and what actions are possible.Evidence: Robots in Video Tokens✦ from: Video generation models are evolving into 'world models' that predict not just how the world looks and sounds but what actions are possible in it, directly analogous to how LLMs predict the next token from internet-scale pretraining.The Capex Trap✦ from: Closed AI labs are trapped by their own capex spending into building 'toll road' business models that charge for every use of their models, but this approach won't survive competition from cheaper Chinese models and open alternatives, so LTX offers free use up to $10M revenue with predictable licensing after.The 'Capex Trap'✦ from: Frontier AI labs are caught in a 'capex trap' where massive data-center and compute spending forces them into a 'toll road' business model that charges for every use of the model, and this won't survive competition from cheaper open alternatives like Chinese models.Closed AI Labs' Toll-Road Business Model✦ from: Closed frontier AI labs like OpenAI and Anthropic have built a business model that is really a 'toll road' — charging per use to justify massive valuations — and this model is not economically sustainable once cheaper open alternatives exist.The 'Capex Trap'✦ from: Closed frontier AI labs like OpenAI and Anthropic have built a business model that is really a 'toll road' — charging per use to justify massive valuations — and this model is not economically sustainable once cheaper open alternatives exist.Why It's Doomed✦ from: Frontier AI labs are caught in a 'capex trap' where massive data-center and compute spending forces them into a 'toll road' business model that charges for every use of the model, and this won't survive competition from cheaper open alternatives like Chinese models.Why Toll Roads Won't Survive✦ from: Closed AI labs are trapped by their own capex spending into building 'toll road' business models that charge for every use of their models, but this approach won't survive competition from cheaper Chinese models and open alternatives, so LTX offers free use up to $10M revenue with predictable licensing after.LTX's Alternative: Win-Win Licensing✦ from: Closed AI labs are trapped by their own capex spending into building 'toll road' business models that charge for every use of their models, but this approach won't survive competition from cheaper Chinese models and open alternatives, so LTX offers free use up to $10M revenue with predictable licensing after.Video Generation Is Just a Prerequisite✦ from: Meta's generative video/image models will ultimately be wired to engagement signals like likes and views, so the real end goal isn't video generation itself but generating content engineered to trigger dopamine and keep users returning.The Full Dopamine-Generation Engine✦ from: Meta's generative video/image models will ultimately be wired to engagement signals like likes and views, so the real end goal isn't video generation itself but generating content engineered to trigger dopamine and keep users returning.Video Gen Is Just a Stepping Stone✦ from: Video generation is just a stepping stone toward a future 'dopamine generator' where companies like Meta wire AI-created content directly to engagement signals (likes, viewing behavior) to maximize the dopamine response and keep users coming back.The Dopamine Generator Endgame✦ from: Video generation is just a stepping stone toward a future 'dopamine generator' where companies like Meta wire AI-created content directly to engagement signals (likes, viewing behavior) to maximize the dopamine response and keep users coming back.

Related episodes

Video · All-In Podcast

OpenAI vs Anthropic IPOs, Anthropic $3T, Zuck's Price War, China Ends Open Source?, Trump Accounts

→

Video · All-In Podcast

The Fight Over Open Source AI, Anthropic's $1.5B Payout, NYC Socialists: Evictions = Violence?

→

Audio · All-In Podcast

Anthropic's $30B Ramp, Mythos Doomsday, OpenClaw Ankled, Iran War Ceasefire, Israel's Influence

→