The Ledger
Every statement that passed quotation and attribution checks. Mix any filter with any other: certainty 1/5, debate potential 5/5, or both at once.
why aren't all 28 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 0 checkable ones are still open, waiting for their date. predictions held up or didn't; assertions are supported or contradicted. on every card: ▮▮▮▮▮ certainty · ▮▮▮▮▮ debate potential. speakers are clickable
Ethan He: Pipeline bug fixes drive more model gains than new algorithms
“And often I find that this is kind of boring, but like a lot of the improvements does not come from new algorithms. It comes from finding small bugs here and there in the data pipeline, in the model training pipeline. Those gave the biggest boost to the model …”
Ethan He: Training video models costs roughly the same as medium-scale LLMs
“So surprisingly video models is like the cost is very, is comparable to language models. And obviously the largest scale is language model. Maybe like a medium scale language models.”
Ethan He: Visual intelligence in video generation models stems primarily from language models
“The visual intelligence are actually mostly coming from language. Like, these video models, especially from now, since the diffusion model technology is more mature, the, like, every time you see there, there's some improvement on these models, I would say mos…”
Ethan He: Peak pre-training learning rate works best for MoE upcycling
“We found that the best is to use the original highest peak learning rate from pre-training, which works the best.”
Ethan He: Video models require image foundations and 100% synthetic caption pairs
“Building a video model. You actually need to build a image model first and building, building these two models. The data you need is a hundred percent synthetic pair of language and image or language to video because on the internet, actually the videos Don't …”
He: Upcycling dense models to MoE beats continuing dense training per FLOP
“By training these upcycled models, you can achieve better accuracy than simply training the dense model further for the same number of flops.”
He: Mixtral's top-k before softmax routing hurts MoE upcycling performance
“We actually found the mix-throughs approach didn't work as well as expected, because the original model, the original switch transformer from Google uses a softmax and topk for a reason. And because of upcycling, if you switch to topk, then softmax, it actuall…”
Ethan He: Video Foundation Models Follow Scaling Laws Like LLMs
“There, once I built the Cosmos one, I realized as this thing also has a scaling law similar to language model.”
Ethan He: Daily iteration speed is the top factor in model training
“When I look at like training models, I don't so actually the top important thing is like how many how many iterations can you do like per, per day? And the more iteration can you do, you can train the model much faster. So if you have a very strong infra and y…”
Ethan He: Coding Models Shift Research Bottlenecks Back to Compute
“Compute might become a bottleneck again, because previously, like if you want to train a new model, say you want to generate new synthetic data and then, or write a new algorithm, it might take a few weeks. And during that period of time, you don't, you might …”
Ethan He: Training generative video models on unlabeled data aids generalization
“For the generative model training, there's also really like a small percentage of unlabeled data. So, so the model is instructed to generate a video without any text instruction. That, that can also help the model generalize.”
Ethan He: Training transformers directly on raw image pixels is impossible
“If you're trying, if you can technically, theoretically train image or video models on pure pixels, but the problem is that the, it's a lot of tokens. So like one image, like it's a thousand by a thousand is like one million tokens, one million pixels. It's im…”
Ethan He: Diffusion transformer training closely mirrors LLM training architecture
“So now the training, training of the diffusion transformer, you already generated models use diffusion transformers. It is actually quite standard. It's very similar to how you train a language transformer models. It's not that much difference. It's just the t…”
Ethan He: Video Models Must Bootstrap From Image Diffusion Models for Semantic Understanding
“After you train such model, such image model, the reason it's a foundation for video models is that image, image models are Cheaper to train and they have much denser connection between language and text. So, sorry, language and images. For example, you train …”
Ethan He: Frame-by-frame video compression enables real-time interactivity, temporal compression adds lag
“That being said, the benefit of the frame per frame compression, we might come back to this later, is real timeliness and interactivity. Because if you strain the output of the model frame by frame, you can As a model can respond to any user request immediatel…”
He: Distillation works because teacher models are simpler than the internet
“I guess the, from the modeling perspective, the strong model, the teacher model is trying to model The image and videos of entire internet. And that distribution is extremely complex. As a step distilled model is just trying to learn from the teacher. The teac…”
Ethan He defines world models as real-time, interactive, long-horizon video
“So word model is like real time, interactive, long horizon videos.”
He: Manual reference video conditioning is a workaround, not true long context
“It doesn't need to have a very long context, but it's, I feel like it's an intermediate solution. It's cheating. Yeah, the model should Be able to like selectively know, like where, where should I draw references?”
He: AI first-principles planning calculates the theoretical minimum days to ship
“If you think about some limitation, for example, the current data, like how, how fast can we acquire the videos? And if you think about training the models, like what's the iteration speed? For training a model end-to-end and how, how would adding more GPUs ac…”
Ethan He: Language models prompt AI models better than humans
“Most of the people were actually not very good at prompting. Actually, language models have a better sense of how to prompt AI models. AI models know AI models better.”
He: Video Agents Are Inherently Costlier Due to Iterative Multi-Sample Generation
“I think the enterprise will have much more budget for video models because the agents are inherently more expensive than the other video models themselves because they do this iterative process. They generate many, many variations.”
Ethan He: External heuristic engineering gets absorbed into models
“From our experience, the heuristic engineering also have the models get absorbed into the models themselves.”
He: Token dropping works in MoE pre-training; dropless excels in fine-tuning
“A lot of pre-training experiments show that Token dropping is very efficient, and it doesn't impact performance, but in some of the, like, the downstream fine-tuning, people realize drop-less is better.”
Ethan He: 64 experts is the sweet spot for MoE upcycling
“We found, 64 experts is kind of like the sweet spot. If you increase the number of experts beyond 64, it provides diminishing return.”
Ethan He: Training models directly on MP4 tokens is extremely difficult
“So people actually have tried that, but the main challenge is the latent space for the MP four tokens are not, we're not very comprehensible for the models. It's extremely hard to train on that.”
Ethan He: Temporal video compression cuts context length up to 4x
“The difference is if you compress the temporal dimension, you get a much higher compression rate. Because there is temporal redundancy between frames, because this frame and the last frame, likely they are mostly similar. So there's only some small difference.…”
Ethan He: Stored continuous VAE latent features match raw video size
“You use a VAE to compress the videos and you also need to store, typically you need to store those continuous feature on also in your storage. That's also comparable size with the videos themselves.”
He: Audio-video models require precise time alignment unlike text-to-video
“So one important thing is like the alignment. So the model has to know like the video and audio, the it has to have a time based alignment, like at which time step the video and the audio token correspond to each other. We actually don't have these kind of ali…”