> This is why it seems that alignment for a generalized, chat-style AI is a very hard problem, perhaps impossible. You can't align it to solve a certain kind of problem and keep it general to any question. The two goals are in conflict with each other.
I think this problem is going to be solved soon (hopefully), check out for the Persimmon model[0] from Humans&. They train it to mimic humans, it may seems bad but it could be _really_ useful to train an AI to be aligned to humans and understand their goals really well since they can use Persimmon to create a "fake human" following a defined goal that their big AI model will learn to estimate.
It's still early but I think this is what they are heading toward.
I think we should start training models to guess intentions. That's the whole game. e.g. If I dislike a particular short form video, what are the specific intentions I am conveying to the recommender system, (general category I don't like, the time at which it was shown, the combination of feed in which it came), these are all intentions I am trying to convey. In websites google analytics helps with clickstream tracking (or clickstream analysis, user journey mapping or behavioral tracking.) — the sequence of clicks and pages you move through is your clickstream, and analyzing it to infer intent. The same thing will need to happen in chatbots. Can a chatbot infer the specific instance of the problem I want to solve?
From the article:
> Let me clarify something: we created Tinybird ClickHouse because we wanted to build an analytics application without all the pain I'm describing. We do not offer ClickHouse® hosting; we solve the analytics problem
EDIT: I'm wrong, their home page clearly states: "Ship fast over a Managed ClickHouse®", though I don't really understand the difference between a managed clickhouse and clickhouse hosting...
It's defensive language for sure, i don't know how much it adds of protection in reality, but i nonethelesss sympathize with the author if he feels the need to protect himself that way or signaling the risk he takes.
> Tinybird is not affiliated with, associated with, or sponsored by ClickHouse, Inc. ClickHouse® is a registered trademark of ClickHouse, Inc.
Yeah, it sucks they need to do this. If I was a visitor to their website, I'd immediately want to know what ClickHouse, Inc. is and you'd realize ---> it's managed clickhouse, direct competitor... why would I use the one that needs all the ®'s
The target model is the original LLM that is large and expensive. It can verify candidate tokens in a single forward pass. It means you give all the context + candidate tokens that passes in parallel in the backbone, then you pass the language head (a matmul transformation to produce the token distribution) on all the candidate tokens and you can keep or drop tokens based on how many "quality" you want.
“This Regulation does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development.”
Does that apply to everyone who doesn't live in the EU as well? And what happens when a major AI company uses this dataset as training data, ignoring consent as usual?
But I've been told that the AI act was a terrible anti-innovation legislation made by clueless and corrupt bureaucrats. How is it possible that they made a sensible decision?
It would be fun to train a small vision neural network to recognize ads instead of relying on HTML or network requests. I don't know if anyone has done this yet.
Mistral has the capability of training such models. Take a look at Poolside[1], they are claiming to pre-train their Laguna series of models on 4,096 NVIDIA H200 GPUs[2].
Mistral has approximately 13,800 NVIDIA GB300 GPUs, which are nearly 2x more efficient for training.
The problem with Mistral is that they do not seem to have aligned incentives to train big open-weight models, even if the teams would like to.
From my experience their capabilities are extremely narrow and generally perform terribly when faced with issues outside comparatively narrow training data.
Yes, the point being made is that poolside is able to train large models with limited resources, which means that Mistral should be able to compete in that space, as they have access to much greater resources than poolside. Mistral simply chooses not to.
And Poolside’s latest models (Laguna S 2.1) are pretty good (not frontier, but competitive with the tier 2 models). Which means that Mistral could certainly compete in that space.
Not sure if they would have received the full number yet, but it's been a few months so they certainly could have. Bit of a moot point when the comparison was against Poolside's Laguna which isn't really "general" SOTA but SOTA-for-the-size, and Mistral is clearly capable of training 700B or 120B models that are that when released considering they have done that... A 2-3T model is probably possible with the GPUs they have but they would need to spend most of their resources on it, and it's not clear why they would want to.
"world model" is, sadly, a new buzzword; most of the world models we are seeing are video generative models (like GPT-Image), they receive an action vector and previous frames and must generate the next frame. It's working of course.
The whole field of world models also have lighter approaches based on using latent representations as the target (and not pixel frames). LeWorldModel[0] and DINO-WM[1] are a good way to learn how these works.
Strangely, most of the startups in the space of world models are focused on video-generative world models.
A good application of world models is in robotics where an action policy generate a distribution of actions, you can then predict the future state of the world and use some reward policies to choose the "best" future state.
Seems like MAI models from Microsoft are not going to be open-weight soon, but they are sharing a lot of details in the making of these models, which is a weird position.
I think the idea is show in-house progress and perhaps position MAI models as the "microsoft office of AI." My take is they might open source models 2-3 release cycles later, given this is supposed to be some new product line.
Such a gem, thanks to the author for sharing it's findings :)
The only problem I have with planing in latent space is that it can be really noisy and not representative of the positions in the game (the latent are trained for semantic, so the optimizer can focus a set of specific features and can skip positions, which means it cannot know "where" to go by optimizing on the latents directly).
Hey, author here! Thanks so much for reading :) I totally agree, the latent captures details useful for prediction, but not necessarily for control, making planning noisy. Still, I was surprised by how well it understood horizontal position after less than two hours of training on one A100!
I think this problem is going to be solved soon (hopefully), check out for the Persimmon model[0] from Humans&. They train it to mimic humans, it may seems bad but it could be _really_ useful to train an AI to be aligned to humans and understand their goals really well since they can use Persimmon to create a "fake human" following a defined goal that their big AI model will learn to estimate.
It's still early but I think this is what they are heading toward.
[0]: https://persimmon.humansand.ai/blog/persimmon.html
reply