How to build a diffusion language model | Hacker News Reader

How to build a diffusion language model

kuleshov-group.github.io

184 pointsvolodia20 comments

I've been studying these a bunch for a project in university. Last week I went over the derivation of the ELBO for a couple hours and it was a very fun and elucidating exercise.

Once you give names to the larger mathematical structures and understand them a bit better it becomes quite simple. I wish some of the blogs/papers I'd read had named "Importance Sampling".

The probability notation can be pretty confusing too. Sometimes it's hard to understand the "types" of some variables. But I'm inexperienced.

ChatGPT was surprisingly helpful. If you put in the work to truly understand the where the gaps are in your mental model (which parts aren't completely intuitive), it can do an amazing job filling in the gaps.

A good video from welch labs on image generation with diffusion models:

https://www.youtube.com/watch?v=iv-5mZ_9CPY&pp=ygUVZGlmZnVza...

Something I've wondered, maybe I should just do it if I can find some time, but... given DeepSeek's nice results on using rendered text as input, I'm wondering if anyone has given serious research efforts towards image-based diffusion methods for text.

As in, instead of all the complexities induced by discrete token generation, just generate the image of the text using standard image diffusion methods, then convert it to text.

If you used a single, monospace font, I bet this would be even pretty efficient, because the OCR problem becomes basically just direct template matching.

But I guess probably there is already a paper out there, I haven't searched. I'd be curious to know if it compares on par with token-based methods.

I've been wondering why we can't skip the entire token embedding step and just feed the model raw Unicode. I suppose the reason for that is efficiency and it's related to your question. Makes one wonder what other steps we can do more efficiently in code than in neurons.

I think it would make training take much much longer but the models would be able to tell you how many Rs are in strawberry.

Yes. By the way I think it is a pity that all the tutorials start with token embedding; by teaching LLMs like that, you don't learn why they are necessary.

There’s has a ton of work on character or byte level encodings for llms. The problem is you expand your input tokens by 3-4x. Expensive.

Also, you still need token embeddings (I think you might be confused how that works).

Could be! I have not (yet) spent much time learning about how llms work, just the occasional blog here and there. My main question is why we _need_ a bit of additional code to massage the input into tokens and especially why the neural network cannot do it, i.e., let the embedding be a latent space that forms naturally when training the network. If that makes sense.

are you aware of n-grams?

“let the embedding be a latent space that forms naturally when training the network”

The embeddings are produced in concert with the network, to serve the network, and not created as a separate step.

It’s actually very cool

The look-up table is a matrix. Each row is an embedding and each row number is a token ID.

You get a differentiable transformation from token ID to token embedding using a “one hot vector” and a matrix multiplication

If you take the transpose of this matrix, you can convert an internal representation back to the same token form, but treat it as logits and give it to the sampler.

So token embeddings are produced on demand in service of the model, according to the model’s needs.

I found an example of this strategy in a paper as far back as 1980!

In the other reply I recommend the Bengio paper. But do bite the bullet and try it.

I'll have a look at that paper, thanks!

Embeddings are there to make continuous-space identities so you can run them through a differentiable model. Without this the tokens (no matter your granularity) are pure surrogate identities, and you can’t run a gradient through them. You also hit the curse of dimensionality hard because the model can’t perceive similarity. “Cat” and “kitten” for example are simply different atoms of text, but with embeddings, you can leverage what you learned about “cat” when you encounter “kitten”. Look at “A Neural Probabalistic Language Model” (Bengio, 2003).

You can actually rig up an embedding variant of a Markov chain with just a few tokens of context, and no position coding, transformers, attention, none of it, and only minutes of training time. As long as you have the embedding lookup table trainable it will do some neat stuff.

Diffusion is already being used in drafters

I’d like to see more of these models.

I’ve been using diffusion Gemma and it is very fast on GPUs in output token/sec.

In the diffusion Gemma whitepaper, they say they could have done better with more time and compute.

Even with those caveats, it is very uses-able as a local model.

What's missing from these discussions is the real weakness of diffusion models: you can have two positions where two tokens need to be coordinated, but both spots are teetering between two valid states. They don't always collapse to the matching pair.

I only learned this the hard way reimplementing diffusiongemma. I had ideas on how to fix it but no cluster to train and experiment, hah.

have u seen the dspark paper, they add a morkov head for light weight dependency, not sure whether it can be extended to multi step refining

This should be much more efficient in theory, right? Why dont we see more leading labs adopt this?

Weak "chain-of-thought" abilities, high error rates and a very bad ability to recover from errors. If they produce a non-sequitur somewhere (which they are highly prone to doing), that global refinement spreads it everywhere like a blood infection superhighway.

Diffusion text models are cool, but they're functionally much less reliable than autoregressive transformers... and man that's really saying something. Right now most research on them is trying to figure out what complementary systems they need to be reasonably useful.

Google was at least trying. Wouldn't be surprised if the others were experimenting with it too. The bar is going to be a lot higher now for for any diffusion model to go from experiment to product since it needs to compete with stuff like glm 5.3 flash and Luna on cost/efficiency for a given quality of output, which is not going to be easy. If it was easy Gemini diffusion would have landed - if it requires a bunch of money and effort it has a much higher bar to make it to market, if it requires some clever breakthrough you have no way of knowing where that's going to come from or what it'll look like/if it'll even seem important when it happens

good stuff, no mention of confidence tho, recommend having a look at diffusiongemma and others.

Reply on news.ycombinator.com