This is how you vibe debug your LLM in the future

this is cool looking but I wonder if it’s actually doing anything or just showing a cool three js particle/fog effect and changing colors with the slider?

It’s a tracing of all embedding vector values through the layers of the network and I’m just “temporarily massaging” the values. There’s actually a bug in the movement so I just made more of a meme post. I think it’s perhaps the start of something that can have real value. The way the transformer works is that each token is given an embedding vector. Every token at every position slowly becomes a representation of the networks best prediction of what will come next, but at every token position, it doesn’t actually know what the next one will be. There are still more geometric concepts to build in. A lot of it I think is thinking creative ways to visualize what’s happening in a way that gets us further away from raw, nearly indecipherable, data. It’s not like a convolutional network classifier. Language models do a lot of weird things in different phases of the network, and I want to bring a little bit more of that to light in an interactive way.

Yes lol

this would make sick water caustics

what's the mapping between text and visualization?

My previous two posts offer a bit more explanation. This is essentially showing the embedding one but through all the layers at once. For fun I made them move a bit, but it's a bit exaggerated here, but I think movement could work if done right

Does this hurt the llm

No it tickles it though

based on probabilities, you could highlight low-confidence tokens (probable hallucinations)