God everything is so DUMB.

Thank you! 🙏🏻

Congrats!

The White House accord on AI that was signed by all the big tech oligarchs today misspells United States as “Unites States.”

also: 🤗 I hope you said that to them and they fell over apologizing!

What.

Makes sense, I looked up Cactus Labs and was wondering if we now have vapes driven by GPUs 😂

There is nothing inevitable about today's GenAI companies or market or product design or power dynamics. It's based on decisions, made implicitly and explicitly, by people. To imagine a better way we must realize how arbitrary all of this is. And who it serves.

Agree 100%. It’s also coming out at a time when enough people are frustrated with LLMs to catch on as a good idea.

🤗 I also tend to read over words like this, I empathize!

It’s “valence”, not violence, but I do like the idea of training it on my own replies and comparing it to others

That’s a super interesting comp, and probably the first thing I would have tried if I were trying to solve this problem from scratch practically!

What are you going to use it for?

Yes! Which is a funny architecture for a classifier, but it does have advantages (I did this to learn).

I have feelings about using the word "trained" here, as it's a LoRA'ed Qwen base model with a classification head. I would have used a different, slower approach if making it from scratch.

Overnight I trained a Jev-like model to output probability distributions over a set of 40 emoji, which runs in RAM at ~150ms per query. This was super easy and cost ~$10. It works pretty well on a weak training set, and with a day or two of work I could make this sing.