Mm-hmm

... to make models better with RLHF is to collect a lot of labeled data, to spend a lot of time having people manually say, like, "This was a good response.

This was a not good response," and then incorporate that into model training, which leads to a lot of, lot of GPU cycles.

Uh, more potentially efficient, uh, area to explore in the future, I think, would be what if the model can understand from just the interactions with the user if it's actually delivering kind of what they want and then allow it to, to self-optimize for that.

So that brings up to, gets to this framework, uh, attention fine-tuning, which is very, very early stage, but think, I think this technique shows promise.

So yeah, the name Attention Fine-Tuning, so we're getting technical, like you said.

Practitioners in the AI world are probably familiar with this paper, "Attention Is All You Need," that came out around 10 years ago that set up the transformer architecture, the deep learning me- model implementation that led to LLMs and ChatGPT and Claude.

So what does it feel like when the math from aerospace engineering starts solving enterprise AI problems?