Chris Olah, Neel Nanda, Catherine Olsson, Nelson Elhage, and a bunch of other people at Anthropic just published “Transformer Circuits,” an application of the Circuits-style interpretability paradigm to transformer-based language models. From their very top level summary:

Can we reverse engineer transformer language models into human-understandable computer programs? Inspired by the Distill Circuits Thread, we're going to try.

They've chosen to publish their work in an interestingly novel format, publishing their first paper, “A Mathematical Framework for Transformer Circuits,” alongside a set of YouTube videos that go into even more detail on their findings. I watched the full YouTube playlist, and I found it absolutely fascinating and would highly recommend it as a way to engage with this research.

Some of my high-level takeaways:

• Their signature finding, I think, is that of induction heads. I'll let you get the explanation of what they are directly from Chris's video, but essentially they're a very basic mechanism that all transformer-based language models seem to use that drives their ability to “meta-learn”—that is, improve their accuracy over the course of seeing a larger context. Similarly to the exploration of early vision in Circuits, induction heads give us insight into the basic building blocks of large language models. While I think the authors' did a great job uncovering this basic building block, I think their explanation of meta-learning as primarily being driven by induction heads mostly leads us to a big open research question, which is how exactly induction heads can be put together to produce the more complex phenomenon we see in large language models.^[1]
• Probably their most fascinating finding, in my opinion, was their discovery of the “induction bump.” I won't try to reexplain exactly what the induction bump is, since I think Catherine's video does such a good job of that, but I will say that, to my knowledge, the authors' exploration of the induction bump is the first time there's ever been a detailed, circuits-level analysis of a phase-change that occurs over the course of training in a discrete way. I think this is especially interesting because a lot of stories that people like to tell about how particular safety problems might arise in neural networks often rely on these sorts of phase-change-style transitions (e.g. the development of deception). It seems to me that the existence of something like the induction bump is suggestive that there might be more of these sorts of phase changes hiding in future large training runs, or even in existing large training runs—as Catherine notes, the bump is pretty easy to miss until you break down the loss into individual tokens.^[2]

Overall, I think this is clearly the most exciting progress in transparency and interpretability in general since Circuits—and I'm really happy to see it happening in language models, as I've previously emphasized I think is important for us to focus on. One thing that I think really sets this sort of transparency and interpretability work apart from the rest is the authors' emphasis on understanding the mechanistic building blocks underlying their models—with the hope of eventually being able to reverse-engineer them—rather than just, for example, trying to give humans tools to predict what models are doing (without the output of those tools necessarily having any correspondence to what the models are actually doing).