This is a linkpost to a new Substack article from MIT FutureTech explaining our recent paper On the Origins of Algorithmic Progress in AI.
We demonstrate that some algorithmic innovations have efficiency gains which get larger as pre-training compute increases. These scale-dependent innovations constitute the majority of pre-training efficiency gains over the last decade, which may imply that what looks like algorithmic progress is driven by compute scaling rather than many incremental innovations.
From the paper, our core contributions are:
... (read more)
- We find most algorithmic innovations we experimentally evaluate have small, scale-invariant efficiency improvements with less than 10× compute efficiency gain overall, and representing less than 10% of total improvements extrapolated to the 2025 compute frontier
Coauthor of the paper here. It's helpful to hear how people have interpreted our work, though I want to try to clear up a few things quickly.
Firstly, we don't claim that data must make up the remaining gap, and we certainly are not claiming it explains capabilities improvements more broadly. We try to be pretty precise about the definition of algorithmic progress under the CEG framework, which is specifically computed using loss rather than downstream performance metrics (I agree, the relationship between these is not straightforward or well understood). We say explicitly that the CEG framework cannot capture a lot of important innovations for performance and efficiency (instruction fine-tuning, constitutional AI, parallelism,... (read more)
Thanks so much for the thoughtful response. I'll certainly reach out and try to participate in BlueDot Impact's course now that I'm more familiar with the content, and will stay on the lookout for anything you document as you go through your own journey! Even just a few of the names and resources so far have been incredibly valuable pointers to the right corners of the internet.
I don't have karma yet, but if I did, I'd gladly open my wallet :)
The timing of this post is quite serendipitous for me. Much of what you wrote resonates heavily. First comment on LW, by the way!
I'm deeply interested in the technical problems of alignment and have recently read through the AI Safety Fundamentals Course. I'm looking for any opportunity to discuss these ideas with others. I've been adjacent to the rationalist community for a few years (a few friends, EA, ACX, rationalism, etc.), but the need to sanity check my own thoughts on alignment has made engaging with the LW community seem invaluable.
I have found the barrier to entry seems high from a career perspective. Despite how I'd like to spend my time, my... (read more)