Tl/Dr:

Generalisation is a capability in its own right, and in our experiments it scales less quickly than specific capabilities.
Our experiments show that on a range of tasks, finetuning Babbage on task A transfers more to task B than finetuning Davinci on task A transfers to task B. This is relevant for identifying the nature of the alignment problem we might encounter (for things in this vague cluster see here, here, and here).
Of independent interest is the fact that a transfer learning metric can give us a fairly non-arbitrary way to measure how similar any two tasks are.

Epistemic status: written quickly, but oh look we did 4×8^2 experiments and we have a lovely graphic

This post makes two contributions. First, we compare results of transfer learning experiments from Babbage and Davinci and suggest that, at least in these experiments, the ability to generalize scales less quickly than do [...]

---

First published:
February 9th, 2024

Source:
https://www.lesswrong.com/posts/qCecZwptWuAEZN3Td/transfer-learning-and-generalization-qua-capability-in

---

Narrated by TYPE III AUDIO.

“Transfer learning and generalization-qua-capability in Babbage and Davinci (or, why division is better than Spanish)” by RP, agg

LessWrong (30+ Karma)

“Transfer learning and generalization-qua-capability in Babbage and Davinci (or, why division is better than Spanish)” by RP, agg