Securing the weights isn't enough, we also need to secure the algos
If final training budgets of AI companies right now are ~10% of compute of US AI companies[1], and the algorithms (e.g. architecture, optimizer, data improvements) are leaking quickly, then Chinese labs can match US performance at any point in time by spending all of their compute on training.
Some related points:
I think this has some big implications:
These ideas aren't very original to me, this is from a bunch of convos with folks at AIFP and Redwood
Most of the compute of US AI companies is going to inference / experiments in order to develop better algorithms. In this scenario China could just do none of that.
The data is an interesting middle ground; naively exfiltrating the data would take a lot of bandwidth, but a lot of the ideas (e.g. better data filtering, synthetic data generation techniques, etc) can probably be summarized compactly. It's unclear what fraction of algo progress is currently data progress; the higher the fraction, the easier algo security looks.