To be clear, I don’t believe we should halt AI progress. Higher education must adapt. But I worry that most universities, already overwhelmed by ongoing crises, lack the agility and foresight to make the tough decisions needed to survive.

Given that roughly half of the academic AI papers published in our top-tier conferences are produced by Chinese universities, this would catastrophically impair AI research in the USA if researchers cannot download code or weights if they were developed by Chinese institutions.

The only barrier is having access to the right kind of chips, and DeepSeek figured out how to more effectively use the chips they have. The learnings from DeepSeek about how to use FP8 will enable AI folks worldwide to get more from NVIDIA's newer chips.

A huge percentage of the PhD students trained in the USA in AI are Chinese, where we only have about 30% domestic students nationwide. We aren't getting domestic applications to do PhDs in the USA. Why people think China wouldn't have AI expertise confuses me.

There are too many unknowns to justify using a fixed compute-based threshold. Policymakers should focus on regulating specific high-risk AI applications, similar to how the FDA regulates AI software as a medical device.

Lastly, many trying to scale LLMs beyond systems like GPT-4 have hit diminishing returns, shifting their focus to test-time compute. This involves using more compute to "think" about responses during inference rather than in model training, and the regulation does not address this trend at all.

It is unlikely that AI progress will remain tied to inefficient transformer-based models trained on massive datasets.

Second, the 10^26 operations threshold appears to be based on what may be required to train future large language models using today’s methods. However, future advances in algorithms and architectures could significantly reduce the computational demands for training such models.

The current regulation seems misguided for several reasons. First, it assumes that scaling models automatically leads to something dangerous. This is a flawed assumption, as simply increasing model size and compute does not necessarily result in harmful capabilities.

I agree, but I don't think it was because I was a student. I think it is because of how enormous the conferences have become. I had a ton of fun at CoLLAs-2024, but it was single-track and only had a few hundred people vs 10-20k people.