This is a very helpful resource and an insightful analysis! It would also be interesting to study computing trends for research that leverages existing large models whether through fine-tuning, prefix tuning, prompt design, e.g., “Fine-Tuning Language Models from Human Preferences”, “Training language models to follow instructions with human feedback”, “Prefix-Tuning: Optimizing Continuous Prompts for Generation”, “Improving language models by retrieving from trillions of tokens” (where they retrofit baseline models) and indeed work referenced in Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing
Ron
I agree treaties, regulations and research are complementary. However, I see a bigger role for regulation than David says—I think it can incrementally build up defenses to stop dangerous releases.
In particular, I advocate requiring safety cases with increasing stringency for increasingly risky capabilities (like loss of control) as the most urgent way to address x-risk. It will be easier to start by establishing a framework for this bottom up e.g., in California, and then build up support in different nations and ultimately establish international coordination. Also, the US is driving the AI race—I believe Chinese progress is mostly downstream of the much larger investments US labs are making (e.g., RL environment and data vendors seeded by US lab customers).
Requiring safety cases would create important incentives for industry to invest in safety and scientific understanding to better build safety cases. The safety cases would block deployment (eventually even training) without real progress on alignment and evidence of efficacy. We can start by regulating the largest labs who are driving the race and extend regulation as capabilities advance.
We can more easily coordinate around concrete safety proposals and convincing other nations to block risky AI activities. A pause regime based on limiting hardware would be challenging (e.g., MIRI’s proposal covers 16 H100 equivalents yet it still doesn’t have a way to catch decentralized training). Likewise for limiting research that advances AI capabilities: where do you draw the line on hardware advances, general algorithm improvements, computational neuroscience.
I wrote an article with more details on how this can work:
https://www.linkedin.com/pulse/who-gets-stop-unsafe-ai-release-ron-bodkin-iinge/