I think this is an underrated idea, and should be considered a good refinement/addition to the FTX theme #2 of “AI-based cognitive aids”. If it’s worth kickstarting AI-based research assistant tools in order to make AI safety work go better, then doesn’t the same logic apply towards:
Supporting the development of brain-computer interfaces like Neuralink.
Ordinary productivity-enhancing research software like better note-taking apps, virtual reality remote collaboration tools, etc.
The idea of AI-based cognitive aids only deserves special consideration insofar as:
Work on AI-based tools will also contribute to AI safety research directly, but won’t accelerate AI progress more generally. (This assumption seems sketchy to me.)
The benefit of AI-based tools will get stronger and stronger as AI becomes more powerful, so it will be most helpful in scenarios where we need help the most. (IMO this assumption checks out. But this probably also applies to brain-computer interfaces, which might allow humans to interact with AI systems in a more direct and high-bandwidth way.)
I think this is an underrated idea, and should be considered a good refinement/addition to the FTX theme #2 of “AI-based cognitive aids”. If it’s worth kickstarting AI-based research assistant tools in order to make AI safety work go better, then doesn’t the same logic apply towards:
Supporting the development of brain-computer interfaces like Neuralink.
Research into potential nootropics (glad to hear you are working on replicating the creatine study!) or the negative cognitive impact of air pollution and other toxins.
Research into tools/techniques to increase focus at work, management best practices for research organizations, and other factors that increase productivity/motivation.
Ordinary productivity-enhancing research software like better note-taking apps, virtual reality remote collaboration tools, etc.
The idea of AI-based cognitive aids only deserves special consideration insofar as:
Work on AI-based tools will also contribute to AI safety research directly, but won’t accelerate AI progress more generally. (This assumption seems sketchy to me.)
The benefit of AI-based tools will get stronger and stronger as AI becomes more powerful, so it will be most helpful in scenarios where we need help the most. (IMO this assumption checks out. But this probably also applies to brain-computer interfaces, which might allow humans to interact with AI systems in a more direct and high-bandwidth way.)