Member-only story
Some Thoughts on AI Tax
This piece is a short response to Brian Slesinsky on AI taxes (from my email), a suggestion from a Marginal Revolution reader posted by Tyler. Check out Zvi’s response as well.
His proposal is to introduce a tax that affects API calls to LLMs. He compares it to a Tobin Tax, which has the goal to tax and therefore discourage short-term currency trading, thus reducing speculation and stabilizing a currency.
The reason for such a tax is to discourage automated use of APIs (for example the model prompting itself, akin to AutoGPT). Right now LLM usage is always in the form of a person interacting with a chat, and the model is frozen, no learning is taking place. This is safe as it happens at a slow pace and because the person interacting with the model can just terminate the session if they see something dodgy going on. The only real danger is if the model is allowed to use its own APIs, or if another model is allowed to do the same without restriction.
I understand the danger here. If you believe that superintelligent models are around the corner or at least are theoretically possible, you would want to have human oversight at the correct place to be able to stop the process if it starts going in a dark direction. However, I don’t see how that would be implemented in practice. How would you find out that Meta’s model has been running in such an “automated…
