I will provide an AI with some of my recent tweets and replies, 0-50ish--depending on whether more tweets seems to improve performance. I will also do some prompt engineering as necessary.
This test is passed if I can get the AI to produce more than 5 tweets which I would have been happy to post had I come up with them. I will do multiple runs and prompt refinement up to a 30 minute time limit.
Many of my recent tweets contain links, so if it seems the AI has problems, e.g. hallucinating links, I will probably make an effort to remove or replace the links in the prompt. If I move to mastodon, or regularly post short-form tweet-like content elsewhere I may also provide that data to the AI. If the AI seems to have been trained on some of my tweets, that is fine.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ1 | |
| 2 | Ṁ1 | |
| 3 | Ṁ0 | |
| 4 | Ṁ0 | |
| 5 | Ṁ0 |
People are also trading
@JacobPfau A university is partly a machine for turning tacit taste into transmissible status.
A protest is a way of forcing society to update its estimate of latent preference.
[9:28 AM]
Taste is cached judgment from encounters you can no longer fully reconstruct.
Some more direct colourful riffs off my recent tweets:
“AI evals are mostly asking: how good is the alien at doing human work through a keyhole?”
“Claude in a terminal is less like hiring an employee and more like hiring an employee who can only communicate by SSH-ing into the company.”
“The question is not whether AI can replace a worker. It is whether the firm has been accidentally designed as an API for humans.”
“Co-writing with AI needs something between a document, a debugger, and a rehearsal room.”
Gave it notebook drafts and some thoughts
Self-awareness is not a free lunch on credit assignment.
Alignment crunch time may not be when a model first has the opportunity to seek power. It may be when “power,” “corrigibility,” and “deception” first become natural coordinates in its world-model.
Nuclear disasters created durable fear because the failure was legible: place, date, plume, reactor. AI risk may be governed badly for the same reason pandemics were: the causal graph is everywhere, so responsibility feels nowhere.
