Michael J Curry
currymj.bsky.social
Michael J Curry
@currymj.bsky.social
Assistant professor of computer science at UIC focusing on AI, economics & computation, mechanism design.
if i tell it "be very careful to make sure the code compiles", seems to be able to get it first try on some other simple tasks. of course no general conclusion on that prompting strategy. it's extremely funny that this is how computers work now
April 27, 2025 at 10:02 PM
i just asked o4-mini-high to code up a self contained Fortran version of running PCA on some random vectors, and it didn't compile first try (did second try after being shown the errors). haven't checked correctness but wouldn't be shocked by subtle error somewhere.
April 27, 2025 at 9:49 PM
imagining a game where people commit to have a proxy play a mixed strategy for them doesn't seem like such a stretch. and once you're in that world, committing to let a proxy play a correlated strategy for everyone seems pretty natural too. i guess that's still pretty mechanism-design-ish though
February 9, 2025 at 10:26 PM
i believe this dialogue is the quota for the irritating AI grammar checker, not compiler. if also failing maybe ui bug or something
January 14, 2025 at 2:45 PM
tried to get gemini deep research to do this and the result was Bart Simpson level

“…Mustard, a key ingredient in the mustard-emulsion method, boasts an ancient history. Its use dates back to ancient Egypt, where it was primarily used for medicinal purposes…”
January 11, 2025 at 3:56 AM
even relaxing a lot of assumptions it seems hard to escape maximizing over discrete sets somewhere. i guess some behavioral models like rational inattention also have the pleasant side effect of smoothing the problem, but i don’t know if people find them credible or not
December 9, 2024 at 11:53 PM
learned the hard way by taking derivative of an optimization problem and watching gradient descent not converge
December 9, 2024 at 11:43 PM
one partial explanation is in physics it’s usually safe to assume everything is nice and smooth. in economics, you have agents that make optimal choices, which tends to create discontinuities, often at exactly the point you care about. so there has to be at least a little more analytic rigor.
December 9, 2024 at 11:26 PM
P-CTC also =PSPACE. in other words a quantum computer would not be useful if you already had a time machine
November 27, 2024 at 6:01 PM
early non-chatbot GPT3 was mostly bad but got a hit occasionally. chat versions seem to never
November 21, 2024 at 3:54 AM
hello!
November 17, 2024 at 3:26 AM