Notices by Mark Gritter (markgritter@mathstodon.xyz)

I think this is very interesting but really very _basic_ research on the capabilities of LLMs at reasoning problems. Any random PhD should be able to move from a static benchmark of math problems to a distribution of similar problems. That's exactly what this paper does, and discovers:

1. All current models do worse at GSM8K-like problems than they do on GSM8K itself, and there is wide variation in success for different samples.
2. #LLM performance varies if you change the names. It changes even more if you change the numbers. Change both, and you get even more variation.
3. Adding more clauses to the word problems makes the models perform worse.
4. Adding irrelevant information to the word problems makes the models perform worse.
5. Even the latest o1-minit and o1-preview models, while they score highly, show the same sort of variability.

I think this is the bare minimum we should be expecting of "AI is showing reasoning behavior" claims: demonstrate on a distribution of novel problems instead of a fixed benchmark, and show the distribution instead of the best results.

It's not that humans don't share similar biases -- plenty of middle-school students are tripped up by irrelevant data too -- but I think results like this show we are very far off from any sort of expert-level LLMs. If they show wide distribution of behavior on tasks that are easy to measure, it's quite likely the same is true on tasks that are harder to measure.

https://arxiv.org/abs/2410.05229

From Max Kreminski at #bangbangcon: Humans are worse at ideation when they use ChatGPT, compared to Oblique Strategies.

The paper he coauthored: https://arxiv.org/abs/2402.01536

Projects he talked about:

Blabrecs: beat the classifier at making up nonsense English-y words. https://mkremins.github.io/blabrecs/

Blabwreckage: start with either a real poem or complete gibberish, then "wreck" it into something vaguely language-shaped. https://mkremins.github.io/blabwreckage/ (I don't think you can provide your own seed? At least, not without hacking the Javascript?)

Savelost: remove one letter at a time from a sentence, attempting to preserve meaning. (I'm curious how a human would do at this task in comparison.) https://barrettrees.com/savelost/

@inthehands So capitalism, but only on existing resources, and without any competition? Like, was he _trying_ to make the point that rent-seeking is bad? Capitalism But Everything is A Monopoly?

Maybe I'm making the unfair assumption that nobody could bring in extra chairs to compete and actually establish some sort of market, but I kind of doubt it.

I am an old now because I remember the last time "run compiled code in the browser" was a thing. But I'm not sure anybody else does, with all the takes about how running a virtual machine (and Java!) in the browser is such a great thing?

Is there a good historically-minded writeup about how #WebAsssembly differs from the last browser VM hype cycle?

TIL that there are Dudes on Quora who feel the need to respond to questions about linear inequalities with anti-woke messaging.

I did not realize you could create a unicursal #maze out of an ordinary branching maze by placing a wall in the middle of each corridor!

https://twitter.com/aemkei/status/1388610729855553549

This makes sense in that it's sort of enforcing the "right-hand rule" by taking away all your choices, guiding you along the path that the RHR would take on the original maze -- and that path is deterministic and thus necessarily branch-free.

@inthehands Very nice!

I'm frustrated by how long we've heard "use composition not inheritance" and yet language support for composition is still so poor. (I like what Swift does -- but my day-to-day language right now is Go and like everything else in Go the solution is "write a bunch of boilerplate so that the function names match up.")