6 comments

[ 3.0 ms ] story [ 18.0 ms ] thread
I'm sorry, but what's the point here ? It's not for a job or improve a LLM or doing something useful per se, just to "enjoy" how version X or Y of an LLM can solve problems.

I don't want to sound grumpy or but it doesn't achieve anything, this is just a showcase of how a "calculator with a small probability of failure can succeed".

Move on, do something useful, don't stop being amazed by AI but please stop throwing it at my face.

I enjoy reading Peter’s ‘Python studies’ and was surprised to see here a comparison of different LLMs for solving advent of code problems, but the linked article is pretty cool.

Peter and a friend of his wrote an article over a year ago discussing whether or not LLMs are already AGI, and after re-reading that article my opinion was moved a bit to: LLMs are AGI in broad digital domains. I still need to see embodied AI in robots and physical devices before I think we are 100% of the way there. Still, I apply Gemini and also a lot of open weight models to both 1. coding problems and 2. after I read or watch material on Philosophy I almost always ask Gemini for a summary, references, and a short discussion based on what Gemini knows about me.

(comment deleted)
> I started with the Gemini 3 Pro Fast model ...

Quiet product announcement.

Odd that it came up with

  pattern_start = 1 if half_digits == 1 else 10 ** (half_digits - 1)
when

  10 ** (half_digits - 1)
is fine.
Last year I used LLM to solve AoC, to see how they could keep up, to learn how to steer them and to see how the open models will perform. When I talk about it, quite a bit of "programmers" get upset. Glad to see that Norvig is experimenting.

p/s, anyone who gets upset that folks are experimenting with LLMs to generate code or solve AoC should have their programmer's card revoked.