points by knollimar 1 day ago | 0 comments
I always go insane trying to tell my software friends this and they wax poetic about how capable the models are.
Trying to insert your own harness into another field makes it very clear and the friends approach AI psychosis levels of overconfidence
[–] api 1 day ago
[–] Personability 1 day ago
Coding itself is really quite easy, for an individual that can properly apply logic. However, it's like learning a language (Because you quite literally are.) in that much of learning to code is remembering what words do what.
The main reason it's considered so hard is that, as I've learned over the years, a large majority of the general population struggles when it comes to formal logic. It's the same reason math is considered challenging.
[–] bsenftner 1 day ago
And then I try explaining to people that writing software and using AI and that combination of using AI to write software are all communications problems. There is an entire formal school called effective communications that zeros in on the communications aspects of, well, doing anything which include by yourself as well as collaboration teams as well as AI in any of these and that is a communications situation, and by focusing on the quality of those communications a majority of the stress and confusion evaporates. Then they argue because they argue at anything, and that ends any forward progress.
[–] knollimar 1 day ago
I think a good harness hides a ton of the abstraction work that makes solving these problems easier.
I work in construction in drafting for electrical and the models go off the rails all the time. Just adding grounds “to be safe” (and it sometimes makes it less safe by doing so)
[–] bigstrat2003 1 day ago
Models go off the rails in programming all the time too. Despite the hype, they still very much suck at programming.
Arcane is not the same as actually intellectually hard. Programming is arcane.
Same with circuit layout, which is also amenable to things like genetic algorithms, and yes I believe math proofs might be easier than we thought.
It seems to be that you can mix some intelligence and a learned bag of motifs with iterative gradient descent and solve math proofs like a hard jagged optimization problem. I feel like this is what has been demonstrated. You get a huge blob of Lean but it works.
We might learn there are domains like, say, nursing or writing (good writing) or blue collar work or being a secretary for a large company that require a lot more fluid intelligence than fields that have historically been smug about their high IQ requirements.