Comment by chaps
15 hours ago
Respectfully, a lot of what you're saying in this thread sounds a lot like the lies that gamblers tell themselves. Saying this as someone with a strong tendency towards addictions.
Some of these things are only possible to really see in hindsight. Yes, you've been working on these things for a while, but these systems are notably different in their capacity and strings they pull on us.
Be well, please.
Yesterday (unrelated to quotamaxxing described in the article), I made an Apple TV macOS menu bar remote app: https://cdn.bsky.app/img/feed_fullsize/plain/did:plc:oxaerni...
Every single prompt worked without issue, and it got most of the way on the first try with the initial prompt (+ a couple visibility bugs due to the agent not having Computer Vision to see said menu bar app) such as:
> Create a SwiftUI menu bar app named `swiftmote` using theto create the most user friendly app following Apple's HID guidelines for creating a remote that can operate a Apple TV on a local network. Instead of reimplementing the protocols needs to interface with an Apple TV, use the Python package `pyatv` and host it within the SwiftUI app as a sidecar along with a Python installation.
I have my own Apple TV I can manually verify that it worked as expected, which is notable because the agent can't test or lie about this pipeline because it does not have access to the Apple TV.
That is not hallucination or psychosis. If you want, I can release all the prompts I used. (EDIT: Sure, why not, here are the prompts. If I don't complain about something in a followup prompt, assume it worked correctly: https://gist.github.com/minimaxir/30fa820daa1392da13026ec6aa... )
A lot of this really seems to be pattern matching on superficial similarities when the meat of the issue is probably the more important one. I think for LLM usage to be compared to gambling, it has to be a) universally negative-sum in the long run, and b) extremely addictive. a) seems very unlikely to be true, even if it is sometimes negative. b) seems to be the case for some subset of programmers, but again I would not assume this just because someone uses some statements that are sometimes uttered falsely by gamblers (the siren song of those statements being that if they were true, the gambling would be a good thing. That means the truth of those statements matters and should not be assumed to be false like they are more obviously so for gambling).
the key comparison is variable reward.
That 'triggers a surge of dopamine and creates highly addictive habits' [thanks Gemini!]
LLM use for code generation does exactly that, sometimes it works amazingly, sometimes it fails inexplicably. Whether it is negative sum or not doesn't really matter. Indeed it may well prove to be negative sum, especially if we step back a bit and consider the business benefit of the code produced, not just lines of code or even features produced.
I think whether it's negative sum or not matters quite a lot, really. The variability of the outcome is a thing that might create some addictive components, but it matters a lot whether it's something with positive effects with some potential negative effects that may need to be managed, or whether it's the same as gambling, and I think it's extremely unlikely that it is the same as gambling.