←back to thread

504 points Terretta | 6 comments | | HN request time: 0.386s | source | bottom
1. Demiurge ◴[] No.45064780[source]
I've actually seem really good outputs from the regular Grok 4. The issue seemed to be that it didn't explain anything and just made some changes, which like, I said, were pretty good. I never wanted a faster version, I just wanted a bit more feedback and explanations for suggested changes.

I recently found it much more valuable, and why I am now preferring GPT-5 over Sonnet 4, is that if I start asking it to give me different architectural choices, its really quite good at summarizing trade-offs and and offering step-by-step navigation towards problem solving. I am liking this process a lot more than trying to "one shot" or getting tons of code completely rewritten, thats unrelated to what I am really asking for. This seems to be a really bad problem with Opus 4.1 Thinking or even Sonnet Thinking. I don't think it's accurate, to rate models on "one-shoting" a problem. Rate it on, how easy it is to work with, as an assistant.

replies(3): >>45064839 #>>45065447 #>>45067954 #
2. cft ◴[] No.45064839[source]
I have the same experience, except while I agree that GPT-5 is better than Sonnet 4 for architecture and deep thinking, Sonnet 4 still seems to be better for just banging out code when you have a well-defined and a very detailed plan.
3. Demiurge ◴[] No.45065447[source]
Sometimes it's obvious, but in this case, why are you downmodding my comment? I'm genuinely curious, what am I saying, that is so offensive or wrong?
replies(2): >>45067052 #>>45073655 #
4. ◴[] No.45067052[source]
5. Szpadel ◴[] No.45067954[source]
I had that issue with gpt-5 that when it wanted to do something in one way that was just plain wrong in this project, and no matter what I said it just kept doing the same action.

it was completely unsterable. I get why people are often upset by "you're right" of Claude models, but that's what I usually want from model.

I guess there is different in expectations depending on experience level of developer, but I want to have final saying what is the right way

6. oblio ◴[] No.45073655[source]
I didn't downvote, but:

1. A lot of people are interesting in maintaining AI hype.

2. People work differently.