|
| ▲ | johnfink8 an hour ago | parent | next [-] |
| I see a lot of the "4.7 is a downgrade" sentiment. 4.7 does (mostly) what you ask it to do. 4.6 does what it thinks it should do. As someone with 20 years writing my own code I want the former, but the loud contingent online wants the latter. When you're on a mature codebase with 500k+ lines of code, I haven't seen anything else be as effective as 4.7. |
| |
| ▲ | onlyrealcuzzo 9 minutes ago | parent [-] | | I can tell you for a fact, Claude 4.7 was NOT doing what I told it to do (in fact the clear and complete opposite - repeatedly), a pretty simple architectural refactor, and that Codex did better and DeepSeek much better. It was given very simple ways to verify success. It simply didn't do that and said it's at a good stopping point, despite moving in the WRONG direction not even doing 1% of the task, and being told to see the task through to completion. Meanwhile, Codex broke it down into 3 steps and just got it done... No, "I'm going to give it to you straight, this is a large risky commit that could go sideways, so I'm just not going to do anything instead." Claude worked on it for almost 200 commits over 2 weeks, needing to typically prompt it 3x to even TRY to make any progress instead of just wasting tokens to ignore me and tell me how big and risky it is. Maybe Claude is just particularly terrible at this type of refactor. I'm not sure why that would be. |
|
|
| ▲ | fendy3002 3 hours ago | parent | prev | next [-] |
| As always, I think this happen more to vibe coder. They don't understand that bigger project means worse AI performance. On top of that Opus felt being nerfed at understanding prompt so if your spec is bad you won't get good result. |
|
| ▲ | solenoid0937 4 hours ago | parent | prev | next [-] |
| It's the same phenomenon as when you learn a new vocabulary word you see it everywhere. People heard "Claude is nerfed" and now they see it everywhere, they notice failures a lot more than they would have otherwise. Doesn't matter that Claude is not, in fact, nerfed. Perception is powerful and most humans are not rational. |
| |
| ▲ | arkadiytehgraet 17 minutes ago | parent | next [-] | | This account is an LLM-hype peddler, shilling for Anthropic (check comment history). If they say that Claude is not nerfed, then most likely it is, in fact, nerfed. | |
| ▲ | fendy3002 3 hours ago | parent | prev [-] | | Oh Opus is nerfed sure, but not that hard. Early this year opus 4.6 can understand your prompt and your intention easily, it got worse around mid April. Opus 4.7 even worse than that. However that's just it, you just need to improve and make clearer of your prompt and it will perform just as good. | | |
|
|
| ▲ | dgellow 6 hours ago | parent | prev | next [-] |
| Opus 4.7 has been a real downgrade for me. I’m back to mid 2025 when I had to catch all the completely intermediary goals/assumptions the model is creating for itself |
| |
| ▲ | kaeluka 35 minutes ago | parent | next [-] | | it's sort of good at thinking, writing specs, etc.. Also debugging. But as a coder: I see no advantage to opus 4.6 and I preferred sonnet most times already over opus 4.6. | |
| ▲ | Wowfunhappy 5 hours ago | parent | prev | next [-] | | You can still use older versions of Opus if they work better for you. Just need to set the environment variable. | |
| ▲ | chantepierre 5 hours ago | parent | prev [-] | | I felt that but find it worked way better by invoking it with `claude --effort max` only |
|
|
| ▲ | colechristensen 6 hours ago | parent | prev | next [-] |
| What it does seem like is that they're tuning some knobs up and down or releasing new versions of models or system prompts that result in the model getting dumber and smarter in waves. Opus has been dumb this week. Claude was having a lot of capacity problems and downtime and then this week that has been much less obvious... and the model is dumber. It could also just be luck and my impressions are false... who knows. |
|
| ▲ | Our_Benefactors 9 hours ago | parent | prev [-] |
| It’s because it’s not true, there’s no evidence for it that passes the sniff test. No lab is “shipping a worse model once they’ve got you”. People have a bad few days and blame the model providers instead of stepping back to fix their workflow. |
| |
| ▲ | raincole 6 hours ago | parent [-] | | When it comes to something with random results (unfortunately that's what LLMs are), people will think the odds are rigged against them. It's a good thing that hype-chasers are cancelling though. So we can use the services with a reasonable latency. |
|
