And I'm sure if you go back to the release of 3.5, you'll see the exact same comments.
And when 5 comes out, I'm sure I'll see you commenting "OK I agree 6 months ago but now with Claude 5 Opus it's great".
It's really the weirdest type of goalpost moving.
I have used Opus 4.5 a lot lately and it's garbage, absolutely useless for anything beyond generating trivial shit for which I'd anyway use a library or have it already integrated in the framework I use.
I think the real reason your opinion has changed in 6 months is because your skills have atrophyed.
It's all as bad as 6 months ago, and even as bad as 2 years ago, you've just become worse.
> Not from people whose opinions on that I respect.
Then you shouldn't respect Antirez's opinion, because he wrote articles saying just that 2 years ago.
> If you think LLMs today are "as bad as 2 years ago" then I don't respect your opinion. That's not a credible thing to say.
You are getting fooled by longer context windows and better tooling around the LLMs. The models themselves have definitely not gotten better. In fact it's easy to test, just give the exact same prompt to 3.5 and 4.5, and receive the exact same answer.
The only difference is that when you used to copy-paste answers from the ChatGPT UI, you now have it integrated in your IDE (with the added bonus of it being able to empty your wallet much quicker). It's a faster process, not a better one. I'd even argue it's worse, since you spend less time reviewing the LLM's answer in this situation.
How do you explain that it's so easy to tell (in a bad way) when a PR is AI-generated if it's not necessary to code by hand anymore?
> Despite the large interest in agents that can code alone, right now you can maximize your impact as a software developer by using LLMs in an explicit way, staying in the loop.
> Anyone claiming "Writing code is no longer needed for the most part" is not a serious software engineer.
You need to recalibrate. Six months ago I would have agreed with you, but Opus 4.5 and GPT-5.2 represent a very real change.
I would phrase this as "typing the code out by hand is no longer needed for the most part", which I think is what antirez was getting at here.
And I'm sure if you go back to the release of 3.5, you'll see the exact same comments.
And when 5 comes out, I'm sure I'll see you commenting "OK I agree 6 months ago but now with Claude 5 Opus it's great".
It's really the weirdest type of goalpost moving.
I have used Opus 4.5 a lot lately and it's garbage, absolutely useless for anything beyond generating trivial shit for which I'd anyway use a library or have it already integrated in the framework I use.
I think the real reason your opinion has changed in 6 months is because your skills have atrophyed.
It's all as bad as 6 months ago, and even as bad as 2 years ago, you've just become worse.
> And I'm sure if you go back to the release of 3.5, you'll see the exact same comments.
Not from people whose opinions on that I respect.
Credible software developers I know were impressed by Claude 3.5 but none of them were saying "I don't type out code by hand any more". Now they are.
If you think LLMs today are "as bad as 2 years ago" then I don't respect your opinion. That's not a credible thing to say.
> Not from people whose opinions on that I respect.
Then you shouldn't respect Antirez's opinion, because he wrote articles saying just that 2 years ago.
> If you think LLMs today are "as bad as 2 years ago" then I don't respect your opinion. That's not a credible thing to say.
You are getting fooled by longer context windows and better tooling around the LLMs. The models themselves have definitely not gotten better. In fact it's easy to test, just give the exact same prompt to 3.5 and 4.5, and receive the exact same answer.
The only difference is that when you used to copy-paste answers from the ChatGPT UI, you now have it integrated in your IDE (with the added bonus of it being able to empty your wallet much quicker). It's a faster process, not a better one. I'd even argue it's worse, since you spend less time reviewing the LLM's answer in this situation.
How do you explain that it's so easy to tell (in a bad way) when a PR is AI-generated if it's not necessary to code by hand anymore?
Claude 3.5 didn't have "reasoning" - Anthropic first added that in 3.7 less than a year ago.
The RL for code problems that supported reasoning modes has been the driving force behind most of the model improvements for code over 2025: https://simonwillison.net/2025/Dec/31/the-year-in-llms/#the-...
> Then you shouldn't respect Antirez's opinion, because he wrote articles saying just that 2 years ago.
Which articles? What did he say?
https://antirez.com/news/154 is one from six months ago where he says:
> Despite the large interest in agents that can code alone, right now you can maximize your impact as a software developer by using LLMs in an explicit way, staying in the loop.
>If you think LLMs today are "as bad as 2 years ago" then I don't respect your opinion. That's not a credible thing to say.
This exact comment started getting old a year ago.
I can't tell if you are agreeing or disagreeing with me here.
What's wrong with you? Let people express their experience without calling them mentally ill. Put yourself together.
The comment was flagged and killed; the system works.
Please don't respond to personal attacks with personal attacks.
Personal attack for calling out the hostility? And btw it was not flagged nor killed at the moment when I wrote my comment.
Language like "What's wrong with you?" is a clear personal attack.