4 ms·

Have you tried Claude Opus 4.5 within Claude Code? it's only been released for two months but its changed the calculus entirely if its been over two months si

Have you tried Claude Opus 4.5 within Claude Code? it's only been released for two months but its changed the calculus entirely if its been over two months since you've tried any LLM generated code solution, or are still occasionally copy pasting code requirements into a browser chat session as if its still 2023, then I can't put any weight into the opinion

input_sh 9mo ago
I've read this same comment for every model since GPT-3.
- yieldcrv 9mo ago
  so have I and this time I wrote it, this broken clock only has to be right once, forever its beneficial to check it
- CamperBob2 9mo ago
  That's because it's been true for every major new model release since GPT-3. HN is full of people who tried the free version of ChatGPT a couple of years ago, got a load of random hallucinated slop, and concluded it was all a bunch of useless hype. They enjoy parroting a lot of obsolete stuff they read once about stochastic parrots, without the slightest sense of irony. When I was growing up, my old man recalled engineers who reacted the same way to transistors, which sucked even more than early-generation LLMs when they first came out. Most of them got over vacuum tubes eventually, though. So will most of the HN'ers, likely including you.
  - senordevnyc 9mo ago
    This is so true, and it's been genuinely painful for me to realize how fucking lazy and close-minded many in our industry actually are when you get down to it.
  - input_sh 9mo ago
    I don't know whose thoughts you're trying to attribute to me, but I've been pretty up-to-date myself (while still keeping myself to a $20/month budget that I switch between providers every few months). Every generation has at best been a 5% improvement, and nothing revolutionary compared to the last generation. Absolutely no significant improvements between me selecting say Claude 3.7 Sonnet, Claude Sonnet 4, or Claude Sonnet 4.5. Still the same somewhat useful tool for specific purposes, still not good enough to be let loose on production, still not better than what I can do myself with 10 or so years of experience. Genuinely useful from time to time? Sure, I agree completely. Anywhere near being revolutionary enough as people here insist on gaslighting themselves to be? Absolutely not. Not even remotely.
    - CamperBob2 9mo ago
      I'll grant that progress in coding hasn't been as impressive, although if you don't see a difference between Claude 3.7 Sonnet and 4.5 Opus (which is the actual SotA for Claude) I'd suspect a skill issue. I haven't seen miraculous progress in baseline code generation, but there has been a lot of progress in tool use. I can tell Claude CLI to install a complex package, for instance, and it will handle all the Python versioning and dependency-hell issues for me, then test and evaluate the results. 3.7 Sonnet couldn't do that, at least not reliably. What's really different now is that LLMs have become useful research tools. Three or four years ago, models couldn't cite their sources at all. Two or three years ago, they started to gain the capability, but a large proportion of the citations would either be hallucinated or irrelevant. About a year ago, significant improvements started to emerge. At this point, I can give Gemini 3 Pro or GPT 5.2 Pro a multilayered research task, and end up with a report indistinguishable in accuracy and bibliographic quality from what a good human might produce in a couple of weeks. It might take a half hour to get the answer, and I'm not sure if you could get the same result at the $20/month level, but the hype and promises we started hearing a couple of years ago are starting to bear fruit. The research models are now capable of performing at grad-student level. Not all the time, and not without making stuff up on occasion... but to argue that no progress at all has been made is nothing more or less than moon-landing denial.
      - 9mo ago

input_sh 9mo ago

I've read this same comment for every model since GPT-3.

yieldcrv 9mo ago
so have I and this time I wrote it, this broken clock only has to be right once, forever its beneficial to check it
CamperBob2 9mo ago
That's because it's been true for every major new model release since GPT-3. HN is full of people who tried the free version of ChatGPT a couple of years ago, got a load of random hallucinated slop, and concluded it was all a bunch of useless hype. They enjoy parroting a lot of obsolete stuff they read once about stochastic parrots, without the slightest sense of irony. When I was growing up, my old man recalled engineers who reacted the same way to transistors, which sucked even more than early-generation LLMs when they first came out. Most of them got over vacuum tubes eventually, though. So will most of the HN'ers, likely including you.
- senordevnyc 9mo ago
  This is so true, and it's been genuinely painful for me to realize how fucking lazy and close-minded many in our industry actually are when you get down to it.
- input_sh 9mo ago
  I don't know whose thoughts you're trying to attribute to me, but I've been pretty up-to-date myself (while still keeping myself to a $20/month budget that I switch between providers every few months). Every generation has at best been a 5% improvement, and nothing revolutionary compared to the last generation. Absolutely no significant improvements between me selecting say Claude 3.7 Sonnet, Claude Sonnet 4, or Claude Sonnet 4.5. Still the same somewhat useful tool for specific purposes, still not good enough to be let loose on production, still not better than what I can do myself with 10 or so years of experience. Genuinely useful from time to time? Sure, I agree completely. Anywhere near being revolutionary enough as people here insist on gaslighting themselves to be? Absolutely not. Not even remotely.
  - CamperBob2 9mo ago
    I'll grant that progress in coding hasn't been as impressive, although if you don't see a difference between Claude 3.7 Sonnet and 4.5 Opus (which is the actual SotA for Claude) I'd suspect a skill issue. I haven't seen miraculous progress in baseline code generation, but there has been a lot of progress in tool use. I can tell Claude CLI to install a complex package, for instance, and it will handle all the Python versioning and dependency-hell issues for me, then test and evaluate the results. 3.7 Sonnet couldn't do that, at least not reliably. What's really different now is that LLMs have become useful research tools. Three or four years ago, models couldn't cite their sources at all. Two or three years ago, they started to gain the capability, but a large proportion of the citations would either be hallucinated or irrelevant. About a year ago, significant improvements started to emerge. At this point, I can give Gemini 3 Pro or GPT 5.2 Pro a multilayered research task, and end up with a report indistinguishable in accuracy and bibliographic quality from what a good human might produce in a couple of weeks. It might take a half hour to get the answer, and I'm not sure if you could get the same result at the $20/month level, but the hype and promises we started hearing a couple of years ago are starting to bear fruit. The research models are now capable of performing at grad-student level. Not all the time, and not without making stuff up on occasion... but to argue that no progress at all has been made is nothing more or less than moon-landing denial.
    - 9mo ago