5 ms·

> ChatGPT is just hastening an already well deserved death. If ChatGPT kills Stack Overflow and Reddit/Twitter block API access, where will it gets its answers

• mike_hearn 3y ago
  Probably by doing the thing that can be used to answer 95% of all questions programmers ask each other: reading the manual. Over time I've got used to it, but my colleagues tend to use me as a personal StackOverflow because I actually read the manuals for things we use/depend on. I regularly "amaze" them by answering obscure questions and when they ask how did you know that, the answer is always some variant on "by reading". Now I just accept that most programmers hate reading and will prefer either just splashing around until they stumble on the answer by accident, or asking other people. It's frustrating but just how it is. I've had a senior programmer at a well known tech company tell me that now LLMs exist they+their team suddenly feel like writing documentation is worth it, because something will actually read it, and he was serious! They work on the assumption that if they write docs, zero programmers will read them! I don't think that's literally true but it might be directionally true. I don't quite understand the assumption that LLMs are learning to code by reading StackOverflow anyway. Sure, it sounds plausible but we don't know how ChatGPT was trained. But it seems irrelevant now. GPT-4 has clearly learned programming to the level that it could generate self-training material just by reading the docs, writing experimental programs to try things out, and then summarizing its own findings which can be fed back into the next training round. At the moment it's not wired up to programming environments closely enough to do this, but it could be.
  • baud147258 3y ago
    > Now I just accept that most programmers hate reading and will prefer either just splashing around Usable documentation has to exists in the first place, though.
    • mike_hearn 3y ago
      It's a bit of a circular problem. If programmers don't enjoy reading, they assume nobody else does either, so they don't bother writing proper docs, so the people who do read find there's no good docs and learn not to bother next time.
  • zer8k 3y ago
    I think this is unnecessarily reductive. When your day job is programming and you have strict sprint deadlines you don't have the hour or two it might take to leisurely stroll through manpages. Quickness is a result of industry demand not because it's correct. But that also doesn't make the programmers "hate reading". I have never had a PM/manager/TL/etc budget "documentation reading" into sprint points. I will hazard a guess and say you haven't either. So if you're not reading manpages in your free time Google/ChatGPT/SO/whatever is usually the best option.
    • baud147258 3y ago
      > I have never had a PM/manager/TL/etc budget "documentation reading" into sprint points. When working with new software components, usually we're able to either work into our sprints an experimentation/PoC task, that'd include reading the doc or we just include into the estimation of the development task some time to read
    • pjmlp 3y ago
      How do you think we used to develop software before the Internet became widespread?
    • mike_hearn 3y ago
      I think reading documentation along with many other tasks is meant to be just budgeted into a task when planning it. You probably wouldn't write buggy code that doesn't work then say the sprint didn't have an explicit line item for debugging?
• twelve40 3y ago
  > If ChatGPT kills Stack Overflow and Reddit/Twitter block API access, where will it gets its answers from? nowhere, they are cannibalizing their own source of data. They own Github so that's helpful, but still mostly just code, not a whole lot of discussion.