|
Top-level
 @david_chisnall your experience is how I expected mine to be if I had actually given the technology a chance. Machine learning has been useful for decades, mostly quietly. The red flag against LLMs for me (aside from the authorship laundering and mass poaching of content) was there scramble by all companies to shoehorn it into their products, like a solution looking for a problem; I’ve yet to see it actually solve. 39 comments
@david_chisnall that’s a nicely eloquent way to put both into perspective. @david_chisnall @carbontwelve this is what has been gnawing at the back of my brain. The purveyors of LLM's have been talking up the latest improvements in reasoning. A calculator that isn't 100% accurate at returning correct answers to inputs is 100% useless. We're being asked to conflate the utility of LLM's with the same kind of utility as a calculator. Would we choose to drive over a bridge designed using AI? How will we know? @david_chisnall @zebratale @carbontwelve "do make mistakes" I wouldn't call that a mistake. The calculator does what it should do according to the spec how to approximate real numbers with a finite number of bits. It's (as you explain) a rounding error. A "mistake" is what Pentiums with the famous Pentium bug made. But maybe it's my understanding of English (as a second language) that is at fault here.  @glitzersachen @david_chisnall @zebratale @carbontwelve the calculator /is/ doing exactly what it's been programmed to... and it is programmed to make specific and defined "mistakes" or errors in predictable and clear cut ways in order to make the pocket calculator run on as little power as possible. An LLM, likewise, is also doing exactly what it was programmed to do... and that is to spew regurgitated nonsense it read off the internet.  @pendell @glitzersachen @david_chisnall @zebratale @carbontwelve floating point finance calculations is a common mistake... @pendell @glitzersachen @david_chisnall @zebratale @carbontwelve programmers and CPU designers are just a tad sensitive and insecure when someone points out the calculator makes a mistake and isn't mathematically perfect 😅  I do find myself building up intuitions for what an LLM does. It's far less reliable than a calculator but humans can build intuitions for other unreliable things that can fail excitingly. This 👉 “The benefits of being able to generate correct code faster 80% of the time are small but the costs of generating incorrect code even 1% of the time are high.” @david_chisnall @carbontwelve This sounds a classic quality management problem: the cost of the shit. @david_chisnall @carbontwelve part of the AI hype brainrot seems to be a complete disregard to the possibility of failure. or if there's a consideration of which, then the next bigger, hungrier model will fix it... @nebulos @david_chisnall @carbontwelve That's capitalism 101: Create a problem that didn't exist before; conjure a solution you can build a business model around; kneecap whatever regulatory mechanisms that would eliminate the problem for good; leech off society for life!  @carbontwelve @david_chisnall this also matches my expectations, and I've seen people mention studies in teams showing no productivity gain, too. So I'm intrigued by the few people who DO report that LLMs help them code, though (eg @simon ). Is there something different about how their brains work so LLMs help? Or (cynically) are they jumping on the bandwagon and trying hard to show the world they've cracked how to use them well, to sell themselves as consultants or something? @kitten_tech @carbontwelve @david_chisnall @simon Something I've found LLMs useful for, and I've seen Simon say something similar, is writing code in a language or situation that I _kinda_ know. I might not have bothered writing it if I had to climb boilerplate mountain first, but the LLM serves as guardrails and crack-filler. And since I _kinda_ know the thing, I don't fall for appalling hallucinations, but I get a chance to learn more about the thing fairly painlessly.  @jt7d @kitten_tech @carbontwelve @david_chisnall @simon this is a really good point - if you treat it as a intern who knows a little about the topic but might be wrong, it can still point you to the correct solution. This is, perhaps the meatspace equivalent of the caching application described elsewhere in the replies to this thread. This is how I also use the (mandatory 🤷🏻♀️) LLM results in Google search - as a way to find links and or search terms that will get me to the actual answer. @jt7d @kitten_tech @carbontwelve @david_chisnall @simon Yeah, earlier I mentioned copilot can help me with the flow. It generates some bs and I am that's not what I want, I want this and do it. Or it gives me ideas I hadn't had before. (Besides autocompleting shorter or simpler bits I can easily approve) @kitten_tech @carbontwelve @david_chisnall @simon This m matches my experience, and I gave up after trying on four small nontrivial projects. The people I personally know using it successfully only use it in limited cases as a kind of hint system to help remind of things or point out new ways of doing things.  @kitten_tech @carbontwelve @david_chisnall I'm actually getting more coding work done directly in the Claude and ChatGPT web interfaces and apps vs using Copliot in my editor The real magic for me at the moment is Claude Artifacts and ChatGPT Code Interpreter - I wrote a bunch about Artifacts here: https://simonwillison.net/tags/claude-artifacts/ Here are all of my general notes on AI-assisted programming: https://simonwillison.net/tags/ai-assisted-programming/ @simon @kitten_tech @carbontwelve @david_chisnall How would you avoid or deal with the issues that David encountered? Specifically, subtle bugs that the process of debugging make the whole process less efficient than writing it yourself. Is there one of your notes that deals with that already? @utterfiction @carbontwelve @david_chisnall He has a few examples where he felt something in the output didn't look right, or ran it and found bugs, and had the LLM try again. Most of his examples are relatively simple things of the form "I didn't want to spend time reading API docs for this quick task", though. I don't find that sort of thing a bottleneck in what I do - and I quite enjoy reading docs, and building a mental model of a tool I can then use to know what its... @utterfiction @carbontwelve @david_chisnall ... limitations and capabilities are. The bits of programming that eat my time, which I'd love a tool to help with, are usually understanding a bug in an undocumented and under commented ball of hundreds of kloc of code, too big for an LLM's context window, and where going and quizzing the people who wrote bits of it is essential to success. The bits Simon gets LLMs to do look like the tasks I do to cheer myself up after that :-) @kitten_tech @carbontwelve @david_chisnall Yeah. A lot of my professional time is spent extending logic, adding new features that follow an existing pattern, refactoring when re-usable abstractions are discovered… so far, they’re just not very good at that. And I don’t think pure LLMs ever will be - limited token windows and no genuine symbolic representation of knowledge. @utterfiction @carbontwelve @david_chisnall If you can suddenly create small throwaway applications far more quickly than before, applications that might be too boring or bothersome to create otherwise, that might allow new ways of working altogether. @utterfiction @kitten_tech @carbontwelve @david_chisnall you have to assume that the LLM will make weird mistakes all the time, so your job is all about code review and meticulous testing I still find that a whole lot faster then writing all the code myself Here's just one of many examples where I missed something important: https://simonwillison.net/2023/Apr/12/code-interpreter/#something-i-missed @utterfiction @kitten_tech @carbontwelve @david_chisnall but honestly, the disappointing answer is that most of this comes down to practice and building intuition for tasks the models are likely to do well vs mess up Manipulating some elements in the HTML DOM with JavaScript? They'll nail that every time Implementing something involving MDIO registers? My guess is there are FAR less examples relating to that in the (undocumented, unlicensed) training data so much more likely to make mistakes @kitten_tech @carbontwelve @david_chisnall Wouldn't touch it with a bargepole myself, but I think a third possibility is that at least some people reporting that haven't had it write a sufficiently hilarious bug _yet_. After all, the OP hit one every four months - one could easily get lucky if that's a typical frequency. @kitten_tech @carbontwelve @david_chisnall I am a couple of years in with copilot. No such bugs yet. Context: Rust. I write lots of tests along with my code (as OP appears to do too), and currently can rely on a massive external test suite. The one ridiculously hard to debug bug I got is when I had to debug a codebase I ported from Java to Rust and I transliterated bits wrong as the human, and had no incremental tests built up along with the code. No LLM to blame. 
If you're looking for plausible and authoritative-sounding pronouncements, you've come to the right place! @moof What kind of boilerplate are you having to write so often that any decent snippet engine couldn't handle perfectly well without the litany of issues of an LLM? We *already have* tools for boilerplate, I don't understand why people are so entranced by LLM's ability to deal with it in an absurdly, grossly inefficient way. @carbontwelve @david_chisnall . Note that how @simon reports using this to generate little projects is an entirely different mode of working with them. I have used copilot for a few years now and like it myself, which is mostly context sensitive autocomplete. A Q&A session to create code for a CLI tool or web app is a very different way of working I started exploring more recently. It's surprisingly capable for little projects and requires a different approach. @faassen @kitten_tech @carbontwelve @david_chisnall Steve Yegge calls it CHOP, for Chat Oriented Programming https://simonwillison.net/2024/Jul/12/the-death-of-the-junior-developer/ |
Gregory's Server
@carbontwelve I used machine learning in my PhD. The use case there was data prefetching. This was an ideal task for ML, because the benefits of a correct answer were high and the cost of an incorrect answer were low. In the worst case, your prefetching evicts something from cache that you need later, but a 60% accuracy in predictions is a big overall improvement.
Programming is the opposite. The benefits of being able to generate correct code faster 80% of the time are small but the costs of generating incorrect code even 1% of the time are high. The entire shift-left movement is about finding and preventing bugs earlier.
@carbontwelve I used machine learning in my PhD. The use case there was data prefetching. This was an ideal task for ML, because the benefits of a correct answer were high and the cost of an incorrect answer were low. In the worst case, your prefetching evicts something from cache that you need later, but a 60% accuracy in predictions is a big overall improvement.