Post AYjzAPsyT0c74C4FvM by rudi@nightmare.haus
(DIR) More posts by rudi@nightmare.haus
(DIR) Post #AYjyg3Z2tQQB9OGlZg by simon@fedi.simonwillison.net
2023-08-15T02:14:06Z
0 likes, 0 repeats
Answered a question on Hacker News about using ChatGPT for programming tasks with a bunch of my own examples... https://news.ycombinator.com/item?id=37126182#37127206Comments so far include "If someone on my team was doing things this unusually they’d probably be let go" and "sorry but these examples are not impressive at all and by no means a representative of any serious programmer's workload"A lot of people out there are absolutely determined not to take the practical utility of this stuff seriously!
(DIR) Post #AYjyrhmTi0AL486gka by simon@fedi.simonwillison.net
2023-08-15T02:15:33Z
0 likes, 0 repeats
There are SO MANY valid criticisms of LLMs that I can respect, but denying how much they can accelerate the act of throwing together useful code isn't one of them
(DIR) Post #AYjzAPsyT0c74C4FvM by rudi@nightmare.haus
2023-08-15T02:19:32Z
0 likes, 0 repeats
@simon LLMs are great when used properly, a lot of people would rather put in as little effort as possible and make fun of the mistakes LLMs produce instead. Doesn't matter how easy or difficult your task is, all that matters is if you and the LLM can discuss it in a way that'll give you good results without wasting unnecessary time
(DIR) Post #AYjzjh9aG0NQl1etwe by evan@cosocial.ca
2023-08-15T02:26:01Z
0 likes, 0 repeats
@simon agreed.
(DIR) Post #AYjzwefanq6qopoIoy by nilesh@fosstodon.org
2023-08-15T02:28:24Z
0 likes, 0 repeats
@simon Absolutely. There's a lot of "easy" but time-taking work involved in building anything.Have you written a post on your current AI tooling/workflows? Is it GitHub CoPilot or something else? Any other VS Code extensions or command-line tools?
(DIR) Post #AYk092rAhdhhJ98OUi by scanner@apricot.social
2023-08-15T02:29:04Z
0 likes, 0 repeats
@simon My take is that "it will get better" (as long as people are trying to make it better.)I mean, I feel that it is going to happen so we should prepare for it. That means ways of testing, verifying, constraining, improving. Even if it does not pain out, the effort in finding ways to validate the code it produces will be valuable regardless.
(DIR) Post #AYk0KD8onyzgOGiG4e by DocBohn@techhub.social
2023-08-15T02:29:52Z
0 likes, 0 repeats
@simon trust, but verify.
(DIR) Post #AYk0WQVKwlnpJK4Qls by simon@fedi.simonwillison.net
2023-08-15T02:31:42Z
0 likes, 0 repeats
@nilesh it's mostly GPT-4, either through the ChatGPT web interface or using my own https://llm.datasette.io CLI tool - I use Copilot mainly as a typing assistant and occasionally have it write me a test or two
(DIR) Post #AYk2BtTiEBBSzxeDMe by NIH_LLAMAS@mastodon.social
2023-08-15T02:53:36Z
0 likes, 0 repeats
@simon Important to notice how asking an LLM to code is the opposite of asking it to get information about reality. The act of coding is creating an environment to support a desired function. As long as the environment you create (built from learned samples) results in the desired function, the created environment is "correct".The chat bots seem to do that, even if you ask them to describe actual reality. They appear to use simple consensus to select an answer, then construct...
(DIR) Post #AYk47gX2cTq9b1KFBA by simon@fedi.simonwillison.net
2023-08-15T03:15:20Z
0 likes, 0 repeats
@NIH_LLAMAS Right - a fascinating thing about code is that it ends up being quite a solid workaround for the hallucination problem - because if the LLM hallucinates a solution it usually won't execute or compile!
(DIR) Post #AYk4LUxYRTfVl8JVLc by cube_drone@mastodon.social
2023-08-15T03:17:39Z
0 likes, 0 repeats
@simon copilot produces a lot of wrong code and I think people underestimate the value of having wrong code appear fast I'm faster at debugging than at coming up with wholly new code a lot of the time, so I feel like it makes me a concretely faster coder
(DIR) Post #AYk5Crtjt7laFWygIy by jrm4@mastodon.social
2023-08-15T03:27:21Z
0 likes, 0 repeats
@simon Wow, working coders and engineers AREN'T? That's madness. I'm presently using it to *blaze* through my personal doofy todo list idea thing that I've always had in my head, and this is just in my free time between semesters...
(DIR) Post #AYk6RSLZGjCnImVwrw by dhoe@sigmoid.social
2023-08-15T03:41:09Z
0 likes, 0 repeats
@simon I think lots of people are sort of working through a denial stage. There are people out there denying there is any valid use case at all for LLMs, which is insane to me
(DIR) Post #AYkDDr3NJT1HYLSxk0 by lewiscowles1986@phpc.social
2023-08-15T04:57:05Z
0 likes, 0 repeats
@simon it depends. They seem alright at plumbing and the esoteric syntax lookup. But given 10 revisions on auto mode I kicked chat gpt ass yesterday for special weighted results in predictable order. I gave it two rounds. It's bias towards using libs like numpy and inability to reason left it making cyclic errors, not learning a damn thing. So YMMV even if it can approximate understanding questions.
(DIR) Post #AYkExEC4E0RgJ3jP0q by tom_armstrong@mastodon.cloud
2023-08-15T05:16:31Z
0 likes, 0 repeats
@simon I mean, I get it. It's such a mixed bag of existential dread, bullshit hype, and actual genuine usefulness.Our company is trialling Copilot, and I find it really interesting how varied the responses from developers using it.Everything from "it's terrible and I hate it" to "on balance it probably helps slightly more than it hurts?" (me) to "it's wonderful and life-changing"
(DIR) Post #AYkSnGqWuG8Gznnsoa by gla@mastodon.social
2023-08-15T07:51:32Z
0 likes, 0 repeats
@simon I agree with you. I’ve written about it here, citing your work as wellhttps://xebia.com/blog/how-i-learned-to-stop-worrying-and-love-llms/
(DIR) Post #AYkUZuPRDOYNYJt6zg by jamwil@fosstodon.org
2023-08-15T08:11:49Z
0 likes, 0 repeats
@simon Hacker News is depressingly cynical.
(DIR) Post #AYkUnZmmgnV7e0KDTc by benjamineskola@hachyderm.io
2023-08-15T08:14:03Z
0 likes, 0 repeats
@simon the usefulness of the code seems to be subjective. I’ve found the need to correct the code they generate to make them no more useful than e.g. StackOverflow, personally. I can see how other people’s development workflows might differ, though.
(DIR) Post #AYkXA2zB0OBWU8xq3E by joelanman@hachyderm.io
2023-08-15T08:40:31Z
0 likes, 0 repeats
@simon isn't the danger that people throw together code that is subtly broken eg accessibility, security. Unless you're an expert you'll miss it
(DIR) Post #AYkXf5OC7RP6utSzAG by tommorris@mastodon.social
2023-08-15T08:46:19Z
0 likes, 0 repeats
@simon I still have some concerns about copyright. Like if I’m doing a work for hire, I can’t legitimately tell a client that it’s all my own work and this assigned to them under the terms of the contract as the legal situation is still somewhat unclear.
(DIR) Post #AYkcuJzGcs9eJ1tJPk by andyide@fosstodon.org
2023-08-15T09:44:50Z
0 likes, 0 repeats
@simon I love that it is very good at writing JavaScript for me, a language I don't care for.
(DIR) Post #AYl2hRmavcptWkU9Zo by simon@fedi.simonwillison.net
2023-08-15T14:33:38Z
0 likes, 0 repeats
@benjamineskola I've found the quality of code it produces gets dramatically better as I learn how best to prompt it - it helps that I work mainly in Python and JavaScript, which I think are the languages it has seen the most of in its training data
(DIR) Post #AYl37s9qC7dxzt4kUK by simon@fedi.simonwillison.net
2023-08-15T14:34:50Z
0 likes, 0 repeats
@joelanman sure, which is why being an expert is so useful in making the most of this tech!
(DIR) Post #AYl3K6jVpUZGg1awKG by simon@fedi.simonwillison.net
2023-08-15T14:35:48Z
0 likes, 0 repeats
@tommorris yeah, there are so many unanswered questions and grey areas thereI certainly wouldn't think poorly of a client that pushed back against using code from LLMs directly
(DIR) Post #AYl3MDLcCwI3lqQClU by simon@fedi.simonwillison.net
2023-08-15T14:34:23Z
0 likes, 0 repeats
@oscherler @NIH_LLAMAS I don't see how that's any different from data loss caused by mistakes in code from humans that makes it past code review
(DIR) Post #AYl3a4VDtFNWOOWVV2 by tommorris@mastodon.social
2023-08-15T14:42:04Z
0 likes, 0 repeats
@simon I think most startups are likely to not care. In finance, big corps etc might be a lot more of a problem.
(DIR) Post #AYlI4PeDEjkg1C5dh2 by joeri_s@mstdn.social
2023-08-15T17:26:05Z
0 likes, 0 repeats
@simon Plenty of people seem incapable of accepting innovation as useful. Douglas Adams described it well: https://www.goodreads.com/quotes/39828-i-ve-come-up-with-a-set-of-rules-that-describe
(DIR) Post #AYlIjwFmufEb17Szce by simon@fedi.simonwillison.net
2023-08-15T17:33:39Z
0 likes, 0 repeats
@tommorris yeah, for my current startup I've decided to take the very small risk that I might one day have to rewrite a bunch of LLM-generated code
(DIR) Post #AYlIwsOiYouap3oRt2 by eichin@mastodon.mit.edu
2023-08-15T15:29:10Z
0 likes, 0 repeats
@tommorrisStartups that take investment do, same as they care about open source licenses, because they don't want them ruining a liquidity opportunity. (Naïve startups might not care about either, sure - but they'll suffer for it.)@simon
(DIR) Post #AYlIwtLv0iGbmfVjA8 by simon@fedi.simonwillison.net
2023-08-15T17:36:04Z
0 likes, 0 repeats
@eichin @tommorris have your heard about any investors who warn their startups against using LLM-generated code?I feel like I've been hearing almost the opposite: stories of investors who pressure their companies to adopt LLMs more aggressively
(DIR) Post #AYm3zeyxahrjvhatGK by eichin@mastodon.mit.edu
2023-08-16T02:23:01Z
0 likes, 0 repeats
@simon @tommorris Not yet, but it isn't until the later rounds that investors start talking audits and intellectual property anyway.
(DIR) Post #AbeB5hr50RHr8mxwoK by stuartcw@mastodon.social
2023-11-10T00:40:28Z
0 likes, 0 repeats
@simon There’s an identical mindset going on on Reddit if you post anything in the learning Japanese subreddits that mention ChatGPT. Most questions on the subreddits can answered by ChatGPT with examples and detailed explanations but the mere mention of ChatGPT raises the ire of naysayers.