Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 149
- 1
Jump in the discussion.
No email address required.
Notes -
Has anyone here used LLMs to generate useful code in a way that's actually saved you time, on net? Or if you're not particularly technical, given you the means to write code that you wouldn't have easily been able to figure out on your own? I've been playing around with o1-mini and it's impressive, enough so that I'm almost starting to get concerned re: job security and all the copium I see on tech Twitter is not quite as convincing as it used to be.
Here's a concrete example: I needed to scrape livestreams in real time and save them. This site isn't supported by any standard tools (yt-dlp, streamlink, etc). Not exactly rocket science, but I primarily dabble in C/C++/Rust and have a bit of an aversion to web technologies generally. This particular site does some wacky stuff with wasm (some kind of hand rolled DRM attempt) and after about 30 minutes of faffing around I gave up on trying to reverse engineer their API. Grabbing m3u8s from my browser's network inspector didn't work and even replaying requests with curl was serving me 403s; they're doing something trickier than I'm used to. I had some vague ideas about how to achieve what I wanted but figured I'd give ChatGPT a chance to weigh in. I outlined what I considered: maybe a Chrome extension that could capture the .ts streams in real time, or failing that, maybe something with Selenium (which I've never used), or possibly even just capturing raw packets with Wireshark or something similar.
o1-mini wrote me a whole Chrome extension in its first response, manifest et al, and provided detailed instructions on how to install it. This solution was flawed, using Chrome's webRequest API to monitor and filter for .ts files and then send a download request when it detects one. This would actually work well for most sites, but not this one, because of the aforementioned authorization shenanigans. To be fair to ChatGPT, I didn't mention anything about the authorization requirements. I asked if it was possible, with a Chrome extension, to intercept incoming network responses and just dump them to a file or something similar: it responded in the negative; apparently that functionality isn't exposed.
Fair enough! I asked for another solution. It suggested 3 options and wrote a nice wall of text with pros and cons for each one. "Use a custom proxy server," "Selenium + Browser DevTools protocol," and "Wireshark or tcpdump". I hadn't considered the first option, but it's obvious in retrospect. So I asked about it and it walked me through setting up mitmproxy (including the custom cert) and writing a Python addon for it to filter .ts files and dump them all to a folder as they come in. All of the code was perfect on the first try. Seriously, the entire process took me about 15 minutes, about as long as I've spent writing up this post about it. I ran mitmproxy -s scrapets.py and pointed my Firefox proxy settings to localhost. Then it just worked.
Now, why do I find this impressive? The solution was ultimately not that technical, the problem was not particularly difficult, and I could have muddled my way to a similar answer with some Googling and maybe 2-3 hours. Yet it feels significant. I remember spending countless hours as a teen ricing my Gentoo distro, scouring wikis and mailing lists and fiddling with configs and reading docs. It seems like we're headed to a point where if you can just clearly state your problem in English and maybe answer a few clarifying questions you'll get an instant solution, and if it's too complex, you can just continue recursively asking for more detailed explanations until you understand it fully. Now it's likely o1's knowledge is ocean wide and puddle deep, but even this is such an improvement over GPT-3 and GPT-4, the slope of the line is starting to get a little scary. When Terence Tao describes o1 as "roughly on par... with a mediocre, but not completely incompetent, graduate student,", well, I'm not that partial to Yudkowsky but I don't see any future where (at the very least!) programming isn't completely different 10 years from now.
Whatever is coming, the kids growing up with access to ChatGPT are going to be cracked beyond belief.
Yes, welcome to the future. I use LLMs all the time. They excel at building prototypes, little tools, and self contained functions. It can't one shot complex projects (neither can I ) but you can still handhold it and prompt and split the complex project into smaller modules that it can reason about.
So far I have not found a good way of getting it to fix bugs in existing codebases, but I do wonder if that's just a UI issue rather than an intelligence issue, since there's no easy way for me to let it track the data flow between several files.
The more context you add, the better. This is a mistake people make. It can't read your mind and often people just don't give it enough information.
Programming will absolutely be different in 10 years time, but so will a lot of the rest of world.
More options
Context Copy link
More options
Context Copy link