Indeed, for smaller self-contained chunks of code, it's very useful and efficient. For more ambitious code, it's too risky to have hallucinated (erroneous) chunks of code.
Perhaps the years ahead will lead to other neuronal system architectures that will complement the "textual-generation-prediction" ones that are actually in vogue. On Thursday, April 13, 2023 at 8:34:45 AM UTC-4 tbp1...@gmail.com wrote: > There was a recent thread about this on python-list, including someone's > experiments. Here's what I wrote - > > " People need to remember that ChatGPT-like systems put words together the > way that many humans usually do. So what they emit usually sounds smooth > and human-like. If it's code they emit, it will tend to seem plausible > because lines of code are basically sentences, and learning how to > construct plausible sentences is what these systems are built to do. That's > **plausible**, not "logical" or "correct". > > The vast size of these systems means that they can include a larger > context in figuring out what words to place next compared with earlier, > smaller systems. > > But consider: what if you wrote code as a stream-of-consciousness > process? That code might seem plausible, but why would you have any > confidence in it? Or to put it another way, what if most of ChatGPT's > exposure to code came from StackOverflow archives? > > On top of that, ChapGPT-like systems do not know your requirements nor the > reasons behind your requests. They only know that when other people put > words and phrases together like you did, they tended to make responses that > sound like what the chatbot emits next. It's basically cargo-culting its > responses. > > Apparently researchers have been learning that the more parameters that a > system like this has, the more likely it is to learn how to emit responses > that the questioner likes. Essentially, it could become the ultimate > yes-man! > > So there is some probability that the system will tell you interesting or > useful things, some probability that it will try to tell you what it thinks > you want [to] hear, some probability that it will tell you incorrect things > that other people have repeated, and some probability that it will > perseverate - simply make things up. > > If I were going to write a novel about an alternate history, I think that > a chatGPT-like system would be a fantastic writing assistant. Code? Not so > much." > > On Thursday, April 13, 2023 at 8:15:59 AM UTC-4 Edward K. Ream wrote: > >> On Thursday, April 13, 2023 at 12:15:17 AM UTC-5 Félix wrote: >> >> Here I am, (simple screenshot below) working on leojs, converting the >> stripBOM function from leoGlobals.py from python to typescript. >> >> Have you tried? Any thoughts or particular productivity tips to share? >> >> >> My impression is that chatGPT does well on small tests. I wouldn't trust >> it with larger tasks. >> >> >> (*I eventually plan to use Leo to organize and automate calls to it's >> API, to make some kind of agi-assistant experiment.*) >> >> >> chatGPT has spawned many creative ideas including yours. Please let us >> know what happens :-) >> >> Edward >> > -- You received this message because you are subscribed to the Google Groups "leo-editor" group. To unsubscribe from this group and stop receiving emails from it, send an email to leo-editor+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/leo-editor/a17b33b0-a3bc-4446-8fe2-fcaa2a20c930n%40googlegroups.com.
