- it's impressive at creating clunky but technically correct creative writing to a prompt, even in constrained forms and about obscure subject matter; the famous examples from twitter so far mostly revolve around people attempting to "jailbreak" its restrictions on which questions it will answer, and include the one-act play about an antagonist who tries and fails to convince the protagonist to do meth through an inspirational haiku and its poems about how to hotwire a car, but I also got it to write a pretty decent song about "james ensor, belgium's famous painter" without just cribbing they might be giants
- it is quite bad at reciting scientific facts beyond the "why is the sky blue?" level; I managed to get it to claim that plants don't have specialized organs for breathing, completely fuck up the series of processes in the coagulation cascade, and claim that carbon-12 is a radioisotope, then when I told it it was incorrect it immediately admitted that it had been incorrect and asked for more context (???)
- despite all of the hype from AI boosters about it being able to generate valid code, most of their examples contain obvious errors; it also seems to have aggressively avoided internalizing anything about calendars.
- @zaratustra embedded in my mind the idea of asking it to conlang and see how far it can continue riffing in a principled way. when I gave it a broad prompt I got it to create one language, Sprock, where "Sprock", "language", and "cat" are all homonyms, as are "on the bed" and "what", and a large chunk of the lexicon was stolen from various languages' words for "language". when I prompted it with specific language groups, it just stole all of its words from German and Welsh; it also appears to have heard of the concept of word orders, but you can tell by the glosses it writes that all of its languages are actually SVO even if it claims they're not.
You must log in to comment.
in reply to @vogon's post:
Honestly, the thing I'm most amused about is the effort the devs keep putting in to bulwark their expensive Eliza program against tricks they'll never be able to actually solve because it is, in the end, just an expensive Eliza program. It gets more and more verbose and has more and more special case logic put in but these problems are fundamentally insoluble and they're going to make the whole thing unusable even for the limited cases it might actually be suitable for.
the code proponents make my head spin. like, a) how is this different from co-pilot??? where is the work attributed to??? and b) what if it's fucking wrong??? if i ask it to check itself and it finds errors, how do i know it's right or wrong? this is just what we're reducing information technology down to???
like, at least if someone posts bad code for a common problem on stackoverflow, other people will point out issues with it. who do i check this work with? and as I ask questions in more specific corner, how much can i be trusting it??? like, ugh, i hate it
I've found it to be honestly really useful in a few circumstances. It mainly requires that you are experienced enough with code to know what you're looking for, almost like using it as a better Google.
I've had it write a few complicated rasterization functions that I've been putting off because they're a pain. They all dropped in with no manual changes. I was also able to ask it to refactor the inputs for me so I didn't need to myself. I've asked it to code review itself and suggest optimizations which has been hit or miss but I would say generally correct.
I've also asked it about algorithm questions like using GOAP for npc ai which it has responded very helpfully with, including asking it to write pseudocode for it that got me into the right headspace of understanding the general premise.
It has suggested some very practical code habits that I agree with like focusing on readability before micro optimizations.
I like it more than copilot because I can correct it and ask it questions and literate to what I want.
As with all of this obviously someone who doesn't know how to code trying to rely on it like a crutch is going to end badly. But I see a huge amount of time saved just using it as a reference and a grunt worker.
so in other words it emulates what eln msk's student papers were probably like
The calendar stuff extends to time in general. It adamantly refuses to believe that any car can accelerate to 100km/h in "under 1 hour" while saying "it's possible in under 10 seconds"