Artificial intelligence still has some way to go

yesterday I was too lazy to explain all the permits required for a grid-tied solar system to somebody, so I threw it into GPT and it pretty much nailed it

I think at some point the online ad market is going to crash a bit after eight zillion websites are created that are just zombie copies of other content spewed out to generate clicks

I was looking for some video game information the other day and half of the websites were just robot-regurgitated copies of the other half

yeah, i was fairly impressed that it did a decent write-up on 401(k) automatic rebalancing, but there are repositories of that information literally everywhere and it's a basic investing concept that's existed for many decades, so it had a lot to plagiarize from.

I still had to edit it because it wrote too much, even while being told in the instructions it was for training material.

ha I generally say "write a BRIEF description of xxx" for that very reason

I was looking for a bar to go to last week and two of the bars listed websites that were no longer functional on their FB page. when I clicked the link, both URLs wound up being zombie blogs about construction jobs and materials

was thinking more about when it came to looking up existing verdicts and legal arguments, like you could just put into ChatGPT "has there ever been a slip and fall case where x, y, and z happened"

― frogbs, Tuesday, March 14, 2023 10:42 AM (nineteen minutes ago) bookmarkflaglink

ah yeah prob true

― lag∞n, Tuesday, March 14, 2023 10:42 AM (eighteen minutes ago)

yeah there are specific use cases for searching where it could be v useful, but, as lagoon said, what you wouldn't want in this scenario is a paragraph (or 9) of unsourced text

I use it for product descriptions, it does a pretty good job but you definitely have to proofread

― obsidian crocogolem (sleeve), Tuesday, March 14, 2023 9:51 AM (fifteen minutes ago) bookmarkflaglink

yeah i think it just depends, stuff i'm doing is always tied to upcoming marketing messaging plans for the next year of product and honestly I always feel like a lot of times in agency work your true audience is the client themselves, so it's not a matter of it just being accurate more "XXX person is weird about this, they don't like XXX" etc, but this is a big brand with a lot of weird baggage and stuff. I personally haven't found it to be less work just because I've internalized so much about the client's preferences. I'd imagine with some brands or businesses it would work. though the meta description thing was kind of a dumb way to start. instinctively I felt like oh this is just busywork but then I realized how short 160 characters is and often I ended up writing prompts that were longer than the meta descriptions themselves and then had to edit them quite a bit anyway.

AI is going to make biopics 1000x better. More shockingly, that'll include biopics that were already made *before* AI (from https://t.co/haZKa3wbJ6) pic.twitter.com/c1sO2MdUE0

— Flo Crivello (@Altimor) February 1, 2023

Yesterday i was thinking a good use of this nonsense would be if you turned captions on a Zoom meeting and were able to export them as a text file, you could ask ChatGPT to summarize the meeting in the form of bulleted minutes

map, I posted this during the ilx poll; I enjoyed his rage: https://davidgolumbia.medium.com/chatgpt-should-not-exist-aab0867abace.

many xps, loved this, ty rob

ai generated meeting minutes would leave out the crucial elements of cringe and agony at getting lost in the weeds for 20 minutes, considered a very important part of meetings where i work

theres already transcription software, the bullet points idk would be a trick to pull them out def

If people said "that's an action for me" etc as they often do it could even call out the actions and attach the names to them because Zoom captions (usually) know who's talking

yeah if people are doing manual tagging then you could do it without ai, at that point you could prob just write down the thing tho

this feels like how people discover over and over again that in a lot of ways paper is better than screens

For a list of actions maybe but it's laborious to minute an hour-long meeting and capture everything that's said - but a full transcript is way too much detail and not organised well enough to be useful as a reference

I'm going to try it

meeting minutes are kind of an art imo and it's a big part of the job of a lot of people in government especially. they're also a permanent record where i am. looking through old meeting minutes can be a pleasure, a lot of stuff between the lines.

automating minutes-taking would also mean one couldn't volunteer for the job to get out of participating

lol its like making a totally new weird computer

lots of mysteries to unravel here pic.twitter.com/8XdiJ9QbKC

— Colin Fraser (@colin_fraser) March 13, 2023

You said it wasn't art
So now we gonna rip you apart

http://images.rapgenius.com/1e2b042bbddb8a55f53f726b7d150183.400x400x1.jpg

🤯🤯Well this is something else.

GPT-4 passes basically every exam. And doesn't just pass...
The Bar Exam: 90%
LSAT: 88%
GRE Quantitative: 80%, Verbal: 99%
Every AP, the SAT... pic.twitter.com/zQW3k6uM6Z

— Ethan Mollick (@emollick) March 14, 2023

at first I thought this was pretty impressive, but on second thought what would a human score if they were allowed to use the internet while taking these tests?

& standardized tests tend to grade you on your ability to think like a computer, so computers taking them should have a headstart on humans anyway.

They are as much about your ability to answer standard questions under time constraints as anything, and human time constraints are meaningless to a computer.

I mean, I'm impressed.

It's telling though that OpenAI isn't bragging about how many parameters GPT-4 has compared to GPT-3 (basically, how big/complex the model is.)

One reason they might not brag about that is: they actually DID use a ton more parameters, like 10x as many. But performance didn't improve anywhere close to that much, suggesting that the current approach with LLMs is approaching a plateau.

Of course even if I'm right that plateau is going to be enough to wreak havoc (in both good and terrible ways.)

jimbeaux is otm. iirc a primary reason we built computers was to store information and retrieve it quickly. GPT is an impressive feat of interface design not doomsday for human brains

These technologies are going to make 90% of human labor obsolete. I spent two years writing travel itineraries for rich people—that job could easily be replaced with chat gpt. My current job, teaching, cannot be, because forming relationships with students is just as important as grading essays, which i assume chat gpt can do or will be able to do. But my side hustle, freelance writing/art criticism, could be done by chat gpt. Not “as well,” I flatter myself by saying, but maybe as well.

Releasing this stuff under current capitalist conditions seems like a nightmare. We need a radically new way of understanding the value of human lives in light of this technology. We can no longer say that people unable to generate value are disposable because that is all of us.

lol cmon man

how many people does it take to write travel itineraries for rich people in this economy

a computer program can’t ever do my job cuz I’m a genius at it boom take that computers

a computer program cant do my job i am a computer program

not sure if I mentioned it before but if you want to see something cool that's probably more impactful:
https://www.deepmind.com/research/highlighted-research/alphafold

tl;dr you give this model a protein sequence and it's insanely good (95%+) at telling you what the structure would look like

yeah formal problems like that are clearly the real "ai" application, making ai talk is just pr to juice openais valuation

feel like it's gotta be good for some climate stuff too

don't worry treeship, rich people won't be satisfied exploiting and dominating AIs

Keep it busy with the three body problem.

making ai talk is just pr to juice openais valuation

I dunno! If it replaces Google (or if Google replaces Google with it) that's a pretty impact right there.

At a minimum, it's going to change how we interact with computers quite a bit. How long have we been talking about interacting with computers using natural language, forever right? And GPT is good at ingesting a bunch of documents and answering questions about them, so you can do things like give it your web history and ask it "what was that article I read about a forest with unusually ancient trees" and it'll be able to give you the URL.

Or say if you're a lawyer and your work is sort of mechanical text transformations anyway, having AI draft something for you to review will save you time (although you will have to review it as if written by a very clever stupid person.)

LLMs ain't all hype.

sure search is a nice application for it too but thats pretty different than the generative performances that are getting all the press, doubt chatgpt will replace google tho im sure they have something better in house

also worth questioning if searching with ai will be much different than whats going on now, the problem with the holy grail of natural language interface is that natural language just isnt that precise, you see people interacting with these programs now and theyre not talking like humans theyre figuring out how to talk computer same as always

google already uses very sophisticated programs for search but thats not "ai" because were not pretending the computer is learning or that it somehow resembles a brain its kind of semantics at some point

for simple stuff nah, it's pretty good (if you're asking it to do something in its capabilities and not treating it like an infallible oracle)

for complicated stuff yeah, but for the complicated stuff it's still wayyyyy less work than writing code to do the same thing

xp

literally it has problems finding movie showtimes and doing long division currently

yeah you have to know what it's good at and what it's not

but like i fed it a bunch of my emails and asked it "what were alex and lukas talking about in 1999" and it gave me a really good summary

that's amazing1 and like I didn't have to try multiple prompts to get something to work. it just worked.

a bunch of friends are using it to write code. and it works! not 100% of the time. but overall it's a ridiculous time saver.

code is good because, like law, it's so regular.

well my calculator is really bad at generating infinite episodes of The Big Bang Theory

I cant even get it to spell Bazinga

idk searching an extremely small data set thats already tagged by name and date for two names and a date xp

... and summarizing it really well. without me needing to like, find a good text summarization library and write code. it's just ... one more thing it can do out of the box.

a bunch of friends are using it to write code. and it works! not 100% of the time. but overall it's a ridiculous time saver.

code is good because, like law, it's so regular.

― official representative of Roku's Basketshit in at least one alternate u (lukas), Tuesday, March 14, 2023 5:11 PM (one minute ago) bookmarkflaglink

its a time saver until you have to debug it and you have no idea how the program works