Derek Powazek ๐Ÿ

I keep hearing people say AI generated text "will get betterโ€ as if thatโ€™s a known, accepted fact.

But it's not a known, accepted fact. Massive language model AI text generators have been around for YEARS, and they've always been this bad, this prone to lying, this full of shit. Weโ€™re just getting it shoved in our faces now.

โ€œIt will get betterโ€ is a bet, not a statement of fact, and I bet against it.

May 30, 2024 at 8:40pmWeb
Show threadDerek Powazek ๐ŸMay 30, 2024

Google did not roll this out now because it was good, they rolled it out now because Bing did it first and they panicked. They also fired all the people who had been saying, no, it's not ready yet, even though those people were right then, and they're still right now.

And Bing only rolled it out because they had a search engine nobody gave a shit about and had to try something desperate because they need eyeballs to put ads next to.

Capitalism is why we're all seeing this shit now. Capitalism.

Show threadDerek Powazek ๐ŸMay 30, 2024
What's amazing to me is that I hear tech skeptics repeating the "it'll get better in the future" line and I'm like why would you think that? It hasn't gotten better recently, or even not-that-recently. It's fundamentally the wrong tool in the wrong place for a search engine. I don't think there's any amount of work that can make it better, even by people who give a shit about truth or providing a good user experience, and there's no evidence that Google and Bing give a shit about those things.
Show threadDerek Powazek ๐ŸJun 1, 2024

People ask me, whatโ€™s it like being right all the time?

And I always tell them, itโ€™s awesome, actually.

https://www.nytimes.com/2024/06/01/technology/google-ai-overviews-rollback.html?unlocked_article_code=1.wU0.D3VD.yfT8H0oCQOSM&smid=url-share

Google Rolls Back A.I. Search Feature After Flubs and Flaws

Google appears to have turned off its new A.I. Overviews for a number of searches as it works to minimize errors.

The New York Times
Show thread`Da ElfJun 1, 2024
@fraying No, it isn't.
Show threadDerek Powazek ๐ŸJun 2, 2024
@elfin you do you
Show threadKoraunuiPakehaJun 2, 2024
@fraying And you're right. Obviously
Show threadClaudius LinkMay 31, 2024
@fraying
Wait until it will be trained by AI generated training data /s
Show threadJWcph, Radicalized By DecencyJun 1, 2024
@fraying I've read that natural language processing keeps hitting a ceiling right around 80% accuracy because better than that requires semantic/symbolic comprehension, which not only isn't currently possible; we have no idea how to get there. I feel this might be related to why LLM/AI keeps failing so spectacularly - which might explain why I can't find any articles about it with the search engines run by companies not interested in anyone knowing about it...
Show threadDerek Powazek ๐ŸJun 1, 2024
@jwcph yup. Everything we have now is not intelligence, itโ€™s just a parlor trick.
Show threadBensGreenGamingAcJun 2, 2024
@fraying If you put profit above function then this is what you get.
Show threadwasabi brainJun 1, 2024
@fraying this is maddening because itโ€™s not even like bing gained any market share from doing it. Literally consumers did not give a shit and just stuck with google. No one asked for LLM search
Show threadDerek Powazek ๐ŸJun 1, 2024
@virtualinanity Right? Itโ€™s all just stock market manipulation at this point.
Show threadRolandOfGileadJun 1, 2024
@fraying are you suggesting that North Korea or Russia would do a better job?
Show thread๐Ÿณ๏ธโ€โšง๏ธPepperTheVixen๐ŸฆฏMay 31, 2024
@fraying AI text will get better just like how Braille will be gone in 10 years and crypto will solve all our problems. It won't. Not if it keeps getting its output fed back into it
Show threadcohomology is FUN!Jun 1, 2024
@fraying itโ€™s survivorship bias. We are surrounded by technology that has improved substantially since first being introduced to market. We forget about the inventions that never gained a foothold, often hyped for some short time before fading away as fads.
Show threadDerek Powazek ๐ŸJun 1, 2024
@cohomologyisFUN Thatโ€™s a good point!
Show thread๐‰๐จ๐ง ๐๐ž๐ซ๐ ๐ž๐ซJun 1, 2024
@fraying Try Googling "does using a parachute decrease the risk of injury from jumping out of an airplane?" (Without the quotes.)
Show threadMalcontent Creator Jun 1, 2024

@jonberger
Hahaha yes it looks like itโ€™s parroting the first result, which looks like a joke paper

ยซ However, the trial was only able to enroll participants on small stationary aircraft on the ground, suggesting cautious extrapolation to high altitude jumps ยป

Garbage in, garbage out
@fraying

Show thread๐‰๐จ๐ง ๐๐ž๐ซ๐ ๐ž๐ซJun 1, 2024
@geist @fraying Indeed it is. You recognized that right away, as did I, and as would pretty much any human reader. So it seems safe to say that AI is guaranteed to be unreliable until it develops a sense of humor.
Show threadื—ื ืŸ ื›ื”ืŸ โ€ข Hanan CohenJun 1, 2024

@fraying research says it cannot get much better

https://m.youtube.com/watch?v=dDUC-LqVrPU

Has Generative AI Already Peaked? - Computerphile

YouTube
Show threadCassandrichJun 1, 2024
@fraying We've also known for decades that statistical chat models get worse past a certain point.
Show threadDerek Powazek ๐ŸJun 1, 2024
@dalias Yes! This. Exactly.
Show threadRichard W. Woodley ELBOWS UP ๐Ÿ‡จ๐Ÿ‡ฆ๐ŸŒน๐Ÿšดโ€โ™‚๏ธ๐Ÿ“ท ๐Ÿ—บ๏ธJun 1, 2024
@fraying
"It will get better" seems to assume the crap being fed into it will get better as they keep adding crappier and crappier stuff to their "learning "feed and expect it to get better.
Show threadepicdemiologistJun 1, 2024
@fraying The more AI text is out there, the more of it will be in the training dataset, and the higher the garbage quotient will get.
Show threadDerek Powazek ๐ŸJun 1, 2024
@epicdemiologist yup!
Show threadHelen LHJun 1, 2024
@fraying Did they forget the 1st lesson of computing? Garbage in Garbage out.
Show threadDerek Powazek ๐ŸJun 1, 2024
@Research_FTW they have indeed forgotten that - in fact theyโ€™ve turned it on its head - they think more input is the answer to bad input.
Show thread(moved)Galbinus Caeli ๐ŸŒฏJun 1, 2024
@fraying they won't get better, because the underlying theory is bad. Without an underlying theory of mind and experience of the objective world, speech is just sounds. They can be pretty, like Abba singing phonetically, but there is no inherent meaning.
Show threadDerek Powazek ๐ŸJun 1, 2024
@SkipHuffman yup
Show thread(moved)Galbinus Caeli ๐ŸŒฏJun 1, 2024
@fraying children learn sounds first. Every adult male is "daddy", every animal is a "kitty". Kids have to learn the meaning behind the sounds. LLMs have no way to do that.
Show threadMr. BitternessJun 1, 2024
@fraying
AI technology will surely improve.
But the things AI says... It's hard to imagine a scenario where that actually gets better.
AI trainined on nothing but things humans have said is kinda neat, but often hilariously wrong. ("Hallucinations").
Now imagine AI trained on a mixture of stuff humans have said and stuff that AI has said. (i.e. Today's internet)
Now imagine AI trained on content that is majority AI-written. (The internet of tomorrow)
๐Ÿ˜ฌ
Show threadXenofact (he/him)Jun 1, 2024

@fraying seriously. If your model is flawed at its base or has built-in issues then it doesn't get better. You might build around it, but then you're building a corrective system atop another system.

Someone built some toys and are telling us they're tools.

Show threadCurtiss Grymala (he/him/his)Jun 2, 2024
@fraying if anything, I honestly imagine it getting worse. They're creating a feedback loop on themselves, consuming AI-generated content, which is most likely going to make things worse over time.
Show threadDerek Powazek ๐ŸJun 2, 2024
@cgrymala Most things bound to capitalism do.
Show threadMichelle MurphyJun 2, 2024
@fraying
Has the internet gotten better?
Show threadDerek Powazek ๐ŸJun 2, 2024
@me Bigger and faster, sure, but better? Not really.