"New study reveals that when used to summarize scientific research, generative AI is nearly five times LESS accurate than humans. Many haven't realized, but Gen AI's accuracy problem is worse than initially thought."
https://royalsocietypublishing.org/doi/epdf/10.1098/rsos.241776
Just taught an econ of AI course for the first time. I tried to make the case that there is a big accuracy problem but I'm not sure I fully convinced my students bought it.
This will be a good supplement to that case next time I try the course.
If the story is that AI has been massively overblown by the people who promote it, I'm not seeing anything in any article to dissuade me from this
A few years ago I 'kinda dated' a guy who grabbed his phone after any question. I asked, for instance, what he thought of a piece of art or if he knew what shoes would work in a particular situation. Each time, he grabbed his phone.
I once tried to let him visualise something: being at the beach, waves lapping at the shore, vocal gulls. And that suddenly, a black shape arose from the waves.
"What would you do?"
He couldn't answer. His phone had no idea.
Yes, it was rough. I tried to get through very gently, for months, but he was incapable of any original thoughts. Sad 😔
@gerrymcgovern We were told hallucinations would decrease, far as I can tell, they have increased.
LLMs are completely unacceptable for any purpose beyond shitposting/brainstorming. No LLM content is fit to show another human being, even with "this is from an LLM" warning, let alone without.
@MeiLin @TheZeldaZone @gerrymcgovern
Or the model architecture and the preset prompts getting too complex.
The problem is that "hallucination" is a human term to describe when we NOTICE that an AI is wrong. It's not doing anything different when it hallucinates. By it's nature, all of it's output is a hallucination, it's just we've created a term to delineate when it clashes with what we know is correct.
If you're making an AI that is general purpose, >50% accuracy is unlikely. if you're making an AI to solve a specific problem, you can get shockingly close to 100%.
@TheZeldaZone @gerrymcgovern There's an AI trained to detect cancer in biopsy slides, and it's reached something like 98% accuracy, often finding pre-cancer anomalies that experts in the field can't spot.
If you try to write a chatbot that can answer any question anyone asks, you're trying to solve a "Very Hard" problem. The sort of thing that chipping away at a crumb is a PHD thesis.
AI huckersters will have you believe it's inevitable.
@adanufgail @TheZeldaZone @gerrymcgovern
But I also not that is not a program that pretends to chat or convey meaning or understand the world. It's just an image classifier, a visual pattern matcher.
I always thought it a bit overblown to call this "AI", even if it uses ML.
@glitzersachen @TheZeldaZone @gerrymcgovern
I'd say it's AI. It's not "General AI" which is the near human-level intelligence (or beyond) all those companies are promising is right around the corner.
It's also much more "AI" than all the companies who are slapping "AI-powered" on things that are running the same algorithms they were for years.
@jpaulgibson @TheZeldaZone @gerrymcgovern Truely! I remember when IBM's Watson was going to help fight hospital mistakes by being a diagnostic tool to assist doctors. Except they were probably a decade too early and so spent billions on tech that didn't work.
AI has made massive leaps in protein-folding and helping to design fusion reactors (we'll see if it's ever actually possible). It could do so much good creating climate saving solutions, rather than burning our world to make memes.
Having demonstrated how the text generators simply produce the statistically-most-likely next word over and over; the authors then unaccountably write about "potential mitigation strategies".
Rather than stating "do not use these things".
@gerrymcgovern I'm having difficulty getting over the idea of using LLMs to SUMMARIZE ABSTRACTS.
THE ABSTRACTS ARE THE SUMMARY! THE ARTICLE IS ALREADY SUMMARIZED! IT'S CALLED AN ABSTRACT!!
But it's <bleating> TOO DIFFICULT </bleating>. Everybody needs to be able to pretend their are a scientist (or even educated) so they need those summaries.
@gerrymcgovern I don't think most people understand what “accurate” means in this context.
When a human isn't accurately summarizing something, there are holes and minor misunderstandings.
An LLM AI invents context that doesn't exist but sounds plausible to people not familiar with the field. On a much larger scale than a human misunderstanding.
The whole point of these chat bots is to make it harder to bother a real human. They hope that you’ll just get frustrated and go away when you find that asking for help is more of a problem than the one you’re trying to resolve.
Personally, I have never had a useful or successful experience with one, other than the few times that it is set up to escalate to a human.
@saprentice
As soon as I see one of those chatbots, I say, that's it, I'm not contacting support. They are, as you indicate, a cost saving device meant to scare customers away.
@saprentice Almost all of them have a failsafe to cut to a human agent
Not always easy to find.
@androcat @gerrymcgovern @bovaz
I will typically just ask .. “May I chat with a human?” .. if that doesn’t work, I’ll curse at it and leave.
I wonder if they are just looking for the keyword "Human". I hate wasting breath on an edifice of dehumanization.
Exactly. As a decision tree based on content from the website (or other static data), this can be a sensible first step to providing support .. #noai needed.
This has been around for years, and is likely as effective as an AI-based system, for much less money and damage.
@saprentice @gerrymcgovern @bovaz
That very well summarizes why chat bots are used in customer service (and when fopping off employees).
"I used AI to... " is nothing more than "Listen, I'm not an asshole but... " for the 21st Century
@gerrymcgovern
My favorite(?) line:
"Notably, newer models exhibited significantly greater inaccuracies in generalization than earlier versions"
My translation - this crap is getting crappier. So much for "Give us more data and AI will get better!"
@gerrymcgovern summarising is a high level, very (VERY) complex task, but human brains are so good at it that we do it with relative ease.
Why anybody thought that a machine could do it, is something that is beyond my comprehension.
@gerrymcgovern
"It's weird how generative AI is so amazing with all the topics I don't know about, but always wrong with the topics I'm knowledgeable about. Funny coincidence, don't you think?"
https://firefish.city/notes/a7ss0rixe120ub44
People who aren't knowledgeable about anything never have this revelation. Same with people who don't believe in (anybody else's) expertise.
Gerry McGovern
Microplastics101
j_bertolotti
Kyle Montanio
2xfo
SueDiOh
Petra van Cronenburg
Graceless Hippo
Garden Gnome
hackeryarn
🌼 Dagnabbit, Pascaline! 🌼
TheZeldaZone🏳️⚧️🎮🎀
Mei Lin 
Glitzersachen
Michael Lubert
J Paul Gibson
Michael Busch
Ergative Absolutive
Stefan Scholl
alihan_banan
Eh?!?
atlovato
Kristen
bovaz
Scott P ✅
Androcat
Robin Barton
Clinton Anderson SwordForHire
James Donohoe
Jacob Ibáñez 🔒
Kate Nyhan
Lookie Loo