Please tell me @ifbookspod.bsky.social is going to cover his new release, because I got dragged to a Futurism conference where they kept hyping it and Iโm afraid my coworkers are going to be asked to read itโฆ
We haven't even mapped out a tenth of our brain and what it does, developing an AI for that isn't just jumping the gun, it's jumping an entire war's worth of weaponry.
"Stupid" is a massive understatement, sir. If there is a word to describe how dumb this really is, please use it now.
I am literally fighting the industry version of this right now (letโs give a gen AI model our existing customer interviews so it can hallucinate fake customers to do product research with)
the brain rot is so fucking pervasive
Am I reading correctly that it reliably predicted the sign correctly but not the magnitude? Ah well, why worry about a few decimal places in scientific research.
You know all those memes about techbros accidentally inventing the bus?
Pretend hypotheses tested on pretend people? This guy just accidentally invented the discipline of economics.
I like that they explicitly *don't* include "counter-intuitive" results, so at best, the finding is that the LLM can confirm your prior beliefs. And like, yeah, that's well-known, and it's a big problem with them, not a strength!
Sounds like a great way to manufacture consent for the bad ideas you have: "no choice but to fire on the protestors/cancel Social Security/nuke Lebanon", the AI told us to do it
So yes, there is sometimes a valid use case for synthetic data in quantitative research, but you have to generate the data somehow. With a LLM you feed it a prompt. And if that prompt is the hypothesis, then you have data leakage. It's like putting leading questions into a survey. ๐
One thing thatโs cool about these kinds of papers is that it removes the illusion of intelligence from folks like this. You may have suspected before that someone like this was full of shit and canโt possibly be smart; but now you KNOW it.
jesus christ, LLMs are not "testing [a hypothesis] with simulated AI humans"! they're just stringing together statistically likely words based on a big database! that's literally all they can do! are these assholes just supernaturally shameless, or genuinely the most gullible humans ever to live?
We don't need no stinking humans. Just words and dreams. And just think of the savings! No grants needed! No one to hire to do any tasks! Just a bunch of word clouds colliding to form new and better clouds.
i love the part where he says heโs super excited about how the LLM predicts the simulations wrong, but then if you make the simulation outcomes part of the dataset, it predicts them correctly
yes? no shit? thatโs literally what an LLM is?
There are dear friends whoโve gone to Wharton, and there are people I wouldnโt piss on if they were on fire whoโve gone to Wharton.
That said, if there were any field of academia that could be replaced with AI it would surely be Business School.
AI: Let's see if the earth is flat (hypothesis)
AI: Er, um... (null hypothesis)
AI: Is the earth flat? (investigation)
AI: Er, maybe, it all depends (results)
AI: Earth is probably flat (conclusion)
we've come such a long way from "the lurkers support me in e-mail" to "the simulated AI human agents agree with my conclusions, at least when I remember to add those conclusions to the prompt that generated them"
He deleted it, with a whiny followup. The dude is a professor at Wharton studying AI with >200k followers and he "didnโt expect the tweet to travel so widely."
That's like training a million ants to climb up on a wall and form themselves into the image of a doorway, then trying to walk through it.
Or not. I have to check which strain I'm smoking brb.
This works ok for drug/molecule discovery because it gives you potentially interesting ideas that you can then go test in actual physical reality.
This is a cargo cult application of that approach.
Psycho-social experiments : "yes it's interesting, but it's way too oversimplified and can't explain everything"
Game theory: "Well same thing... not inherently bad, but a lab isn't a mini society"
AI generated experiments :โฏ"Ah, come on now !"
Maybe we can use a more primitive approach. I make a hypothesis, design an experiment and graph what I think the data will be if I perform it... Oh wait no that's not science it's data fabrication
'AI' academics to face the reality that most people in academia would not in fact enjoy having to sift through nonsense-filled procedurally-generated papers as part of their work.
Soโฆ heโs suggesting that we just make up data?
Because thatโs what this is.
Using AI to generate hypotheses is fine. I could even see using an AI chat bot to test them. Itโs probably better than a surveyโฆ. but to also make AI the participant?
What the actual fuck?
I just woke up so I might be wrong here, but did some asshole just make AI that works like fucking Facebook posts where it just turns into a confirmation bias circle jerk regardless of how right it actually is?
Can I get the future where the computers clean up the floors and toilets and things like that and humans do the creating things and inventing?
I feel like this is AI going in the wrong direction
โIn silicoโ
๐๐๐๐๐
I donโt think thereโs directly translatable a Latin word for โsilicon.โ โSandโ would be โharenaeโ. I meant, if we were to approach this seriously, which we should not do.
