I'm halfway through this article & like (and agree with) the points made thus far. But I want to have an aside on how people uncritically use the term "foundation models" & "reasoning," when it is very likely that the models memorized all these benchmarks. 🧵 prospect.org/power/2025-0... - ThreadSky

timnitgebru.bsky.social • 3 days ago

I'm halfway through this article & like (and agree with) the points made thus far.

But I want to have an aside on how people uncritically use the term "foundation models" & "reasoning," when it is very likely that the models memorized all these benchmarks. 🧵

https://prospect.org/power/2025-03-25-bubble-trouble-ai-threat/

Comments

digelsrud.bsky.social•2 days ago

📌

markusbuhmann.bsky.social•2 days ago

GenAI platforms are pump and dump scams.

toroidal.bsky.social•2 days ago

There’s panic setting in. The hype bombardment is increasing, leading to this whacky stuff. I feel like I’m taking a daily regiment of crazy pills.

edzitron.com•3 days ago

While I agree - what do refer to these projects as instead?

timnitgebru.bsky.social•2 days ago

We had terms like base models, pre-trained models, large language models etc which are specific to the type of model discussed & didn't signal "foundation" or other hype. That term was proposed by Stanford which also created a center by that name capitalizing on our firing after writing about LLMs.

timnitgebru.bsky.social•2 days ago

Now "foundation" isn't even enough and they say "frontier models" and people just adopt that term.

timnitgebru.bsky.social•2 days ago

I just saw that this term is also used in this article smh.

wat3rm370n.bsky.social•2 days ago

then whatever you do, don't watch this Maureen Dowd interview at about 5 minutes 30 seconds into this clip unless you actually want to feel like your head is exploding 😬

https://youtu.be/fjaUB6-m2rE?si=bYjwZZ1vWaQEtyGA

wat3rm370n.bsky.social•2 days ago

new species... 🫠🫠🫠🫠🫠🫠🫠🫠

ahumbleman.bsky.social•1 day ago

📌

nhegday.bsky.social•2 days ago

You make a good point - nonetheless it’s an insightful piece. Coming to the bubble part of it, this pretty much sums it up …”OpenAI is deeply dependent on hypothetical breakthroughs from future models that unlock more capabilities to boost its subscription price and grow its user base.”

timnitgebru.bsky.social•2 days ago

Yes agreed.

hagenblix.bsky.social•2 days ago

The political problem with this argument is that it focuses on the investors - they're the ones who are being over-sold on hype, and the ones who won't be able to recoup their investment. I think that may very well be true, but the technology will stay and make things shittier anyway.

hagenblix.bsky.social•2 days ago

Focusing on the 'bubble' aspect overprioritizes a narrow set of technical claims (yes, the models can't do what the public is told they can do), but it underestimates the effects on class power that they can have even if they don't work all that well (which will exist even if OpenAI goes bankrupt

hagenblix.bsky.social•2 days ago

and we all are thrown into another 2008 style crisis - in fact, a crisis will likely exacerbate the need for low quality super cheap services, and create a doom spiral of low wages, and bad products). Call it the IKEA mode: it's not about making things better, its about making stuff a lot cheaper.

hagenblix.bsky.social•2 days ago

And there's no way that these models won't make for shit AI legal clerks, shit AI therapists, shit AI adjuncts, and so on - they already are, and they already are much cheaper. So, even if the bubble pops, even if the market corrects - it won't save us, the shit will still come for us.

hagenblix.bsky.social•2 days ago

So I think as a political task, we should prioritize talking about the effects that AI will have on good jobs, and on working class people - the bubble talk can be part of that, but focusing exclusively is dangerous. We can't ask the VCs to save us, we'll have to do it ourselves

hagenblix.bsky.social•2 days ago

I'm not saying the hype/bubble claims are wrong. But as a political focus, they come with severe limitations that I think are really dangerous. They underestimate the resonances of shit AI and fascism. And at worst, the invite complacency - "oh, it'll all blow over, don't worry".

timnitgebru.bsky.social•2 days ago

I wonder how long the making things cheaper part will continue to be true, because investors are subsidizing the astronomical costs thinking that magic will come any day now. What happens when magic doesn't come any day now?

hagenblix.bsky.social•2 days ago

Fair, but I think those costs are already sunk. So, even we were to find ourselves in the most radical scenario - OpenAI goes bankrupt, sells for the symbolic dollar, stops investing: most of their models, at worst the charge for inference would double or triple, no?

victorhg.com•2 days ago

📌

sorocchungus.bsky.social•2 days ago

Pop the bubble

timnitgebru.bsky.social•3 days ago

It truly is like the emperor's clothes. Everyone seems to be in on it & you're the crazy one being like but HE HAS NO CLOTHES.

- There is no difference between the likes of Stanford & any of these companies, they're one & the same. So schools like it make 💰 from the hype & will perpetuate it.

timnitgebru.bsky.social•3 days ago

- The McKinseys and other huge consulting orgs are raking in bank on the hype, akin to all the people who made money during the gold rush, except for the people looking for gold.

All the things people call "laws" aren't laws and were never "laws".

timnitgebru.bsky.social•3 days ago

-Scaling laws? Some people looked at some plots and came up with that.

-"Emergence"? Take a look at this paper showing how that is nothing but hot air: https://proceedings.neurips.cc/paper_files/paper/2023/hash/adc98a266f45005c403b8311ca7e8bd7-Abstract-Conference.html

timnitgebru.bsky.social•3 days ago

- "Reasoning"? Lets set aside how they don't even have a definition for this. But literally change some minor thing on the benchmarks like a number, and you see how these models completely fail.

https://arxiv.org/pdf/2410.05229

timnitgebru.bsky.social•3 days ago

-"Understanding"? Just watch this debate to see the rigor with which Emily discusses the topic vs those who make these wild claims: https://www.youtube.com/watch?v=YtIQVaSS5Pg&t=1s.

If you come up with a new benchmark they'll just guzzle it as part of the training data and then claim to do "reasoning" on that.

ambivalentricky.bsky.social•2 days ago

is this basically the same strategy Volkswagen got busted for using in their turbo diesel engines, where they were programmed to detect emissions tests?

timnitgebru.bsky.social•3 days ago

It is so mind-boggling to me that people have to even spend time debunking these claims. Such a waste of resources that could be going to doing actual science and engineering work.

metteurenscene.bsky.social•3 days ago

📌

Comments

Posting Rules

Reply