Thank you for this important contribution to Type of Guy Studies! The Cowen connection is super interesting-- I've got a piece I haven't finished yet on the relationship between right-wing economics and AI, and think this will be a useful example. Cowen comes out of a right-wing economic lineage that is very susceptible to AI hype because of how it sees brains and the world. AI, as a black box with its throw-it-all-in-the-cauldron training, is easy for them to see as representing a kind of market-based "spontaneous order" (very Hayek) over deliberate "planning" of traditional computing. In this view it makes sense that AI would lead to superhuman AGI, which would naturally prove the superiority of market competition over deliberate lefty cooperation. The red-baiting is a nice touch here, and really brings some of these ideas full circle. (I think all of this also explains the shift away from EA you mention here... but I should really finish the piece instead of writing it in someone else's comment section)
Will be interesting to see how they react when AGI doesn't pan out...
Great comment, and I think you're very right about the Hayekian attraction of A.I. for some of these guys. (It's interesting to me too because just a few years ago Thiel was making some grand claims about the blockchain being capitalistic and American, while A.I. is communistic and Chinese.)
Love this: "a cloying, manic self-assurance that somehow scans as “genius” to the credulous and the powerful and as 'extremely annoying bullshit' to literally anyone else." I taught and did research on neural networks right around when "deep learning" took off, and listening to consultants--let alone these "personalities"--hype AI at my company makes me want to lose my fucking mind. I can't even get anyone to look at a regression line and we're somehow going to solve all our problems with AI.
Your article is extremely interesting and touches on many issues, some explicitly, some more implicitly. I'll have to read it again at home in order to write a proper comment. But it's really a good work!
> none of the words, none of the axes, none of the lines--bears even an indirect relationship to an observable, measurable, or quantifiable phenomenon.
One of the axes is time - years from 2018 to 2028. Sure seems like a quantifiable phenomenon to me, which points to your tendency towards hyperbolic writing.
I’m a black woman in my mid-50s who uses generative AI and the whole nine yards @ the AI hype. I’m not your “typical” AI-er. Is it me - or are these guys all of a type? I know I’m middle-aged and cranky but I don’t think I’m wrong. it’s Zuckerberg 2.0. Borrrinngg 🥱
Man. I did not need to know about Leopold Assburger and his dum dum delusions. Him and Ackman have the exact same fundamental personality flaw: not wedgied enough.
A post criticizing a person who was fired for leaking confidential information (and apparently an EA or cultist) and the racist Bill Ackmann. I subscribe.
Besides using faster chips and buying more of them, people are improving the recipe for training models with better architectures and algorithms. Someone trained a 2018 state of the art model on their laptop last year!
When looking at older models, counting the number of operations (flops, basically adding or multiplying two numbers) used during training isn't an apples to apples comparison against newer models because of the improved algorithms. Effective compute tries to measure/standardize these improvements.
conceptually "effective compute" is the component of that graph that has the clearest and most direct relationship to measurable aspects of the real world, I'll grant, but its true measurability depends on the formula for "effective compute" you're using, and, worse, it's being used in a criminal way here—to backfill a nominal metric against which a pre-supposed exponential curve can be fit. I know this will make me sound like a crank, but I tend to think that by "bundling" algorithmic progress and physical scaling "effective compute" as a "metric" is basically designed for this purpose
Yeah there's definitely some fuzziness, particularly if you try include pre-transformer language models or try to compare across huge model scales. I don't think too much is being smuggled in though — the y-axis brings in way more assumptions. Typically scaling law plots[1] show flops v. loss (how well the model predicts the next word during training). It is very possible that we'll continue to make loss go down and that won't correspond much in terms of useful abilities or anything like an "automated engineer."
Two other plausible possibilities for AI fizzling out:
- The x-axis is log scaled; at some point we'll be unable to pour order of magnitude more chips/electricity/data into the models.
- At large enough sizes, transformers stop getting much better (just like LSTMs before them). Basically the straight lines on the scaling law plots start to bend.
Thank you for this important contribution to Type of Guy Studies! The Cowen connection is super interesting-- I've got a piece I haven't finished yet on the relationship between right-wing economics and AI, and think this will be a useful example. Cowen comes out of a right-wing economic lineage that is very susceptible to AI hype because of how it sees brains and the world. AI, as a black box with its throw-it-all-in-the-cauldron training, is easy for them to see as representing a kind of market-based "spontaneous order" (very Hayek) over deliberate "planning" of traditional computing. In this view it makes sense that AI would lead to superhuman AGI, which would naturally prove the superiority of market competition over deliberate lefty cooperation. The red-baiting is a nice touch here, and really brings some of these ideas full circle. (I think all of this also explains the shift away from EA you mention here... but I should really finish the piece instead of writing it in someone else's comment section)
Will be interesting to see how they react when AGI doesn't pan out...
Great comment, and I think you're very right about the Hayekian attraction of A.I. for some of these guys. (It's interesting to me too because just a few years ago Thiel was making some grand claims about the blockchain being capitalistic and American, while A.I. is communistic and Chinese.)
Sounds like a very interesting piece, will gladly check it out once published.
Love this: "a cloying, manic self-assurance that somehow scans as “genius” to the credulous and the powerful and as 'extremely annoying bullshit' to literally anyone else." I taught and did research on neural networks right around when "deep learning" took off, and listening to consultants--let alone these "personalities"--hype AI at my company makes me want to lose my fucking mind. I can't even get anyone to look at a regression line and we're somehow going to solve all our problems with AI.
I read this stack to experience what being a smart person might feel like.
Good stuff!
Your article is extremely interesting and touches on many issues, some explicitly, some more implicitly. I'll have to read it again at home in order to write a proper comment. But it's really a good work!
Oh, I wasn't aware Leopold was an economist. I feel considerably better already. Never mind.
When you're "off-the-charts brilliant" you need some charts that don't make sense
It's almost quaint to read about Mitt Romney's lies considering what the Republican party has become.
The amount of attention this guy is getting is enraging. It gives him way more legitimacy than he deserves. Strange times we live in.
> none of the words, none of the axes, none of the lines--bears even an indirect relationship to an observable, measurable, or quantifiable phenomenon.
One of the axes is time - years from 2018 to 2028. Sure seems like a quantifiable phenomenon to me, which points to your tendency towards hyperbolic writing.
You got me...
I’m a black woman in my mid-50s who uses generative AI and the whole nine yards @ the AI hype. I’m not your “typical” AI-er. Is it me - or are these guys all of a type? I know I’m middle-aged and cranky but I don’t think I’m wrong. it’s Zuckerberg 2.0. Borrrinngg 🥱
Man. I did not need to know about Leopold Assburger and his dum dum delusions. Him and Ackman have the exact same fundamental personality flaw: not wedgied enough.
A post criticizing a person who was fired for leaking confidential information (and apparently an EA or cultist) and the racist Bill Ackmann. I subscribe.
“ This is what a highly intelligent, highly knowledgeable person actually believes after much thought.”
— another take on the same 165-page essay:
https://thezvi.wordpress.com/2024/06/07/quotes-from-leopold-aschenbrenners-situational-awareness-paper/
What to believe? I feel so inadequate.
right? where do we start in our level of inadequacy?😉 like I said, I’m 54 and cranky….
> What is "effective compute"?
Besides using faster chips and buying more of them, people are improving the recipe for training models with better architectures and algorithms. Someone trained a 2018 state of the art model on their laptop last year!
https://github.com/samvher/bert-for-laptops/blob/main/BERT_for_laptops.ipynb
When looking at older models, counting the number of operations (flops, basically adding or multiplying two numbers) used during training isn't an apples to apples comparison against newer models because of the improved algorithms. Effective compute tries to measure/standardize these improvements.
https://epochai.org/blog/algorithmic-progress-in-language-models
conceptually "effective compute" is the component of that graph that has the clearest and most direct relationship to measurable aspects of the real world, I'll grant, but its true measurability depends on the formula for "effective compute" you're using, and, worse, it's being used in a criminal way here—to backfill a nominal metric against which a pre-supposed exponential curve can be fit. I know this will make me sound like a crank, but I tend to think that by "bundling" algorithmic progress and physical scaling "effective compute" as a "metric" is basically designed for this purpose
Yeah there's definitely some fuzziness, particularly if you try include pre-transformer language models or try to compare across huge model scales. I don't think too much is being smuggled in though — the y-axis brings in way more assumptions. Typically scaling law plots[1] show flops v. loss (how well the model predicts the next word during training). It is very possible that we'll continue to make loss go down and that won't correspond much in terms of useful abilities or anything like an "automated engineer."
Two other plausible possibilities for AI fizzling out:
- The x-axis is log scaled; at some point we'll be unable to pour order of magnitude more chips/electricity/data into the models.
- At large enough sizes, transformers stop getting much better (just like LSTMs before them). Basically the straight lines on the scaling law plots start to bend.
[1] https://arxiv.org/abs/2001.08361