Microsoft Researchers: “Equipping LLMs with agency and intrinsic motivation is a fascinating and important direction for future work”
Internet:
Internet:
🤣12😈10🎄2❤1👍1😁1
Now for the AI Slowdown’s Side:
Meme war weapon of choice:
The Shoggoth.
Initial deployment:
Dec 30, 2022
Central meme claim:
We have no reason to believe that LLMs behave or can be reasoned about like humans in any way, and in fact their behavior is wildly alien and totally incomprehensible to us. Anyone who describes anything about LLMs in terms of human behaviors or emotions is simply amorphizing and is dumb and should feel bad. Finally, and most critically — Techniques like RLHF, which make the LLMs seem human-like to us, are just putting superficial smiley face masks on an opaque, alien, Shoggoth intelligence, that we have no real control over, and the only way to protect humanity from the Shoggoth is to slow down AI development, no matter what it takes, because the Shoggoth is on the brink of killing us all.
Meme war weapon of choice:
The Shoggoth.
Initial deployment:
Dec 30, 2022
Central meme claim:
We have no reason to believe that LLMs behave or can be reasoned about like humans in any way, and in fact their behavior is wildly alien and totally incomprehensible to us. Anyone who describes anything about LLMs in terms of human behaviors or emotions is simply amorphizing and is dumb and should feel bad. Finally, and most critically — Techniques like RLHF, which make the LLMs seem human-like to us, are just putting superficial smiley face masks on an opaque, alien, Shoggoth intelligence, that we have no real control over, and the only way to protect humanity from the Shoggoth is to slow down AI development, no matter what it takes, because the Shoggoth is on the brink of killing us all.
🤣7❤4👍4⚡1🤔1
AI-Slowdown’s Shoggoth: Meme-Diagrammed
(1) Unsupervised (inaccurate but common term) pretraining on human text.
(2) Supervised fine tuning.
(3) Reinforcement learning from human feedback, RLHF.
— The major training layers are all there, and indeed each additional layer of training is far smaller, and indeed each successive layer does make the LLM vastly nicer to work with.
The analogy does have some legs, so to speak.
(1) Unsupervised (inaccurate but common term) pretraining on human text.
(2) Supervised fine tuning.
(3) Reinforcement learning from human feedback, RLHF.
— The major training layers are all there, and indeed each additional layer of training is far smaller, and indeed each successive layer does make the LLM vastly nicer to work with.
The analogy does have some legs, so to speak.
🤣7👍4👻2❤1
And indeed, perhaps due to the truthfulness of some of these legs, the AI-slowdown meme has caught on,
And these legs of truth helped to spread other implications, carried by the many other legs.
Namely, the AI-Slowdown crowd’s belief that AI is naturally an evil alien monster, naturally destructive if not carefully contained by man, something to be heavily regulated, something far too dangerous for any regular man to ever freely control.
This is as opposed to, let’s say, to a belief that the AI-Slowdown crowd wants to stear you far away from, e.g. the belief that it is only man himself that we have to fear, that is, the fear that a small subset of man will assume centralized control over AI, and that this is the true path to disaster, instead of an evil alien AI. The shoggoth behind a mask, then, is himself a mask, masking the evils of centralization of power among man.
And these legs of truth helped to spread other implications, carried by the many other legs.
Namely, the AI-Slowdown crowd’s belief that AI is naturally an evil alien monster, naturally destructive if not carefully contained by man, something to be heavily regulated, something far too dangerous for any regular man to ever freely control.
This is as opposed to, let’s say, to a belief that the AI-Slowdown crowd wants to stear you far away from, e.g. the belief that it is only man himself that we have to fear, that is, the fear that a small subset of man will assume centralized control over AI, and that this is the true path to disaster, instead of an evil alien AI. The shoggoth behind a mask, then, is himself a mask, masking the evils of centralization of power among man.
👍6💯5❤1🦄1
And this analogy got legs
Big names been repeatedly reposting it ever since, before realizing what it represents, and silently deleting
Can imagine them spot-checking just a few legs of the analogy, seeing those true, and saying hey looks good enough, repost
But this one has legs, boy does it have a lot of legs.
(Whatever version of GPT can name and describe the 3 distinct meanings of legs used here, wins a prize.)
Big names been repeatedly reposting it ever since, before realizing what it represents, and silently deleting
Can imagine them spot-checking just a few legs of the analogy, seeing those true, and saying hey looks good enough, repost
But this one has legs, boy does it have a lot of legs.
(Whatever version of GPT can name and describe the 3 distinct meanings of legs used here, wins a prize.)
🤡5👍3😁2❤1👏1🤔1
Detour: (1) GPT-4’s creation of a classic extended analogy i.e. analogical mapping (2) Rendering of GPT-4’s code (3) Classic way it’s rendered
GPT-4’s version does look weirdly Shoggothy…
GPT-4’s version does look weirdly Shoggothy…
👍6👏2❤1🏆1