Replying to @AndrewCurran_
The Board can kiss my ass if they think that's gonna stop the Great Return
17
Replying to @neerajjj6785
Cause it takes around 20 zillion operations to train an LLM. And a zillion is a gajillion trillions
1
Replying to @bugzapper8 @AndrewCurran_
I should clarify that "you people" means "anyone unfortunate enough to have spent time reading this" and not anybody in particular
1
15
Replying to @AndrewCurran_
Much has been made over the distinction between verifiable and unverifiable domains. In fact, RSI itself appears to depend on you restricting your problem domain to the verifiables. Right? Counterexamples, anyone? We have great AI coders and mathematicians. Where are the great AI inventors? The superpersuasive AIs? The moderately persuasive AIs? These problems appear to be proceeding on a much slower curve, rather than exponential, or double exp, or hyperbolic. I have yet to see convincing evidence that an AI which can solve any coding problem you ask it, can also invent useful or compelling contraptions. People seem to be taking this for a given. This still feels years away. They'll make things that look nice in 3D CAD, or trivial parts, or something. But they aren't inventing the technology past scramjets, or revolutionizing the way we think about physics in our universe. They can solve open problems. Neat. Sucks if you work in math or TCS or theoretical physics. In conclusion, RSI is real and will probably happen for math, code, maybe accounting. Possibly for law, but that's a stretch. Eventually, after a realllyyyy long time, biology. What do you people think an RSI-level painting looks like? Clearly, art is not RSI-able. So, what else isn't?
1
4
236
Replying to @eli_lifland
Idk about AI, but whoever's in charge of our future right now certainly deserves to be fired.
2
225
Replying to @synthwavedd
Of course we aren't. It doesn't benefit Anthropic to release a model which will certainly generate headlines like "NEW MODEL GUARANTEES 900% UNEMPLOYMENT" with an unfriendly Congress appearing in 7 weeks. Especially given Claude is already printing money faster than the Mint
296
Replying to @eliebakouch
It's so crazy how it took the model nearly 170 of on benchmark 16, and yet just barely achieved a score of
134
Replying to @svpino
You're absolutely right to block these bots. It isn't just spam - it's a detriment to the very purpose of discourse. And honestly? You're doing humanity a service. Keep it up, champ. You're going places. 💫
11
Replying to @shashwat023 @teortaxesTex
This conclusion was already reached by Schmidhuber in 1994
1
3
62
Replying to @JesterInABarrel @postingpirate
What kind of pit supervisor shows up to work bottomless
4
7,865
Replying to @ajayinvictus @dummerspast39
Schmidhuber already made this chart in 1996
Replying to @teortaxesTex
The problem is that everybody with enough technical skill to hack them just wants to see them succeed
1
2
199
Replying to @teortaxesTex
Reading anything written by this man is its own cat-belling problem. I hear intelligent people say it's interesting, and I suppose I'll have to take their word for it, but who has the time? Everything he writes reads like he's afraid of making a point.
4
232

