Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

There is no hard takeoff

geohot.github.io

Post reply on HN
Print view View original

131–140 of 190 posts

Re: There is no hard takeoff

#131

Post by brandall10 »

holmesworcester wrote: post #33

> The problem is your model needs to include all the computers playing the market, and it also needs to include the other hedge fund bros themselves. The error here is equating hard takeoff with such granular and expensive world prediction. Human intelligence finds more efficient compression than this. When Donald Trump decides what provocative thing to say in a speech, he isn't granularly modeling the human intellig…

The Trump analogy doesn't quite work unless we include a bunch of other actors in the field that have his societal cache and approach. He slaughters other republican nominees because there simply is no competition, no matter how much some try to out-Trump him.

I've seen discussion about RenTech being so dominant because they are indeed observing the competition's trades in a more sophisticated manner and are able to more quickly adjust strategy.

Re: There is no hard takeoff

#132

Post by rafaelero »

WithinReason wrote: post #2

The main issue with this argument that current AI models are extremely inefficient. Models evaluate all weights on every pass, LLMs recall their entirety of knowledge just to output a part of a word and then do it all again. Often a hand crafted algorithm can achieve what a neural network can in a small fraction of compute. There is likely a 1e4x-1e6x compute gap that can be achieved with the right algorithm, maybe e…

> Models evaluate all weights on every pass

Not true for GPT-4.

Re: There is no hard takeoff

#133

Post by isodnhionio »

zenkat wrote: post #71

"And yet, here we are in 2023 and self driving cars still don’t work." LOL. Someone should tell the California CPUC that: https://www.theverge.com/2023/8/10/23827790/waymo-cruise-cpu...

This is yet another reason I want Musk in prison.

Getting drivers off the road will save lives, and not a small number. Most companies investing in automated driving have built a system that is safe and largely effective. These systems have worked safely in the places they've been deployed. We should be heavily investing in expanding their capabilities.

Yet Tesla's system is the most famous and most widely-deployed and basically doesn't work at all. They have made themselves the face of the technology, and that face is uuuuugly. When people think of self-driving cars, they don't think of Cruze or Waymo. They think of a techbro giggling at his Tesla making fart sounds while it barrels into someone's kitchen.

Tesla's marketing is correct in that self-driving cars will have huge economic benefits and save lives. Their grift is harming the reputation of even legitimate companies. The best way I know to speed up self-driving car development is to shut Tesla down.

Re: There is no hard takeoff

#134

Post by epups »

I don't agree with the premise that a 51% compute threshold is transformative, but it's scary to consider that collusion among a small number of big tech companies (Google, Amazon, Microsoft, Nvidia, AMD, Apple) would easily pass that threshold when it comes to GPU compute. More so when you consider they are also two steps ahead of everyone else when it comes to developing improvements.

Re: There is no hard takeoff

#135

Post by reducesuffering »

AbrahamParangi wrote: post #114

Wild to me that people will take seriously people like Eliezer Yudkowsky and Nick Bostrom who are both "idea guys" who have accomplished exactly nothing in the field of AI. Their strength is in the world of discourse, not in the material world. Meanwhile, George Hotz, an engineer and a technologist - a guy who actually works with and understands technology, but isn't an acclaimed harry potter fanfiction writer or a s…

Why not wait until their debate on Aug. 15th? It's going to be a disaster for Geohot as he realizes Eliezer has been "mind-mazing" this problem for 20 years, and comes to realize the barrel of the gun we're all facing...

Otherwise, in your appeal to technological authority, surely you'd rather listen to Yoshua Bengio and Geoff Hinton's perspectives rather than TFA's ramblings... right?

Re: There is no hard takeoff

#136

Post by perlgeek »

Remember the early days of COVID-19? There were some scientists that predicted exponential growth, and I was torn between "it's not really bad yet, maybe there's still hope" and "every exponential growth feels very slow at the start".

I feel like we're with AI, we're in the part of the curve where you can feel it arcing upwards. There's definitely been an uptick in the AI capability. Still no fully self-driving cars, but good enough text generation that the Turing test is in the rear-view mirror. Previously, I've always written "machine learning" or "statistical models" or so, because calling it AI always seemed silly. Not anymore.

Not everything that slowly arcs upwards must become an exponential curve, but we have theories that predict that. There can be many dampening factors that can make it a slower exponential curve, but even that will look like a brick wall if you zoom out far enough.

I think a question we have to ask ourselves is: at what zoom level should we look at the AI capability curve? At the "can re react to an emergency?" time scale, which is, like, days to month? Or at the "can we change society to cope with the new situation?" scale, which is more like decades? At the first time scale, we probably won't even notice a hard takeoff. Viewed over decades, it'll probably very obvious, in retrospect, when it will have happened.

Re: There is no hard takeoff

#137

Post by isodnhionio »

greatpostman wrote: post #129
KnobbleMcKnees wrote: post #77

Earlier quoted context omitted.

My problem with this isn't even that it's a bunch of strung together ideas without a clear narrative or conclusion; enough practice at this combined with actual deep field knowledge results in works like Godel, Escher, Bach . My problem is that GHotz is a shitty writer, has puddle deep knowledge of anything that isn't iPhone firmware, and watching him play at being a freelance tech journo and produce this kind of hol…

Ask yourself why you’re so repulsed by him, it’s probably not due to his writing skill

I would guess it's that GHotz is a shitty writer, has puddle deep knowledge of anything that isn't iPhone firmware, and plays at being a freelance tech journo and produces this kind of hollow pseudo-philosophy that is meant to sound epiphanal but comes across as deeply naïve to anyone with any actual experience in these topics.

Re: There is no hard takeoff

#138

Post by slashdev »

ilaksh wrote: post #82
slashdev wrote: post #53

Earlier quoted context omitted.

> We should start thinking now about limiting the performance of AI hardware. Congratulations, you just ceded the market for AI, and with it any hope of controlling it, to anyone who doesn’t limit it. Like China. That kind of strategy is doomed to failure.

That's my point. That's what makes it dangerous. You can't limit the performance if you want to compete. And like I said, you can't even really keep humans in the loop. So eventually you get something like a 200 IQ GPT tightly integrated with SuperAlphaZero (DeepMind) controlling hypersonic drones, missiles, satellite weapons, etc. planning and operating at 50-100 times human thinking speed engaging in autonomous war…

That seems like a stretch, bit off slippery slope fallacy there.

Will we see AI use in warfare? Absolutely. Will humans trust warfare totally to AI? Not any time soon. I wouldn’t trust my lunch order to AI yet.

Re: There is no hard takeoff

#139

Post by fxtentacle »

KnobbleMcKnees wrote: post #77
scaramanga wrote: post #40

Am I the only one who sees a jumbled stream of barely coherent semi-thoughts and word salad?

My problem with this isn't even that it's a bunch of strung together ideas without a clear narrative or conclusion; enough practice at this combined with actual deep field knowledge results in works like Godel, Escher, Bach . My problem is that GHotz is a shitty writer, has puddle deep knowledge of anything that isn't iPhone firmware, and watching him play at being a freelance tech journo and produce this kind of hol…

I believe you're being unnecessarily mean here. He managed to write enough of a low level GPU driver to compile and execute neural network kernels. Calling that "puddle deep knowledge" is pretty dismissive considering that CUDA is Nvidia's most profitable moat.

Re: There is no hard takeoff

#140

Post by slashdev »

audunw wrote: post #90
slashdev wrote: post #53

Earlier quoted context omitted.

> We should start thinking now about limiting the performance of AI hardware. Congratulations, you just ceded the market for AI, and with it any hope of controlling it, to anyone who doesn’t limit it. Like China. That kind of strategy is doomed to failure.

> and with it any hope of controlling it, to anyone who doesn’t limit it. Like China. I'm a bit annoyed by the use of China in these examples. China absolutely the last country that would allow development of AIs with no limits. There may be other countries willing to allow limitless AI development. China is not it. I would expect the development of AGI to be developed in a similar way to GPT.. you need to feed it lo…

I don’t agree with your characterization of China, but if you don’t like it, pick another. It doesn’t matter. As long as some country isn’t artificially restricting AI, then the research and development moves to that country. Your strategy is completely flawed from a game theoretic point of view.

We’re still far from needing to worry about alignment.

We need to see some real intelligence first. LLMs are not intelligent at all. Doing massive research into how to align them seems pointless.

Post reply on HN