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

The sigmoids won't save you

astralcodexten.com

Post reply on HN
Print view View original

271–280 of 297 posts

Re: The sigmoids won't save you

#271

Post by dehrmann »

rdedev wrote: post #139
LarsDu88 wrote: post #31

I think an interesting thing about recent AI developments is that its all happening right as we hit the diminishing returns side of another "exponential that's actually a sigmoid" which is Moore's law. The naive expectation is that AI will slow down b/c Moore's law is coming to an end, but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell. At so…

IMO we are either limited by data or reaching the limits of what's possible with a transformer architecture. Hardware will get us efficiency but I am not sure if it will lead to smarter models

My doubts in the architecture is how different they are from human intelligence. They need an inordinate amount of training data and lack any sort of generational architectural intelligence.

Re: The sigmoids won't save you

#272

Post by imtringued »

red75prime wrote: post #181
andy99 wrote: post #105

AI has scaled well according to convenient measures. It (neural networks) have the property that whatever you define, they can rapidly be trained master it. We’re able to show that various tasks of increasing complication do not require intelligence and can be framed as autoregressive RL problems. I personally don’t think AI is any closer to sentient intelligence than LeNet; it’s almost trivially clear, we know how i…

> basically how well a universal function approximator can fit to a function we define That's what you've got wrong. We don't define functions that an LLM approximates. Autoregressive pretraining approximates an unknown function that produces text (that is what the brain does). RL doesn't approximate functions, it optimizes objective by finding an unknown function that performs better.

What an interesting perspective! There is just a small problem. How does the human brain work in the phase between being born and learning written language? Seems like a significant bootstrapping problem for your theory.

Re: The sigmoids won't save you

#273

Post by imtringued »

LarsDu88 wrote: post #31

I think an interesting thing about recent AI developments is that its all happening right as we hit the diminishing returns side of another "exponential that's actually a sigmoid" which is Moore's law. The naive expectation is that AI will slow down b/c Moore's law is coming to an end, but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell. At so…

>but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell.

I'm not sure this follows? Research has gone into two bit quantizations that only need a scale factor per block and each parameter merely takes up two bits which means that the operations can be mapped directly onto adders rather than multipliers.

>but if you really think about the models and how they are currently implemented in silicon, they are still inefficient as hell.

The vast majority of analogue components are both space and energy inefficient. Digital won for a reason. You can simply keep scaling to lower voltages and smaller transistors since you only need to distingish between one and zero.

Re: The sigmoids won't save you

#274

Post by imtringued »

LarsDu88 wrote: post #84
cyanydeez wrote: post #33

Earlier quoted context omitted.

they already did put a model into the silicon and it's crazy fast. https://chatjimmy.ai/ I'm pretty sure there's a 3 year design goal starting this year that'll do that to any of the qwen, deepseek, etc models. There's a lot you could do with sped up models of these quality. It might even be bad enough that the real bubble is how much we don't need giant data centers when 80-90% of use cases could just be a silicon c…

And this is an asic that is still operating digitally. Imagine a chip with baked it weights that does its math analogue with 20x reduction in number of circuit elements needed to do a multiplication op. If there's a breakthrough in memristors, you could end up with another 20x reduction in circuit elements (get rid of memory bottlnecks, start doing multiplication ops as log transform voltage addition) The ceiling is…

fp8 and fp4 are already the standard data types, where are you going to find a 20x reduction?

At 1 sign bit, 1 exponent and 2 mantissa bits, there is barely any work done during multiplication into a bf16 accumulator. You're performing two shifted adds at this point.

Re: The sigmoids won't save you

#275

Post by ToValueFunfetti »

RaftPeople wrote: post #261
naasking wrote: post #239

Earlier quoted context omitted.

This is all speculative. We don't understand intelligence, so you literally have no idea whether what we recognize as intelligence is some suitable arrangement of "statistical token generation", especially once you add feedbacks loops.

> "We don't understand intelligence, so you literally have no idea whether what we recognize as intelligence is some suitable arrangement of "statistical token generation"" Do you mean "token" as in the LLM sense? Or are you thinking that thoughts in the human brain are also constructed out of some sort of underlying "token" even though the abstract thought happens and is held before any words are used to try to comm…

LLMs also don't run on tokens internally, they're just the inputs and outputs. The reasoning models do operate (partially) in the token space, but then so do I.

Re: The sigmoids won't save you

#276

Post by tsimionescu »

jfjfnfnttbtg wrote: post #128
tsimionescu wrote: post #100

Earlier quoted context omitted.

While this is very fun as a mathematical exercise, it's completely irrelevant as a real tool for getting a better understanding of unknown processes in the real world. The law only applies for certain types of processes, and is completely wrong for other types (e.g. a human who has lived 50 years may live 50 more, but one who has lived 100 years will certainly not live 100 more). So the question becomes: what type of…

> The law only applies for certain types of processes Did you even read the post? It’s an estimate in the context where you have zero information on which to base an accurate estimate. The author’s point is that if you’re making a different estimate you need to actually say what information is informing that. Human lifespan is obviously not a case where we have zero information, so what is your point in bringing that…

My point is that some processes you encounter will tend to show the Lindy effect, and some won't. If you want to claim that the Lindy effect should be the default assumption, it is up to you to show that it's more likely than not that a random process you encounter shows the Lindy effect. This is not at all an obvious fact, and is quite likely not even true.

So, if most processes are in fact like human lifespans, which do not show a Lindy effect, then it is completely wrong to assume that a random process that you encounter will have this property.

Re: The sigmoids won't save you

#277

Post by FergusArgyll »

energy123 wrote: post #21
krupan wrote: post #17

News flash: predicting the future is hard

The individual who is the best at predicting the future is predicting ASI and full labor automation by 2040: https://xcancel.com/peterwildeford/status/202963666232244661...

Wait till you read his piece about Fordow

Re: The sigmoids won't save you

#278

Post by RaftPeople »

ToValueFunfetti wrote: post #275
RaftPeople wrote: post #261

Earlier quoted context omitted.

> "We don't understand intelligence, so you literally have no idea whether what we recognize as intelligence is some suitable arrangement of "statistical token generation"" Do you mean "token" as in the LLM sense? Or are you thinking that thoughts in the human brain are also constructed out of some sort of underlying "token" even though the abstract thought happens and is held before any words are used to try to comm…

LLMs also don't run on tokens internally, they're just the inputs and outputs. The reasoning models do operate (partially) in the token space, but then so do I.

LLM's generate their output words sequentially based on probability (from learned stats).

Human's don't operate the same way, the thought happens and then the words are generated to reasonably describe that thought.

Re: The sigmoids won't save you

#279

Post by elektronika »

naasking wrote: post #236
stego-tech wrote: post #187

Earlier quoted context omitted.

You’re 100% correct, which is why I opted for a broad investment approach rather than trying to pick “winners”. My thought process RE: superintelligence/AGI is generally this: * I personally don’t believe it’s likely to happen with silicon-based computing due to the immense power and resource costs involved just to get to where we are now; hence why I invest broadly to capitalize on what gains we actually attain usin…

I think this is a bit too pessimistic. Progress in algorithms has matched or exceeded progress in hardware, so the same number of FLOPS spent training GPT-3 years ago would produce a much better model today. Ditto for energy use, and hardware is more efficient at delivering FLOPS. > the widening wealth gaps creating a new underclass not seen since before WW2 I go back and forth on this. I think the reality is that "u…

> AI and automation makes things so cheap that today's underclass lives better than kings ever did

I suggest you go share this opinion with the people living on the street because they can't afford housing.

Re: The sigmoids won't save you

#280

Post by ogogmad »

whatshisface wrote: post #153

If you want a model, here's one: LLMs have never demonstrated the ability to go obviously beyond interpolating their training data. It takes an army of paid data producers solving homework problems to give ChatGPT the ability to do your homework. All vibecoded apps that turned out to be successful could put on a geological soil chart with other apps, probably on GitHub somewhere, on the corners. The prediction? They…

See the recent breakthrough "Erdos Problem 1196" which experts couldn't solve for 60 years until ChatGPT Pro did. ChatGPT's key idea was the use of the "Von Mangolt function" which it showed could finally settle the problem. Terry Tao has condensed the AI's proof to around a page. The problem was well-known to experts (in the field of number theory) but it was a Large Language Model that ultimately solved it - which it did without human help!
Post reply on HN