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

What we know about LLMs

willthompson.name

Post reply on HN
Print view View original

111–120 of 173 posts

Re: What we know about LLMs

#111

Post by Lich »

llm_nerd wrote: post #71

"Crypto VCs & ”builders” making a hard left into AI" This is a humorous intro graphic caption, but this sentiment appears on here constantly and it's self-destructive. This response might seem a bit over the top to a funny graphic, but I am replying to the general "ha ha AI like crypto amirite?" sentiment that is incredibly boring and worn out. When confronted with challenging new technology that we don't understand,…

> "ha ha AI like crypto amirite?"

I don’t think that was the meaning at all. I think the image was supposed to convey how the crypto grifters and con artists were veering into AI to run scams under the guise of AI.

Re: What we know about LLMs

#112

Post by dgant »

bilsbie wrote: post #106
bradfox2 wrote: post #94

Earlier quoted context omitted.

It's still loss being backproped, but the loss is calculated over a different criteria

Ok that makes a lot of sense. Why do they call it reinforcement learning then? Is it not traditional RE such as Q learning?

The distinction making it RL is that the model is training on data produced by the model itself.

The benefit of RL in general is that you're training on states the agent is likely to find itself in, and the cost is needing an agent which explores salient states. Which is why we keep seeing RL as a finishing step after imitation (eg AlphaStar first learning StarCraft from replays)

Re: What we know about LLMs

#113

Post by killernap »

lolinder wrote: post #45
killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

That 8 months seems like a long time to you is indicative of just how fast tech has been moving lately. I expect at least another year before we have a good sense for where we actually are, probably more. However, I'll hazard a guess: I think we haven't seen many real new apps since then because too many people are focused on packaging ChatGPT for X. A chatbot is a perfectly decent use case for some things, but I thi…

News Minimalist LLM vs News Maximalist LLM, garbage-in, garbage-out. ChatGPT giveth and taketh away. Ouroboros levels of disruption.

Re: What we know about LLMs

#114

Post by JustBreath »

FishInTheWater wrote: post #69
JustBreath wrote: post #60

Earlier quoted context omitted.

There's some argument to be made that a form of reasoning happens in a roundabout way when the AI is told to explain it's reasoning. For example if you tell it "Do " and then open a new context and say "Do , explain your reasoning beforehand." you will often get a more accurate response. Granted, it's not that any "Hmm, let me think about that." Deep Thought reasoning occurs, but simply that predicting what the reaso…

This is where the terminology becomes a bit annoying, but there is a key difference in the kinds of reasoning at work here. When you ask LLMs to provide a reasoning, the actual reasoning performed is linguistic; The LLM has (is) a model about language and performs some (limited) reasoning on that model to get an output. But that is explicitly different from reasoning about the abstract question at hand, thus the answ…

Seems like a blurry line between "reason" and "guessing."

Kind of like how an educated guess by a professional is often more accurate than a well reasoned opinion of a layman.

The professional may not have reasoned it so much as intuited, but within that intuition is a lot of wisdom.

I suppose "predicting" is a more precise word than guessing or reasoning.

Guessing implies an arbitrary nature, reasoning implies understanding the concepts at some level.

Re: What we know about LLMs

#115

Post by killernap »

iainctduncan wrote: post #98
killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

I work in tech diligence so I look at companies in detail. I have seen a couple where good machine learning is going to make a massive difference (whether it will keep them ahead of everyone is a separate question). I think it really boils down to: "Is this a problem where an answer that is mostly right and sometimes wrong is still a great value proposition?" This is what people don't get. If sometimes the answer is…

That's interesting. Quite the needle to thread. I wonder how big the market will be for niche models that aren't commodities.

It needs to be something lucrative enough that training the model is not-trivial but not so lucrative Microsoft/Google would care enough to go after. And it somehow needs to stay in that sweet spot even as Nvidia chips away at that moat with each new hardware generation.

Re: What we know about LLMs

#116

Post by killernap »

IKantRead wrote: post #54
killernap wrote: post #13

ChatGPT was announced November, 2022 - 8 months ago. Time flies. Question for HN: Where are we in the hype cycle on this? We can run shitty clones slowly on Raspberry Pi's and your phone. The educational implementations demonstrate the basics in under a thousand lines of brisk C. Great. At some point you have to wonder... well, so what? Not one killer app has emerged. I for one am eager to be all hip and open minded…

> Where are we in the hype cycle on this? Can we stop acting like the Gartner "hype cycle" is anything more than a marketing gimmick created Gartner to validate their own consulting/research services? While you can absolutely find cases that map to the "hype cycle", there is nothing whatsoever to validate this model as remotely accurate or valid for describing technology trends. Where is crypto in the "hype cycle"? I…

Gartner are so great at their job you think they own the concept of hype cycles and rage against them being mentioned while being the one to introduce them to the conversation in the first place :)

Re: What we know about LLMs

#117

Post by famouswaffles »

FishInTheWater wrote: post #95
famouswaffles wrote: post #82

Earlier quoted context omitted.

>Thus we can observe that LLMs do not abstractly reason about the question and it's model. Your conclusion makes no sense. Humans provide increasingly wrong answers as questions get more complex too. Jumping from that to "incapable of abstract reasoning" is silly. You have not "trivially proven" anything at all >The LLM has (is) a model about language and performs some (limited) reasoning on that model to get an outp…

Humans provide increasingly wrong answers as questions get more complex too. Human this, Human that. LLMs aren't humans. "My model is crap but the human brain isn't very good at this either" is irrelevant when we have machines that are not only very good at these tasks but almost perfect at them. Humans make such mistakes precisely because they are not perfect reasoning machines. To compare LLMs to humans is not only…

>Human this, Human that. LLMs aren't humans.

You said you trivially proved something and made up nonsensical lines of reasoning to justify it. If your "proof" can't port to Humans then it's not proof. You are just rambling.

>Humans make such mistakes precisely because they are not perfect reasoning machines.

Nobody is calling LLMs perfect reasoning machines. Your "point" was that they don't reason at all which of none of your ramblings has been able to "prove".

>If these LLMs are not, as I claim, reasoning on language rather than the abstract model of the query, then how come they fail miserably in exactly the ways you would expect where that the case?

They don't. The idea that you must make no mistake reasoning before you can be considered to be reasoning has no ground.

>LLMs generalize to non linguistic patterns. Yes, congratulations, if you turn a problem into a linguistic one LLMs can deal with them.

Can you read ? Did you even bother looking at the link? LLMs don't need patterns to be linguistic to reason over them lol. None of those patterns are turned linguistic. Some of them are arbitrary numbers that resemble nothing like the data they've been trained on.

Re: What we know about LLMs

#118

Post by swyx »

dsubburam wrote: post #105
swyx wrote: post #99

> Transformers can be generally categorized into one of three categories: “encoder only” (a la BERT); “decoder only” (a la GPT); and having an “encoder-decoder” architecture (a la T5). Although all of these architectures can be rigged for a broad range of tasks (e.g. classification, translation, etc), encoders are thought to be useful for tasks where the entire sequence needs to be understood (such as sentiment class…

'Formal Algorithms for Transformers'[1] is a proper account of the architectures and what tasks they naturally lend themselves to, by authors from DeepMind. See sections 3 (Transformers and Typical Tasks) and 6 (Transformer Architectures). Not much on empirical observations, though. [1] https://arxiv.org/abs/2207.09238

ty!

Re: What we know about LLMs

#119

Post by swyx »

pseudonom- wrote: post #104
swyx wrote: post #99

> Transformers can be generally categorized into one of three categories: “encoder only” (a la BERT); “decoder only” (a la GPT); and having an “encoder-decoder” architecture (a la T5). Although all of these architectures can be rigged for a broad range of tasks (e.g. classification, translation, etc), encoders are thought to be useful for tasks where the entire sequence needs to be understood (such as sentiment class…

There's some discussion in section 3.2 of https://arxiv.org/pdf/1910.10683.pdf

ty!
Post reply on HN