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

Are language models deprived of electric sleep?

blog.cbs.dk

Post reply on HN
Print view View original

71–80 of 95 posts

Re: Are language models deprived of electric sleep?

#71

Post by vvrm »

tehsauce wrote: post #13
benreesman wrote: post #2

I’m well aware that this is an unpopular opinion. When your game plan is take a multi-head attention transformer and overfit it on as much of the Internet as we have A100s to fit it on, you’re going to get a distribution, and if you sample from it, it’s going to sound like the Internet. People read too much into this. The technology is obviously powerful borderline dangerous, but it’s still just maximizing P(this-sen…

Worth noting that in the limit the distribution P(human did/said A | context B) is a complete definition of human behavior. If you could model this perfectly, that would be a perfect model of a human ie identical.

LLMs for the most part learn P(someone on internet said A | internet specific context B) given ginormous amounts data. There’s no other type of A, B with that much training data at hand.

Re: Are language models deprived of electric sleep?

#72

Post by benreesman »

soraki_soladead wrote: post #63
benreesman wrote: post #61

Earlier quoted context omitted.

OPT-6.7B is good, but not even close to GPT-3. If you can get GPT-like performance out of a 17B model, you should publish that.

I’m referring to post-training pruning not smaller models. This is already well-studied but it’s not as useful as it could be on current hardware. (Deep learning currently works better with the extra parameters at training time). Retrieval models (again, lots of published examples: RETRO, etc.) that externalize their data will bring the sizes down by about that order as well.

I agree that RETRO is cool. I think you might be stretching it a bit with the applicability, but I take your point.

Re: Are language models deprived of electric sleep?

#73

Post by benreesman »

trention wrote: post #66
Veedrac wrote: post #42

Earlier quoted context omitted.

1. They aren't overfit. 2. It was far from obvious a priori that this could work at all. The existence of the sentiment ‘lol you trained a big model on X so obviously it produces a good probability distribution on X’ only exists because big models proved to be extraordinarily more effective than anyone expected.

>The existence of the sentiment ‘lol you trained a big model on X so obviously it produces a good probability distribution on X’ only exists because big models proved to be extraordinarily more effective than anyone expected. No, the "popularity" of that sentiment exists because of their "effectiveness". That sentiment was existing and being voiced 10 years ago.

You’ve stated this point better than I did.

Otherwise brilliant and rational people just go all mystical when we start talking about meaningless words like “intelligence” or “consciousness”.

An animal or a human or a big matrix perform X well on Y task. That’s quantitative and objective.

All this “is it smart” bullshit is thinly-veiled “how does the world still rotate around my subjective experience given this thing writes better Tolkien fan fiction than I do”.

Performance on tasks. Everyone else shuffle over to the Philosophy department.

Re: Are language models deprived of electric sleep?

#74

Post by benreesman »

trention wrote: post #68
the8472 wrote: post #48

Earlier quoted context omitted.

Part of the evaluation procedures for the models like these is checking whether the output they produce is present in the training set. The fraction of samples being verbatim reurgitations is low unless you prompt-engineer your way to that, copilot being able to recite the fast inverse square root function being a common example. But the ability to recite some samples is not proof of inability to create novel outputs…

>The fraction of samples being verbatim reurgitations is low Copilot spitting out a function out of its training data with changed variable names to match those in your file(s) - and no one is actually testing what proportions of results those are - is still regurgitation.

Copilot is weird.

The N in the NLP training set and arch implies a fuzzy match.

Why anyone would rather start with a plausible but broken-in-a-subtle way buffer rather than a blank one is beyond me.

Re: Are language models deprived of electric sleep?

#75

Post by benreesman »

So green text is easier to spoof than HN comments, but my, uh, friend tells me that the HN comment corpus is floating around.

Someone is eventually going to speed run 10k karma on HN and YouTube it. This year? Next? Who knows. But not a decade.

https://youtu.be/efPrtcLdcdM

Re: Are language models deprived of electric sleep?

#76

Post by wwweston »

boredemployee wrote: post #8

Just a side note, I was showing some friends gpt3 and asking some trivial questions, trying to impress them, when one of them asked me: "Is it any different from like asking google for the same information?" I mean, In a pure naive and practical point of view she wasn't that wrong.

Maybe from a theoretical point of view too — the internet is a generative network, yes?

Re: Are language models deprived of electric sleep?

#77

Post by moonchild »

joshspankit wrote: post #62
boredemployee wrote: post #8

Just a side note, I was showing some friends gpt3 and asking some trivial questions, trying to impress them, when one of them asked me: "Is it any different from like asking google for the same information?" I mean, In a pure naive and practical point of view she wasn't that wrong.

I watched Google grow up, and for people who’s main experience is with the last decade or so, I’d say they have a very insightful view. Yes, some have argued that Google can’t make content and from a technological standpoint that’s right, but here’s the fulcrum: from the perspective of “I asked a question and got a result that seems right though I’m not sure”, there’s shockingly little difference between GPT-3 genera…

> low-effort SEO-optimized content found on the first page of Google

often, this is also mechanically generated

Re: Are language models deprived of electric sleep?

#78

Post by telotortium »

benreesman wrote: post #2

I’m well aware that this is an unpopular opinion. When your game plan is take a multi-head attention transformer and overfit it on as much of the Internet as we have A100s to fit it on, you’re going to get a distribution, and if you sample from it, it’s going to sound like the Internet. People read too much into this. The technology is obviously powerful borderline dangerous, but it’s still just maximizing P(this-sen…

You better watch what you're saying - soon the champion (in the medieval sense) of ML, Gwern, will come to defend his liege's honor. I see the other champion, Veedrac, has already come out :)

Re: Are language models deprived of electric sleep?

#80

Post by strogonoff »

MontyCarloHall wrote: post #6
benreesman wrote: post #2

I’m well aware that this is an unpopular opinion. When your game plan is take a multi-head attention transformer and overfit it on as much of the Internet as we have A100s to fit it on, you’re going to get a distribution, and if you sample from it, it’s going to sound like the Internet. People read too much into this. The technology is obviously powerful borderline dangerous, but it’s still just maximizing P(this-sen…

I completely agree. Nonetheless, this approach of taking a fairly complex writing prompt and simply regurgitating contextually similar text from the internet often yields 100% meaningful results. This suggests that our intelligence is a lot more mimetic than we'd like to believe, which is perhaps an even more unpopular opinion.

Even if a sentence is meaningful standalone in vacuum, no communication between humans ever happens in such a way. It always matters who said it and why; we never take the substance independently from the agenda and the mind behind it, the context in which that mind existed and its relationship with our own, and so on.

This strikes me as a crucial part of what intelligence (whether in humans or other creatures) means.

In case of an algorithm, there is no intent of its own—except the intent and the minds of humans who trained it, ran it and supplied inputs.

IMO the onus is on AGI believers to prove that material world is somehow the source of consciousness or intelligence (it wouldn’t hurt defining the terms first either). Otherwise it’s a philosophical position, and while one is entitled to hold their own one is not entitled to force it onto others.

Post reply on HN