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

What we still don’t know about how A.I. is trained

newyorker.com

Post reply on HN
Print view View original

181–190 of 211 posts

Re: What we still don’t know about how A.I. is trained

#181

Post by anileated »

kzrdude wrote: post #138
airstrike wrote: post #130

Earlier quoted context omitted.

that paper is unreal... section 6 on theory of mind is downright scary

It is. Some things I wonder about, it says things like this: > GPT-4 successfully passes the classic Sally-Anne false-belief test from psychology [BCLF85] (which was modernized to avoid the possibility the answer was memorized from the training data) But it's a language model, generalizing text and performing substitutions on it, is what it excels at. "The car is yellow" is "the is " and it can substitute in other th…

The meanings we assign to words like “understanding”, “sentience”, “consciousness” imply really a very high degree of shared context and cultural baggage. If we expand those terms to include systems radically unlike us, then they could be literally anything and everything, from weather systems to laws of physics—but we don’t think of those words in this way; if we can’t reason about sentience or understanding from our vantage point, then for all intents and purposes there is no sentience or understanding. This catch-22 brings together the fallacies of believing in aliens and believing in AGI: we yearn for sentient non-humans, but we only recognize human-like sentience.

One human mind is a lot like another human mind, markedly less like one of an octopus or a dog (still enough similarity that the concept of, e.g., “hurt” kind of makes sense), but really unlike an LLM (and I’m not going to get into an argument as to why an LLM is fundamentally different from a human mind and why we are not even close and possibly can never be to achieving that, the only way of producing new systems like us remaining childbirth; if you don’t agree on that part then it’d be useless to discuss the matter further). We have an uncanny situation where a system radically unlike us can produce output that is mostly similar to what another human mind might produce, but unless we accept that everything around can be conscious (and believe in gods and spirits again) it’s not even a question as to whether the system can understand any of the symbols it produces or have consciousness in the commonly accepted meaning of those terms: it’s only a tool.

Note that I’m not against widening our concept of sentience, just saying it needs to happen if we want to grant an LLM sentience; and if this widening does happen, a sentient language model would be small beans compared to a philosophical revolution we’d have on our hands then.

Re: What we still don’t know about how A.I. is trained

#183

Post by euroderf »

precompute wrote: post #146
woeirua wrote: post #77

Earlier quoted context omitted.

I'm not so sure about this. There is speculation that GPT4 may utilize additional specialized models underneath it for specific tasks.

Exactly, this has been my guess as well. They must have trained the model specifically to write poems, haikus and other things. So that the output looks much more polished than it really is.

I for one want a module that plays around with physics models and proposes practicable FTL-capable mechanisms.

Re: What we still don’t know about how A.I. is trained

#184

Post by noelherrick »

liamconnell wrote: post #173
kossTKR wrote: post #125

Earlier quoted context omitted.

From a european perspective even asking the question "why should be scared of these so called elites" is so bizarre it's almost frightening, i'm sorry. It's a testament to the absurd amount of philanthropic whitewashing, PR and media control these billionaries hold. "Elites" have conspired to exploit the masses throughout 5000 years of civilisation, it's simply a fact of history. It's almost physically impossible to…

If you study history you’ll notice that groups and their leaders are rising and falling, conquering and pillaging and then losing it all. The world is too dynamic for your reductive theory to fit in. Elites compete with each other. They don’t sing kumbaya and cooperatively share the keys to power.

While they will absolutely backstab each other given an opportunity (see the VCs who caused a bank run that primarily affected other VCs), they do seem pretty chummy. Davos is quite literally the summit of the elites.

Re: What we still don’t know about how A.I. is trained

#186

Post by versteegen »

2OEH8eoCRo0 wrote: post #147

> GPT-4’s predecessor, GPT-3, was trained on forty-five terabytes of text data Is this correct? I've seen varying reports of the training set size.

Good catch, that's way off the mark. GPT3 was actually trained over 300B tokens, but some of those were repeats (which is now considered bad practice). Based on Table 2.2 in the GPT3 paper, I calculate the actual training set was only ~238B tokens (assuming they shuffle as they should, rather than selecting each minibatch with replacement). That's approximately 1TB of text (~4 characters per token).

The source of the 45TB number is this quote from the paper:

> The CommonCrawl data was downloaded from 41 shards of monthly CommonCrawl covering 2016 to 2019, constituting 45TB of compressed plaintext before filtering and 570GB after filtering, roughly equivalent to 400 billion byte-pair-encoded tokens.

Re: What we still don’t know about how A.I. is trained

#187

Post by versteegen »

Jensson wrote: post #157
fasterik wrote: post #89

Earlier quoted context omitted.

Isn't this just the AI effect? Whenever there is a breakthrough in AI research, it's no longer considered AI. This happened with search algorithms, game playing, speech recognition, computer vision, etc. etc. https://en.wikipedia.org/wiki/AI_effect Maybe as "tech people" we should give the public a realistic picture of what AI research is. It's solving problems using a diverse set of techniques that include search, o…

The AI effect is actually: Whenever there is a breakthrough in AI research, the AI researchers will consider it AI and lament that others don't agree. That wikipedia article is just from the AI researchers point of view, it isn't what actually happens.

I never heard that one. It's pretty much true. Is it interesting though? But it's totally compatible with the other "AI effect", which also seems pretty much true in my observation.

Re: What we still don’t know about how A.I. is trained

#188

Post by versteegen »

gitpusher wrote: post #127

> "Leaving aside [all of AI's potential benefits] it is clear that large-language A.I. engines are creating real harms to all of humanity right now [...] While a human being is responsible for five tons of CO2 per year, training a large neural LM [language model] costs 284 tons." Presuming this figure is in the right ballpark – 284 tons is actually quite a lot. I did some back of the napkin math (with the help of GPT…

> - 42 years of energy usage by a typical U.S. household

Focus on that one! OpenAI (for example) has approx 375 employees. By your calculations, the CO₂ emissions of those employees driving to work, etc, already dwarfs the quoted 284t CO₂.

Re: What we still don’t know about how A.I. is trained

#189

Post by r_hoods_ghost »

frrae wrote: post #66

GPT Is Not A.I. We tech people should actively go on the offence and educate whomever we can that text inference is not intelligence.

By your definition of intelligence. By my definition of intelligence it is, but then I consider any system that displays goal oriented behaviour and an ability to react to changes in its' environment to be at least minimally intelligent. What is your definition?

Re: What we still don’t know about how A.I. is trained

#190

Post by kossTKR »

liamconnell wrote: post #173
kossTKR wrote: post #125

Earlier quoted context omitted.

From a european perspective even asking the question "why should be scared of these so called elites" is so bizarre it's almost frightening, i'm sorry. It's a testament to the absurd amount of philanthropic whitewashing, PR and media control these billionaries hold. "Elites" have conspired to exploit the masses throughout 5000 years of civilisation, it's simply a fact of history. It's almost physically impossible to…

If you study history you’ll notice that groups and their leaders are rising and falling, conquering and pillaging and then losing it all. The world is too dynamic for your reductive theory to fit in. Elites compete with each other. They don’t sing kumbaya and cooperatively share the keys to power.

I agree, and the world is dynamic, it's just not as dynamic as especially american collective consciousness would have you believe.

There is very little social mobility, wealth transfers very solidly between generations at the absolute top and organisation around PR and Politics is tightly integrated in this class.

Off course you can always fall from grace, but that does not in any way diminish the collective power of this class. That's why it's called a class and not a "person" or one singular family or group of people that clownish conspiracy theories would have you believe.

A good primer to the historical context could be this new book from Cambridge: "The Power of Ritual in Prehistory: Secret Societies and Origins of Social Complexity".

Elites have always formed tightly knit clubs that most couldn't get into.

This is not a reductive theory, it's based on solid academic research on wealth transfer and academic books like the one above.

It's a almost like a biological or physical property of advanced civilizations - like social patterns seen emerging in larger groups of monkeys, or a precursor to the labour divisions seen in ant hives - there's clear distinctions set fourth for an individual at birth or because of location or family, and no amount of ideology is able to change this unless very, very lucky - this is mirrored in the social mobility data.

Post reply on HN