Item — HN Reader

> Some people have completely unrealistic expectations about what large-scale language models such as GPT-3 can do.

Just want to point out that he's saying the people on the upper end of the expectation distribution are wrong, not the people in the middle of it. So if you're takeaway from this is that GPT3 is nothing special, that's probably the wrong message.

2 comments

His next paragraph claims that Nabla "debunks" the idea that "large language models" can be used in healthcare.

That's not just "some people have unrealistic expectations" it's "this tool, when when more advanced and find tuned, will never be appropriate to use in a very broad class of use cases".

He also says "GPT-3 has no knowledge of how the world works", which is clearly an overstatement meant to clear up hype, but is untrue. For example, GPT-3 knows more trivia than I do.

I'm having trouble wrapping my head around LeCun's thinking regarding the Nabla reference. The Nabla link is just a blog post by three people without any technical details provided at all. How can this possibly "debunk" anything?

>For example, GPT-3 knows more trivia than I do.

no it doesn't, GPT-3 is a very sophisticated parrot. it doesn't know any trivia, it knows how to put the most likely string of characters next to the one it just saw, it doesn't matter what the text represents. That's the difference between you and the model.

It's basically the Chinese room. You can make an analog GPT-3 by asking a question, recording your answer, handing someone who doesn't understand a word of your language the giant box of tapes, and she tries to match them together until she appears to make sense to listeners

AlphaGo doesn't know anything about the game of go. It "just" manipulates symbols, runs instructions on its CPUs and GPUs, illuminates pixels on the screen for a human to see.

It just also "happens" to be the case that if you interpret those pixels as go moves and play those moves against the world champion human go player then eventually that human will hold a press conference announcing to the world that AlphaGo has won the match.

Yes that is entirely true, but the world of Go is very small and unambiguous in terms of goals, rules and so on. A little less than chess, but it's still a game with a well structured world and simple rules. AlphaGo understands as little about Go as Stockfish does about Chess or your TI-83 understands about maths. That they're good at it has almost nothing to do with (generalised) intelligence, and it's no coincidence that the performance of RL say, in games is insanely better than it is when applied to real-world, open-ended problems.

Deep Mind made a 'superhuman' Breakout bot, but if you messed with the panel at the bottom it sucked. It hadn't learned what the panel is or what the game is about, it had just learned to rearrange pixels. You can get away with this in one version of Breakout, but in the real world it's brittle and useless

Ok, fine. It doesn't "know" any trivia. It'll still crush me in a trivia contest:

"Q: Who was president of the United States in 1801? A: Thomas Jefferson was president of the United States in 1801.

Q: What city is in the northwest corner of Ohio? A: Toledo is in the northwest corner of Ohio.

Q: What highway do you take to get to North Lake Tahoe from Oakland? A: Highway 80 is the highway you take to get to North Lake Tahoe from Oakland.

Q: What are Darius's abilities in League of Legends? A: Darius's abilities in League of Legends are: Apprehend, Decimate, Noxian Guillotine, Crippling Strike, Apprehend [sic], and Hemorrhage.

Q: Who did Capablanca defeat to become world chess champion? A: Capablanca defeated Lasker to become world chess champion."

Of those 5 questions, I could answer #1 after deploying a mnemonic and some math, and #2 if you gave me multiple choice of the top 5 cities in Ohio, and I would miss the rest.

> which is clearly an overstatement meant to clear up hype, but is untrue

It all depends on your definition of knowledge. Under a certain definition you could say that GPT-3 knows basically nothing.

If someone teaches me to repeat perfectly something very smart in a language I don't know, without explaining to me what that thing is, do I have knowledge about this?

The same argument can be made about those kind of models, the knowledge they have is about the structure of the language and what word is most likely to come next, but they have no way to ground those words in actual relation with the world.

Aka. the Chinese room argument. However, I'm not so sure us people are little more than just pattern matching machines. When I start to talk (or write, as I'm doing now), the words kind of just flow out. I can make the argument, that I understand the "real" world, but do I really?

> I'm not so sure us people are little more than just pattern matching machines

Yes that's what leads to multiple definition of what knowledge really is. Yann LeCun believe we are more than just that, hence why he is saying GPT-3 would have no knowledge.

Yes, I think the focus on "getting to the moon,", to use his analogy, ignores the fact that GPT-3 is an SR-71 in a world of 19th century balloons. It may not get to the moon, but it definitely points the way to lots of useful stuff. There is a lot of boilerplate text in the world that is well-suited to auto-generation by a GPT-3-style model. And currently a lot of people employed to generate that text, at significant aggregate expense.

I hate this concept. I have a lot of “boilerplate” real life interactions but I would never replace them with a robot. If people really think nuts and bolts writing is not a worthwhile enough endeavour for humans to do, then shame. What an immense forfeiture, to never again be surprised by a furniture catalogue.