Hacker News Reader

Comment by abrax3141 | Hacker News Reader

This test of compositionality is utterly lame. (FtR: I am a cognitive scientist and AI researcher and my PhD was building computational models of how humans do compositionality - which neither I, nor anyone else can spell, and therefore I will hereinafter refer to simply as C! :-) Anyway, the kind of C that they are seeking is trivial compared to the breadth of the capabilities of human C. Here’s a better example:

You are engaged in a long conversation with someone, perhaps a friend of a friend who you met for lunch. At some point in the conversation they mention that they have a startup and are seeking someone like you. This revelation colors the whole conversation from that point onward. Indeed, each sentence colors the conversation from moment to moment.

But, you reasonably respond, we can’t test that sort of C, modern AIs don’t do even ELIZA-level dialog yet!

What’s the phrase??? “I rest my case?”

View on HN

What kind of insights do you expect a machine to be able to extract? I passed your example to GPT-3, and got back results that seem about the same as I'd expect from a human:

PROMPT:

> This is a test of reading comprehension. Read the following passage and answer the questions below in order.

> Passage:

> "You are engaged in a long conversation with someone, perhaps a friend of a friend who you met for lunch. At some point in the conversation they mention that they have a startup and are seeking someone like you. This revelation colors the whole conversation from that point onward. Indeed, each sentence colors the conversation from moment to moment"

> Questions:

> 1. What is the "revelation" referenced?

> 2. What do you think the person is hoping to achieve by inviting you to lunch?

> Answers:

GENERATED OUTPUT:

> 1. The revelation is that the person has a startup and is seeking someone like the reader.

> 2. It is possible that the person is hoping to recruit the reader for their startup.

And if you ask it, "What does it mean for the conversation to be colored?", what does it answer with?

Or to be tricky, if you were to ask it, "What color was the conversation from one moment to the next?", what would it say?

Oh. Sorry. I seem to have taken the conversation off in a different direction than I had intended. (Foreshadow: The previous sentence is carefully shaped!) When I said "color" I didn't mean to be indicating qualia, although there is that too. What I meant to be indicating is just that each discourse contribution folds into a semantically sensible whole that one can speak (only metaphorically here) as the color of the conversation. One might say (to oneself, if one had a mind to, or was asked): "Oh, but hold on, I didn't realize that this was an interview. I thought that [our mutual friend] just thought that we would get along well. But now that I see it's really an interview, and you're the CEO, well, that makes this a very different situation, and I'll have to put on my 'interview with the boss' face..." and thus like. Again, I don't mean that you would say this explicitly, nor could you, probably, unless you were pressed to do so, in which case you couldn't really completely explain all the nuances, most likely. Nb. (per foreshadow above), each contribution in this, or any discourse, "colors" (or perhaps rather, kneeds together, if you prefer a cooking metaphor) an ongoing collage of situational understanding which comes together to direct, for better or worse, the ongoing complexities of the discourse. But, and here I want to be perfectly clear: Not in a fully forward-going way, because then you'll simply say: well the blah blah state of the whole blah blah network incorporates all that, like Dall-E, etc. melding together everything and smearing it into something sensible. But that's not how people work! In constructing their next action (e.g., sentence) they foreground some aspects of the composed whole, and background others in a goal-directed manner...at least if they're not too drunk.

That's clever. But my response is colored such that I wonder if you didn't generate that text by using some of the other comments, shaping it such that I can't tell whether you're on the way to inebriation or just messing with my head. Or perhaps composing a point. In which case, I myself could use a drink.

Okay, lol literally! If only I could click the up arrow twice, this definitely would deserve it! :-)

It's a lame test, but I don't think most people were claiming that it proves general compositionality. What it does prove is that compositionality is possible with these models, and will likely improve rapidly, as everything else has that they've gotten a toehold into.

Ironically, the very fact that there is now a compositionality benchmark, as Gary points out, is all you really need to know that it's going to fall in the next decade, and probably sooner than that. I'm not aware of any major benchmark dataset upon which enormous progress has not been made in the last few years. And i'd be more than willing to bet anyone anything they'd like that a great deal of progress will be made on this one over the next few.

Marcus seems to be treating it as the hallmark of intelligence (I think he actually uses that phrase), so arguing about whether the hack manages to get the tree into the effective object slot vs the effective subject slot is really not much of a hallmark.

This just fundamentally feels like a bad hill to die on. Compositionality feels like it is:

A) Something AI is currently known to be bad at

B) A matter of degree, not a categorical stumbling block

C) Vague enough of a concept that AI skeptics will continue to complain about it even after the field has moved on

Your example feels less like a description of "compositionality" and more a description of "qualia." It feels an awful lot like dualists trying to carve out a place for magic that no artificial process can reach.

Qualia isn't magic. It's a philosophical term for what it feels like. Colors, sounds, pains are what it feels like. It's dismissive to call it magic. How about instead come up with a good physical explanation of consciousness, showing how the hard problem is mistaken?

Similarly, if compositionally isn't a categorical stumbling block, then show how that's the case. Making a future prediction about what you think computers will accomplish doesn't do that.

I was not implying qualia were magic, insofar as there can be monistic descriptions of qualia. My criticism was that the example appeared to be steering the conversation into a crash course with dualism by invoking qualia-like explanations of compositionality when it was unnecessary to do so.

I believe the "hard problem" is far easier than the dualists' interface problem. It can be explained by viewing consciousness not as the driving force behind our thoughts/feelings, but rather an after-the-fact log the brain keeps for itself. Qualia are therefore distinct from the immediate physics of signals reaching the brain; they are instead the brain's own shorthand description of the impact those signals had on the brain.

Compositionality isn't a categorical stumbling block because the machines are actually getting incrementally better at it. The Winoground paper the article references explicitly says that score on their compositionality benchmark does in fact scale with training dataset size, suggesting that while it is more difficult for AI to discover compositionality during training, there is no reason to think it is impossible

Not only compositionality, but for autopilot I would want an appropriate apprehension, emergency response, courtesy & safety emulator through-and-through and I would want it to more accurately resemble human behavior than what all other AI accomplishments have reached at any one time.

Then it must further be tuned to perform even more appropriately on those points than the most effective human driver.

The average of human drivers is not a valid benchmark.

When you think about it, natural language efforts could probably use more effort tackling this same type of challenge.

Humans too IRL so maybe when it's good for the man and good for the machine in completely undeniable ways, then you're on the right track.

I'd love to read your PhD thesis and papers! I'm also an AI researcher, currently doing a Master's in something else, but compositionality and representation learning is very interesting to me.

Reply on news.ycombinator.com