Hacker News Reader

Comment by saurik | Hacker News Reader

This model doesn't learn and abstract: it just pattern matches and replicates; that's why it was shown exactly replicating regions of code--long enough to not be "de minimis" and recognizable enough to include the comments--that happen to be popular... which would be fine, as long as the license on said code were also being replicated. It just isn't reasonable to try to pretend Copilot--or GPT-3 in general--is some kind of general purpose AI worthy of being compared with the fair use rights of a human learning techniques: this is a machine learning model that likes to copy/paste not just tiny bits of code but entire functions out of other peoples' projects, and most of what makes it fancy is that it is good at adapting what it copies to the surrounding conditions.

View on HN

Have you used Copilot? I have not, but I have trained a GPT2 model on open source projects (https://doesnotexist.codes/). It does not just pattern match and replicate. It can be cajoled into reproducing some memorized snippets, but this is not the norm; in my experience the vast majority of what it generates is novel. The exceptions are extremely popular snippets that are repeated many many times in the training data, like license boilerplate.

Perhaps Copilot behaves very differently from my own model, but I strongly suspect that the examples that have been going around twitter are outliers. Github's study agrees: https://docs.github.com/en/github/copilot/research-recitatio... (though of course this should be replicated independently).

So, to verify, your claim is that GPT-3, when trained on a corpus of human text, isn't merely managing to string together a bunch of high-probability sequences of symbol constructs--which is how every article I have ever read on how it functions describes the technology--but is instead managing to build a model of the human world and the mechanism of narration required to describe it, with which it uses to write new prose... a claim you must make in order to then argue that GPT-3 works like a human engineer learning a model of computers, libraries, and engineering principals from which it can then write code, instead of merely using pattern recognition as I stated? As someone who spent years studying graduate linguistics and cognitive science (though admittedly 15-20 years ago, so I certainly haven't studied this model: I have only read about it occasionally in passing) I frankly think you are just trying to conflate levels of understanding, in order to make GPT-3 sound more magical than it is :/.

What? I don't think I made any claim of the sort. I'm claiming that it does more than mere regurgitation and has done some amount of abstraction, not that it has human-level understanding. As an example, GPT-3 learned some arithmetic and can solve basic math problems not in its training set. This is beyond pattern matching and replication, IMO.

I'm not really sure why we should consider Copilot legally different from a fancy pen – if you use it to write infringing code then that's infringement by the user, not the pen. This leaves the practical question of how often it will do so, and my impression is that it's not often.

The argument I was responding to--made by the user crazygringo--was that GPT-3 trained on a model of the Windows source code is fine to use nigh unto indiscriminately, as supposedly Copilot is abstracting knowledge like a human engineer. I argued that it doesn't do that: that GPT-3 is a pattern recognize that not only theoretically just likes to memorize and regurgitate things, it has been shown to in practice. You then responded to my argument claiming that GPT-3 in fact... what? Are you actually defending crazygringo's argument or not? Note carefully that crazygringo explicitly even stated that copying little bits and pieces of a project is supposedly fair use, continuing the--as far as I understand, incorrect--assertion by lacker (the person who started this thread) that if you copied someone's binary tree implementation that would be fair use, as the two of them seem to believe that you have to copy essentially an entire combined work (whatever that means to them) for something to be infringing. Honestly, it now just seems like you decided to skip into the middle of a complex argument in an attempt to made some pedantic point: either you agree that GPT-3 is a human that is allowed to, as crazygringo insists, read and learn from anything and the use that knowledge in any way they see fit, or you agree with me that GPT-3 is a fancy pattern recognizer and it can and will just generate copyright infringements if used to solve certain problems. Given your new statements about Copilot being a "fancy pen" that can in fact be used incorrectly--something crazygringo seems to claim isn't possible--you frankly sound like you agree with my arguments!!

I think a crucial distinction to be made here, and with most 'AI' technologies (and I suspect this isn't news to many people here) is that – yes – they are building abstractions. They are not simply regurgitating. But – no – those abstractions are not identical (and very often not remotely similar) to human abstractions.

That's the very reason why AI technologies can be useful in augmenting human intelligence; they see problems in a different light, can find alternate solutions, and generally just don't think like we do. There are many paths to a correct result and they needn't be isomorphic. Think of how a mathematical theorem may be proved in multiple ways, but the core logical implication of the proof within the larger context is still the same.

Statistical modelling doesn't imply that GPT-3 is merely regurgitating. There are regularities among different examples, i.e. abstractions, that can be learned to improve its ability to predict novel inputs. There is certainly a question of how much Copilot is just reproducing input it has seen, but simply noting that its a statistical model doesn't prove the case that all it can do is regurgitate.

It's not really comparable to a pen. Because a pen by itself doesn't copy someone else's code/written words. It's more like copying code from Github or if you wrote a script that did that automatically. You have to be actively cautious that the material that you are copying is not violating any copyrights. The problem is Copilot has enough sophistication to for example change variable names and make it very hard to do content matching. What I can guarantee it won't be able to do is to be able to generate novel code from scratch that does a particular function (source: I have a PhD in ML). This brute-force way of modeling computer programs (using a language model) is just not sophisticated enough to be able to reason and generate high level concepts at least today.

One way to look at these models is to say that they take raw input, convert it into a feature space, manipulate it, then output back as raw text. A nice example of this is neural style transfer, where the learnt features can distinguish content from style, so that the content can be remixed with a different style in feature space. I could certainly imagine evaluating the quality of those features on a scale spanning from rote-copying all the way up to human understanding, depending on the quality of the model.

Imagine for a second a model of the human brain that consists of three parts. 1) a vector of trillion inputs, 2) a black box, and 3) a vector of trillion outputs. At this level of abstraction, the human brain "pattern matches and replicates" just the same, except it is better at it.

Human brains are at least minimally recurrent, and are trained on data sets that are much wider and more complex than what we are handing GPT-3. I have done all of these standard though experiments and even developed and trained my own neural networks back before there were libraries that have allowed people to "dabble" in machine learning: if you consider the implications of humans being able to execute turing complete thoughts it should be come obvious that the human brain isn't merely doing pattern-anything... it sometimes does, but you can't just conflate them and then call it a day.

The human brain isn't Turing-complete as that would require infinite memory. I'm not saying that GPT-3 is even close, but it is in the same category. I tried playing chess against it. According to chess.com, move 10 was its first mistake, move 16 was its first blunder, and past move 20 it tried to make illegal moves. Try playing chess without a chessboard and not making an illegal move. It is difficult. Clearly it does understand chess enough not to make illegal moves as long as its working memory allows it to remember the game state.

>The human brain isn't Turing-complete as that would require infinite memory

A human brain with an unlimited supply of pencils and paper, then.

Hmm... but a finite state machine with an infinite tape is Turing complete too. If you're allowed to write symbols out and read them back in, you've invalidated the "proof" that humans aren't just doing pattern matching.

> The human brain isn't Turing-complete as that would require infinite memory.

This is wrong, this is not what Turing completeness is. It applies to computational models, not hardware.

https://en.wikipedia.org/wiki/Turing_completeness

How so? The page you link offers three definitions[1], and all of them require an infinite tape.

You could argue that a stack is missing in my simplified model of the human brain, which would be correct. I used the simple model in allusion to the Chinese room thought experiment which doesn't require anything more than a dictionary.

[1]: https://en.wikipedia.org/wiki/Turing_completeness#Formal_def...

Turing completeness applies to models of computation, not hardware. Otherwise, nothing would be Turing-complete because infinite memory doesn't exist in the real world. Just read the first sentence of what you linked to:

In computability theory, several closely related terms are used to describe the computational power of a computational system (such as an abstract machine or programming language)

Thank you for pointing that out, I was indeed wrong to assume it was used to classify hardware rather than a model.

Human thought isn't anything like GPT thought - humans can spend a variable amount of time thinking about what to learn from "training data" and can use explicit logic to reason about it. GPT is more like a form of lossy compression than that.

This is called prompt engineering. If you find a popular, frequently repeated code snippet and then fashion a prompt that is tailored to that snippet then yes the NN will recite it verbatim like a poem.

But that doesn't mean it's the only thing it does or even that it does it frequently. It's like calling a human a parrot because he completed a line from a famous poem when the previous speaker left it unfinished.

The same argument was brought up with GPT too and has been long debunked. The authors (and others) checked samples against the training corpus and it only rarely copies unless you prod it to.

I don't know if I agree with your argument about GPT-3, but I think our disagreement seems to be besides the point: if your human parrot did that, they would--not just in theory but in actual fact! see all the cases of this in the music industry--get sued for it, even if they claim they didn't mean to and it was merely a really entrenched memory.

The point is that many of the examples you see are intentional, through prompt engineering. The pilot asked the copilot to violate copyright, the copilot complied. Don't blame the copilot.

There also are cases where this happens unintentionally, but those are not the norm.

Transformers do learn and abstract. Not as well as humans, but for whatever definitive of innovation or creativity you wanna run with, these gpt models have it. It's not magic, it's math, but these programs are approximating the human function of media synthesis across narrowly limited domains.

These aren't your crazy uncle's Markov chain chatbots. They're sophisticated bayesian models trained to approximate the functions that produced the content used in training.

Are they Bayesian? It would be good if ML models were Bayesian (they'd be able to show uncertainty better) but they usually aren't.

The model and attention mechanism produces Bayesian properties, but transformers as a whole contain non-Bayesian aspects, depending on how rigorous you want to be in defining Bayesian.

> this is a machine learning model that likes to copy/paste not just tiny bits of code but entire functions out of other peoples' projects

Github could make a blacklist and tell Copilot never to suggest that code. Problem solved. You use one of the other 9 suggestions.

Reply on news.ycombinator.com