For anyone else who doesn’t want to deal with Facebook, here’s the post:
Some people have completely unrealistic expectations about what large-scale language models such as GPT-3 can do.
This simple explanatory study by my friends at Nabla debunks some of those expectations for people who think massive language models can be used in healthcare.
GPT-3 is a language model, which means that you feed it a text and ask it to predict the continuation of the text, one word at a time. GPT-3 doesn't have any knowledge of how the world actually works. It only appears to have some level of background knowledge, to the extent that this knowledge is present in the statistics of text. But this knowledge is very shallow and disconnected from the underlying reality.
As a question-answering system, GPT-3 is not very good. Other approaches that are explicitly built to represent massive amount of knowledge in "neural" associative memories are better at it.
As a dialog system, it's not very good either. Again, other approaches that are explicitly trained to perform to interact with people are better at it.
It's entertaining, and perhaps mildly useful as a creative help. But trying to build intelligent machines by scaling up language models is like a high-altitude airplanes to go to the moon. You might beat altitude records, but going to the moon will require a completely different approach.
It's quite possible that some of the current approaches could be the basis of a good QA system for medical applicatioms. The system could be trained on the entire medical literature and answer questions from physicians.
But compiling massive amounts of operational knowledge from text is still very much a research topic.
To remind people: Yann LeCun is an engineering superstar who was working on neural networks at Bell labs long before they were cool.
Those handwritten digits that are a scourge today (e.g. 'would any of these methods work on a different set of symbols?' is unasked) came from a competition to develop a commercial zip code reader for the U.S. Postal Service post back in the day of the Apple Newton. He won it!
I was a bagman for text classification data in the early 2000's and his reviews of the results you got using methods of the time (Naive Bayes, Rocchio, Perceptron, SVM) showed a depth of thought and attention to detail which helped me pick and choose tools to make classifiers with fairly predictable performance and development paths.
GPT-3 on the other hand does a good job of spouting nonsense like Peter Thiel and that has something to do with it's emotional appeal. People make fun of it and laugh at the mistakes it makes like those videos where somebody kicks down one of those Boston Robotics dogs: it's just good enough to be an object for those sort of feelings.
To remind people: Yann LeCun worked on artificial neural networks (ANN) during the period where they were actively shunned by most of the scientific community. You could barely publish a paper on ANN.
Just to demonstrate, one the most common books during period, "Artificial Intelligence: A Modern Approach, 2nd ed" by Norvig, 1080 pages, has less than one (1!) page dedicated to ANNs. I personally think Norvig is an idiot with regards to Artificial Intelligence, and his book (used in 1500 schools in 135 countries and regions) singlehandedly slowed down the progress of AI by a few years, until a new generation of students outgrew this archaic book.
> Artificial Intelligence: A Modern Approach, 2nd ed
Published in 2002. At that point, ANN research had reached a pretty hard plateau with very few tangible results. Faulting Russel and Norvig for not going into depth about ANNs is kind of like faulting Richard Feynman for not going into depth about quantum computers in the Feynman Lectures.
Also, a lot of the subsequent work and breakthroughs on ANNs has been done at Google under Norvig's leadership as Director of Research.
As opposed to the AI techniques taught in the AIMA book (KR and logic reasoning), which had plateau'ed in the 70s...?
Norvig had to be pretty clueless to decide ANNs are such a dead-end, that they don't deserve even a chapter in his book, where all around him there are biological living proofs that neural networks are probably a pretty good bet for AI...
(Note: I held the same opinion in the mid 90s when I reviewed his 1st edition and I'm definitely no Feynmann-level. It's just common sense.)
> all around him there are biological living proofs that neural networks are probably a pretty good bet for AI...
100 years ago you would have been arguing that all around you are living proofs that ornithopters are probably a pretty good bet for artificial flight. You would have been wrong about that too.
We seem to be re-enacting the Symbolic vs Connectionist AI debate of the 80s, poorly.
All I'm saying is, Norvig should have been more humble and included a chapter or two about ANNs, with all the research accumulated thus far, instead of betting 100% for the symbolic approach. Let the next generation of students learn both approaches and decide for themselves. It's sad that a whole generation of students was taught AI from this archaic book.
> All I'm saying is, Norvig should have been more humble and included a chapter or two about ANNs
No, that is not all you're saying. You opened with this:
"I personally think Norvig is an idiot with regards to Artificial Intelligence,"
Not only did you lob an ad hominem at one of the most respected members of the community simply for making an editorial decision 18 years ago that you happen not to agree with today, you did it from a newly created anonymous HN account, and then you tried to deny it. Your conduct here has been thoroughly dishonorable. You should be ashamed of yourself.
I think Norvig is a smart person, I enjoy his books, papers and jupyter notebooks, but I always thought he was pretty clueless regarding AI, as history indeed demonstrated. That's not an ad hominem.
What is shameful was the extreme shunning of the mainstream scientific community to ANNs, in 90's-00's decades, to the point that it was considered career suicide to publish a ANN paper. I believe Yann LeCun has said similar things in the past[0], reminiscing the time when it took him several years(!) to get an ANN paper accepted for publication.
It is not fair to call Norvig an idiot. When I was studying AI in grad school, around 2002-2003, just about everyone thought that artificial neural networks were a dead end, compared to approaches like support vector machines. Sometimes the scientific consensus is wrong, and it takes a few heroic figures plugging away to prove it. That doesn't mean that everyone in the mainstream is an "idiot".
applied rigorous math (e.g. when computer science was new) to prove that a certain kind of single-layer neural network couldn't solve certain problems. (Can't learn XOR) It is like proving that it takes N log N comparisons to sort N items.
This dampened interest in neural networks for a long time but the "geometrical thinking in hyperdimensional space" is what the field is all about today.
Interest in neural networks was renewed with Werbos's (1975) backpropagation algorithm. There was continued progress in ANNs all this time.
I think the aversion to ANNs during the 90s was more philosophical and aesthetic - ANNs math is indeed "ugly" compared to symbolic logic, bayesian inference, SVM (in the 00's), and many other traditional AI methods.
For some context. The 2nd edition was published in 2002 (so maybe written in 2001-2002?). The fourth edition published in 2020 seems to have a bunch more things on NNs.
High altitude planes going to the moon is a beautiful analogy. I think this is what I’ll use to explain to less technical friends why I think we’re still many years from self driving cars.
The question isn't whether high-altitude planes can go to the moon, it's whether human intelligence is closer to the clouds or to the moon. For all the talk about how language models "just" learn correlations, there's a remarkable dearth of evidence that humans do something qualitatively different.
> For all the talk about how language models "just" learn correlations, there's a remarkable dearth of evidence that humans do something qualitatively different.
GPT3 doesn't know the difference between a given set of characters and the idea/object the characters represent. It can associate "river" and "stream" and "water" but has no understanding beyond that they appear in patterns together. It couldn't possibly make the connection that river and streams are bodies of water, because there is no association with reality.
GPT3 wouldn't even know the difference between human language and characters derived from some random data source.
The only thing it does is identify deeply complex patterns, as long as there are humans around to notify it when it's doing a good job. It's going to be very useful for auto-complete, and jumping in to help users finish repetitive tasks, along with the other stuff ML is good at, but it's simply a GIGO pattern recognition system.
So I think you have it exactly backwards -- there is a dearth of evidence that AGI is even remotely possible. We have known the full anatomy of the C. Elegans ringworm since 1984 -- it's 1mm long and has 300 neurons. There is a foundation dedicated to replicating it's behavior[1], and all they have achieved is complex animation.
> It couldn't possibly make the connection that river and streams are bodies of water, because there is no association with reality.
Yet if you asked it if rivers and streams are bodies of water it would probably say that they are.
Likewise if I asked you if black holes and neutron stars are both celestial bodies you would say yes... but you've presumably never seen them, only read about them.
Now I think you could argue that you could ultimately tie your knowledge back to some reality, like star ties to the sun and how you've seen the sun, but I haven't been so convinced that we know enough about how the mind works to be sure that the distinction between form and meaning is real.
Assuming there is no "soul" which makes meat life special, I don't see any fundamental problems with building simulated intelligence.
Are you just dumbing down humans to match the model? LeCunn's post is very sparse on detail, but the point is that humans can easily reason about a vast number of things that any form of sequential language model cannot. That alone is evidence that humans are doing something qualitatively different.
It isn't conclusive evidence however, and larger models may produce significantly more human like results. But from what we know about how gpt-3 works, all the evidence is on the side of it not resembling human intelligence.
Exactly - the analogy fails if a few assumptions we have about ourselves or what GPT-3 is actually “doing” are wrong. Until we hit some asymptotic limit on training these kinds of language models, I’m withholding judgement on what such a model will be capable of representing if/when that limit arrives.
> there's a remarkable dearth of evidence that humans do something qualitatively different.
Perhaps now, but if history is any indication, when we (as humans) think we have a good grip on how something really works (like human intellect in this example), we've been wrong.
We model the world around us from observation and testing, find our errors, remodel, and improve over time.
Then at some point we find some piece of information that shows us our model was a decent approximation, but fundamentally wrong, and that we need to start from scratch.
If we find that we want to go beyond the moon (and we eventually will), or that the moon is further than we think, we'll again need a different approach.
I always feel like there's a certain beauty and cosmic humor to it.
I have been using that analogy to explain why Tesla is many years away from a self driving car. Several others are building something that is fundamentally different.
Several things. Mostly that it has to be rolled out slowly because people lives are at stake so testing can run for quite a bit longer than you’d expect. Also that everyone wants to be the one who makes the breakthrough so companies will claim its right around the corner (like fusion) repetitively, i.e. Tesla saying it’d be here in 2018. We’re just at a point where these things can use parking lots so I wouldn’t expect a complete rollout several years as systems are built on top of other systems that have been widely tested and confirmed to work.
> GPT-3 doesn't have any knowledge of how the world actually works. It only appears to have some level of background knowledge, to the extent that this knowledge is present in the statistics of text. But this knowledge is very shallow and disconnected from the underlying reality.
Without excessive effort, humans don't have any knowledge of how the world actually works. They only appear to have some level of background knowledge, to the extent that this knowledge is present in their faint memories. But this knowledge is very shallow and disconnected from the underlying reality.
For example, all humans have the notion of object permanence, developed by about six months of life. Object permanence is the notion that things don't go away just because you can't see them.
ML systems need to be specifically trained to have object permanence, and GPT-3 almost certainly does not possess it.
Like, I get that it's hip to booster ML and GPT-3, and all of the stuff humans can do seem trivial, but it's really not the case and is something that is holding progress in AI back massively.
Object permanence isn't even defined in the stateless world of the GPT-3 API.
My comment was merely a jab at Yann's poor argument. I don't find humans to be trivial at all, but neither do I believe that they are infinitely complex.
The linked Nabla article is fair, albeit I would appreciate more technical details. It seems to be using the API in zero-shot fashion, which is not what one would do to get the most out of it.
I understand that you are being sarcastic, but I'll pretend that you're not ;)
There was a video showing the pattern recognition layer of a car 'forget' about objects it had previously correctly identified. There clearly is room for improvement, and some level of concern is warranted, but it isn't as much of a big deal as it seems, because surely at the higher level objects are remembered. To draw an analogy with human cognition, humans also aren't consciously aware of all the objects in their field of view all the time, but that doesn't mean they forget their existence.
I'm honestly not. I've seen that video, and given my experience of vision models, am actually concerned that object permanance is not being focused on.
> There clearly is room for improvement, and some level of concern is warranted, but it isn't as much of a big deal as it seems, because surely at the higher level objects are remembered.
I would be very surprised if this were true.
> To draw an analogy with human cognition, humans also aren't consciously aware of all the objects in their field of view all the time, but that doesn't mean they forget their existence.
This is exactly object permanence, in that babies start looking in the right direction for objects that disappear. I assume that Waymo at least have some kind of model for this, but am not convinced that it's particularly effective.
Personally, I think that self-driving will require a limited theory of mind model, and we are nowhere near that.
At the very least there must be some kind of a trajectory model which implies that multiple observations of the same object are analyzed as a sequence. How would it work without object permanence? Would it just erase objects undetected in the current frame and pretend they never existed? I suspect you're too focused on the vision aspect of the vehicle control system.
I agree that GPT-3 should not be used for medical applications, but I disagree that it's "not very good" as a dialog system. I've found it to be insanely good though it may require effective prompt engineering to work well.
As both a GPT-3 sceptic and LeCun-sceptic (nothing personal, I'm just sceptical of a lot of the modern hype and seeming cliquey-ness of the AGI "progress" centres), this summary is very useful, but sounds a bit sour-grapesy from LeCun...
Some people have completely unrealistic expectations about what large-scale language models such as GPT-3 can do.
This simple explanatory study by my friends at Nabla debunks some of those expectations for people who think massive language models can be used in healthcare.
GPT-3 is a language model, which means that you feed it a text and ask it to predict the continuation of the text, one word at a time. GPT-3 doesn't have any knowledge of how the world actually works. It only appears to have some level of background knowledge, to the extent that this knowledge is present in the statistics of text. But this knowledge is very shallow and disconnected from the underlying reality.
As a question-answering system, GPT-3 is not very good. Other approaches that are explicitly built to represent massive amount of knowledge in "neural" associative memories are better at it.
As a dialog system, it's not very good either. Again, other approaches that are explicitly trained to perform to interact with people are better at it.
It's entertaining, and perhaps mildly useful as a creative help. But trying to build intelligent machines by scaling up language models is like a high-altitude airplanes to go to the moon. You might beat altitude records, but going to the moon will require a completely different approach.
It's quite possible that some of the current approaches could be the basis of a good QA system for medical applicatioms. The system could be trained on the entire medical literature and answer questions from physicians. But compiling massive amounts of operational knowledge from text is still very much a research topic.