A machine learning researcher writes me in response to yesterday’s post, saying:
I still think GPT-2 is a brute-force statistical pattern matcher which blends up the internet and gives you back a slightly unappetizing slurry of it when asked.
I resisted the urge to answer “Yeah, well, your mom is a brute-force statistical pattern matcher which blends up the internet and gives you back a slightly unappetizing slurry of it when asked.”
But I think it would have been true.
A very careless plagiarist takes someone else’s work and copies it verbatim: “The mitochondria is the powerhouse of the cell”. A more careful plagiarist takes the work and changes a few words around: “The mitochondria is the energy dynamo of the cell”. A plagiarist who is more careful still changes the entire sentence structure: “In cells, mitochondria are the energy dynamos”. The most careful plagiarists change everything except the underlying concept, which they grasp at so deep a level that they can put it in whatever words they want – at which point it is no longer called plagiarism.
GPT-2 writes fantasy battle scenes by reading a million human-written fantasy battle scenes, distilling them down to the concept of a fantasy battle scene, and then building it back up from there. I think this is how your mom (and everyone else) does it too. GPT-2 is worse at this, because it’s not as powerful as your mom’s brain. But I don’t think it’s doing a different thing. We’re all blending experience into a slurry; the difference is how finely we blend it.
“But don’t humans also have genuinely original ideas?” Come on, read a fantasy book. It’s either a Tolkien clone, or it’s A Song Of Ice And Fire. Tolkien was a professor of Anglo-Saxon language and culture; no secret where he got his inspiration. A Song Of Ice And Fire is just War Of The Roses with dragons. Lannister and Stark are just Lancaster and York, the map of Westeros is just Britain (minus Scotland) with an upside down-Ireland stuck to the bottom of it – wake up, sheeple! Dullards blend Tolkien into a slurry and shape it into another Tolkien-clone. Tolkien-level artistic geniuses blend human experience, history, and the artistic corpus into a slurry and form it into an entirely new genre. Again, the difference is how finely you blend and what spices you add to the slurry.
“But don’t scientists have geniunely original ideas?” Scientists are just finding patterns in reality nobody has ever seen before. You say “just a pattern-matcher”, I say “fine, but you need to recognize patterns in order to copy them, so it’s necessarily a pattern-recognizer too”. And Einstein was just a very good pattern-recognizer.
“But don’t humans have some kind of deep understanding that pattern-recognition AIs don’t?”
Here’s a fun question: the human brain is undoubtedly the most powerful computer in the known universe. In order to do something as simple as scratch an itch it needs to solve exquisitely complex calculus problems that would give the average supercomputer a run for its money. So how come I have trouble multiplying two-digit numbers in my head?
The brain isn’t directly doing math, it’s creating a model that includes math and somehow doing the math in the model. This is hilariously perverse. It’s like every time you want to add 3 + 3, you have to create an entire imaginary world with its own continents and ecology, evolve sentient life, shepherd the sentient life into a civilization with its own mathematical tradition, and get one of its scholars to add 3 + 3 for you. That we do this at all is ridiculous. But I think GPT-2 can do it too.
Janelle Shane asks GPT-2 its ten favorite animals:
Prompt: My 10 favorite animals are: 1.
My ten favorite animals are:
1. Zebras with a white scar on the back
2. Insiduous spiders and octopus
3. Frog with large leaves, hopefully black
4. Cockatiel with scales
5. Razorbill with wings hanging about 4 inches from one’s face and a heart tattoo on a frog
3. Cockatric interlocking tetrabods that can be blind, cut, and eaten raw:
4. Black and white desert crocodiles living in sunlight
5. Zebra and many other pea bugs
Forget the particular animals for now (don’t worry, the razorbill with wings hanging about 4 inches from one’s face will be back in your nightmares). Notice the way it numbers its list: 1, 2, 3, 4, 5, 3, 4, 5. Last week the two-year-old child who lives next to me counted 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 8, 9, 7, 8, 9, 7, 8, 9 (his mother warns this is an inexact transcription). GPT-2 is instantiated on giant supercomputers; it’s a safe bet they could calculate the square root of infinity in a picosecond. But it counts more or less the same way as a two-year old. GPT-2 isn’t doing math. It’s doing the ridiculous “create a universe from first principles and let it do the math” thing that humans do in their heads. The fact that it counts so badly suggests it’s counting human-style, which makes it amazing that it can count at all.
I find something similar in this post from Tumblr user antinegationism, playing with the lobotomized public-release version of the system:
The American Association for Suicide Prevention and Life-Threatening Behavior (AAPSLB), a professional organization with an overarching goal of promoting suicide prevention, released a recent video encouraging the public to think before they act, on the public’s part. “When we say we’re the only reason to commit suicide, we’re saying they’re wrong. It’s the right thing to do,” said AAPSLB president Steven A. Anderson.
The American Association For Suicide Prevention And Life-Threatening Behavior is not a real organization; the AI made it up as the kind of organization that it thought would feature in a story like this. And AAPSLB is not quite the right way to acronymize the organization’s name. But it’s clearly an attempt at doing so. It’s very close. And nobody taught it how to do that! It’s not just that nobody programmed it in. It’s that nobody thought “Today I shall program an AI to learn how to acronymize on its own in an unsupervised way”. GPT-2 was just programmed to predict text from other text, nothing else. It’s second-level not programmed in. It just happened!
And, uh, it seems to have figured out how to translate things into French. This part is from the official paper:
We test whether GPT-2 has begun to learn how to translate from one language to another. In order to help it infer that this is the desired task, we condition the language model on a context of example pairs of the format ENGLISH SENTENCE = FRENCH SENTENCE and then after a final prompt of ENGLISH SENTENCE = we sample from the model with greedy decoding and use the first generated sentence as the translation. On the WMT-14 English-French test set, GPT-2 gets 5 BLEU, which is slightly worse than a word-by-word substitution with a bilingual lexicon inferred in previous work on unsupervised word translation (Conneau et al., 2017b). On the WMT-14 French-English test set, GPT-2 is able to leverage its very strong English language model to perform significantly better, achieving 11.5 BLEU. This outperforms several unsupervised machine translation baselines from (Artetxe et al., 2017) and (Lampleet al., 2017) but is still much worse than the 33.5 BLEU of the current best unsupervised machine translation approach(Artetxe et al., 2019). Performance on this task was surprising to us, since we deliberately removed non-English webpages from WebText as a filtering step.
In other words: GPT-2 is very bad at translating French into English. But the researchers were surprised to see it could do this at all, since they didn’t design it as translation software, didn’t ask it to learn translation, and didn’t show it any material in French. It seems to have picked up this ability from noticing a few naturally-occurring examples of French in English text:
And here’s One Weird Trick to make GPT-2 summarize articles:
We test GPT-2’s ability to perform summarization on the CNN and Daily Mail dataset (Nallapati et al., 2016). To induce summarization behavior we add the text TL;DR: after the article and generate 100 tokens with Top-k random sampling (Fan et al., 2018) with k=2 which reduces repetition and encourages more abstractive summaries than greedy decoding. We use the first 3 generated sentences in these 100 tokens as the summary. While qualitatively the generations resemble summaries, as shown in Table 14, they often focus on recent content from the article or confuse specific details such as how many cars were involved in a crash or whether a logo was on a hat or shirt. On the commonly reported ROUGE 1,2,L metrics the generated summaries only beginto approach the performance of classic neural baselines and just barely outperforms selecting 3 random sentences from the article. GPT-2’s performance drops by 6.4 points onthe aggregate metric when the task hint is removed which demonstrates the ability to invoke task specific behavior in a language model with natural language.
Again, GPT-2 isn’t good at summarizing. It’s just surprising it can do it at all; it was never designed to learn this skill. All it was designed to do was predict what words came after other words. But there were some naturally-occurring examples of summaries in the training set, so in order to predict what words would come after the words tl;dr, it had to learn what a summary was and how to write one.
If you’re me, the obvious next question is “can it do poetry”. I am very interested to know if the real thing can manage this, but for now I will have to satisfy myself with the lobotomized public-release version. Given Alexander Pope’s Essay On Criticism as a prompt, it says:
O man, who pains the tongue magnified by tymolle,
Though all the fair melodies let no door go as dark as good art.
Yet it would have been, unless the mastermind is so full
Of his own want that this literary country might grow.
Perfect student, thou hast trodden out.
Happy students ruin the purity of what you write ‘Tis art,
For unordered use of decade,’ but art only failing:
A surprise, for new veins which build behind;
It understands there should be line breaks, it understands the approximate correct length of a line of iambic pentameter, it understands how to talk like an overeducated 18th-century dandy – but it doesn’t appreciate rhyme or meter. In retrospect this isn’t surprising; GPT has no idea words sound like anything; it would be shocked to learn anyone uses language as anything other than text strings.
But human poets haven’t appreciated rhyme or meter either for at least a century, so GPT-2 should be able to compose some wicked modern poetry. Friends-of-the-blog Chelsea V. and Qiaochu Y. got it to write a few more stanzas of Allen Ginsberg’s Moloch (human-curated to remove fluff and improve flow):
Moloch whose brief mind is the barrel of gleaming stars! Moloch whose soul is choices of fragrance! Moloch whose framing is glitz and half-cheerios!
Moloch whose daily life is private; his daily strength hoarse! Moloch whose soft desires become culture burning smoke in a palace of eyes! Moloch whose breeding in the caves sweeps wild and rippling flower-like shapes!
Moloch whose pleading is the breath of youth! Moloch whose prudence is the urine of generations! Moloch whose nerves are not privileged; movies abound in quiet caves; people are dressed like ministers! Moloch whose hope receives some help too. Moloch whose eyes are pitiful outstretched, but not blinded. Moloch whose lungs are burning with fire, but not a cautious electrical girl!
Moloch whose sanity is just filled with atoms, gluing beings in physics and complicated libraries. Moloch whose eyes are all innocent animals! Moloch whose sweat and efforts make pure the ocean with burning annihilation! Moloch whose fields dissolve to lightning-ravaged storage houses! Moloch whose eyes sit haunting up Scottish springs!
Moloch whose arms and men are Judah condemns and listens, having understood whom he was and what he man’s weaknesses oppose with iron. We will feast to it like Lord Kippa; we shall play with it like our Pokemon.
Moloch flickering under the pounds and sustaining more fights! Sports and yoga, equally greedy for wealth as for lives! Moloch who had to split his limbs and society into fifty-two atoms! Moloch who cannot survey duty! Moloch for I do not believe in humanity! Moloch for I don’t believe in god! Moloch who, fought Sins and made Sin out of Sin!
Moloch who left me in the glacier! Moloch who helped me to judge my behavior! Moloch who lured me out of the eternal anguish. Moloch who makes all my faults my undoings! Moloch whose deep conversation filled my soul with love and consolation! Moloch who curls up my brains and crumps my heads and screams my scream! Moloch who prepares oil in my hands and burns me alive! Moloch who cries out, “Why do men scream so loud?” Moloch who unleashes his poison in all rivers at high tide! Moloch whose ruins are filled with its cubits! Moloch who fiddles at insidious clockwork!
Aside from this being obviously one of the best things ever, what’s the point?
The point is, GPT-2 has faculties. It has specific skills, that require a certain precision of thought, like counting from one to five, or mapping a word to its acronym, or writing poetry. These faculties are untaught; they arise naturally from its pattern-recognition and word-prediction ability. All these deep understanding things that humans have, like Reason and so on, those are faculties. AIs don’t have them yet. But they can learn.
From the paper:
Prompt: Who was the author of The Art Of War?
Prompt: State the process that divides one nucleus into two genetically identical nuclei?
Prompt: Do you have to have a gun permit to shoot at a range?
Nobody told the model to learn Chinese history, cell biology, or gun laws either. It learned them in the process of trying to predict what word would come after what other word. It needed to know Sun Tzu wrote The Art Of War in order to predict when the words “Sun Tzu” would come up (often in contexts like “The Art of War, written by famous Chinese general…). For the same reason, it had to learn what an author was, what a gun permit was, etc.
Imagine you prompted the model with “What is one plus one?” I actually don’t know how it would do on this problem. I’m guessing it would answer “two”, just because the question probably appeared a bunch of times in its training data.
Now imagine you prompted it with “What is four thousand and eight plus two thousand and six?” or some other long problem that probably didn’t occur exactly in its training data. I predict it would fail, because this model can’t count past five without making mistakes. But I imagine a very similar program, given a thousand times more training data and computational resources, would succeed. It would notice a pattern in sentences including the word “plus” or otherwise describing sums of numbers, it would figure out that pattern, and it would end up able to do simple math. I don’t think this is too much of a stretch given that GPT-2 learned to count to five and acronymize words and so on.
Now imagine you prompted it with “P != NP”. This time give it near-infinite training data and computational resources. Its near-infinite training data will contain many proofs; using its near-infinite computational resources it will come up with a model that is very very good at predicting the next step in any proof you give it. The simplest model that can do this is probably the one isomorphic to the structure of mathematics itself (or to the brains of the sorts of mathematicians who write proofs, which themselves contain a model of mathematics). Then you give it the prompt P != NP and it uses the model to “predict” what the next step in the proof will be until it has a proof, the same way GPT-2 predicts the next word in the LotR fanfiction until it has a fanfiction.
The version that proves P != NP will still just be a brute-force pattern-matcher blending things it’s seen and regurgitating them in a different pattern. The proof won’t reveal that the AI’s not doing that; it will just reveal that once you reach a rarefied enough level of that kind of thing, that’s what intelligence is. I’m not trying to play up GPT-2 or say it’s doing anything more than anyone else thinks it’s doing. I’m trying to play down humans. We’re not that great. GPT-2-like processes are closer to the sorts of things we do than we would like to think.
Why do I believe this? Because GPT-2 works more or less the same way the brain does, the brain learns all sorts of things without anybody telling it to, so we shouldn’t be surprised to see GPT-2 has learned all sorts of things without anybody telling it to – and we should expect a version with more brain-level resources to produce more brain-level results. Prediction is the golden key that opens any lock; whatever it can learn from the data being thrown at it, it will learn, limited by its computational resources and its sense-organs and so on but not by any inherent task-specificity.
Wittgenstein writes: “The limits of my language mean the limits of my world”. Maybe he was trying to make a restrictive statement, one about how we can’t know the world beyond our language. But the reverse is also true; language and the world have the same boundaries. Learn language really well, and you understand reality. God is One, and His Name is One, and God is One with His Name. “Become good at predicting language” sounds like the same sort of innocent task as “become good at Go” or “become good at Starcraft”. But learning about language involves learning about reality, and prediction is the golden key. “Become good at predicting language” turns out to be a blank check, a license to learn every pattern it can.
I don’t want to claim this is anywhere near a true AGI. “This could do cool stuff with infinite training data and limitless computing resources” is true of a lot of things, most of which are useless and irrelevant; scaling that down to realistic levels is most of the problem. A true AGI will have to be much better at learning from limited datasets with limited computational resources. It will have to investigate the physical world with the same skill that GPT investigates text; text is naturally machine-readable, the physical world is naturally obscure. It will have to have a model of what it means to act in the world, to do something besides sitting around predicting all day. And it will have to just be better than GPT, on the level of raw power and computational ability. It will probably need other things besides. Maybe it will take a hundred or a thousand years to manage all this, I don’t know.
But this should be a wake-up call to people who think AGI is impossible, or totally unrelated to current work, or couldn’t happen by accident. In the context of performing their expected tasks, AIs already pick up other abilities that nobody expected them to learn. Sometimes they will pick up abilities they seemingly shouldn’t have been able to learn, like English-to-French translation without any French texts in their training corpus. Sometimes they will use those abilities unexpectedly in the course of doing other things. All that stuff you hear about “AIs can only do one thing” or “AIs only learn what you program them to learn” or “Nobody has any idea what an AGI would even look like” are now obsolete.