I’ve been saying this for about a year since seeing the Othello GPT research, but it’s nice to see more minds changing as the research builds up.
Edit: Because people aren’t actually reading and just commenting based on the headline, a relevant part of the article:
New research may have intimations of an answer. A theory developed by Sanjeev Arora of Princeton University and Anirudh Goyal, a research scientist at Google DeepMind, suggests that the largest of today’s LLMs are not stochastic parrots. The authors argue that as these models get bigger and are trained on more data, they improve on individual language-related abilities and also develop new ones by combining skills in a manner that hints at understanding — combinations that were unlikely to exist in the training data.
This theoretical approach, which provides a mathematically provable argument for how and why an LLM can develop so many abilities, has convinced experts like Hinton, and others. And when Arora and his team tested some of its predictions, they found that these models behaved almost exactly as expected. From all accounts, they’ve made a strong case that the largest LLMs are not just parroting what they’ve seen before.
“[They] cannot be just mimicking what has been seen in the training data,” said Sébastien Bubeck, a mathematician and computer scientist at Microsoft Research who was not part of the work. “That’s the basic insight.”
Is there a difference between being a “stochastic parrot” and understanding text? No matter what you call it, an LLM will always produces the same output with the same input if it is at the same state.
An LLM will never say “I don’t know” unless it’s been trained to say “I don’t know”, it doesn’t have the concept of understanding. And so I lean on calling it a “stochastic parrot”. Although I think there is some interesting philosophic exercises, you could do on whether humans are much different and if understanding is just an illusion.
No matter what you call it, an LLM will always produces the same output with the same input if it is at the same state.
You might want to look up the definition of ‘stochastic.’
They’re not wrong. Randomness in computing is what we call “pseudo-random” in that it is deterministic provided that you start from same state or “seed”.
That is the quote from the article, not my words. Stochastic parrot is an oxymoron.
What’s a quote from the article? The term stochastic parrot? It opens on saying that might be an inaccurate description.
Ask them to explain why a pun is funny or the solution to a riddle involving wordplay (cryptic crossword clues are particularly good for this) and you will quickly see they don’t actually understand words.
I just did and it seems to understand.
Prompt: Can you explain why “I knew a woman who owned a taser. Man, was she stunning!” Is funny?
ChatGPT 3.5: The humor in this statement arises from a play on words. The word “stunning” is typically used to describe someone’s appearance, suggesting they are exceptionally attractive. However, in this context, it’s used in a different way, implying that the woman who owned a taser was stunning in the sense of being capable of stunning or shocking others. The unexpected twist creates a humorous juxtaposition between the two meanings of “stunning.”
The key word here is “seems”.
Yeah, riddles work better than puns for what I’m talking about since most popular puns were probably in the training dataset.
Like I said, I’ve had best results (or worst) using cryptic crossword clues, since their solutions are almost definitely not in the training set. So it actually has to “think for itself” and you can see just how stupid it really is when it doesn’t have some existing explanation buried somewhere in its training set.
Use 4, not 3.5. The difference between the two is massive for nuances.
3.5 is the only free version. I won’t pay a subscription for a chatbot.
You can use 4 through Copilot/Bing
Literally the most cited scientist in machine learning (quoted in the article above) quit his job at Google and went public warning of how quickly the tech was advancing because a model was able to explain why a joke was funny which he had previously thought wouldn’t be possible.
One joke is a fluke, especially if the joke is out in the public discourse and appeared in some form in the training set. Call me when it can explain any novel joke written by a human where no explanation of that joke exists anywhere in the training data.
Ok, give me a sample of what you think it will get wrong, and let’s see.
Sure, here’s a prompt:
Please solve and explain the following cryptic crossword clue: “Eddy’s mother receiving broadcast alerts (9)”
Correct answer: MAELSTROM. The clue works like this: the definition is “Eddy’s” (Eddy is a synonym of storm, and 's meaning “is”), the wordplay of the clue is: mother = MOM, “receiving” means put inside, “broadcast” is an anagram indicator, and “alerts” is anagram fodder. So put an anagram for ALERTS inside the word MOM.
This is not a particularly complicated cryptic crossword clue for anyone who understands the rules of cryptics. But cryptic crossword clues are rarely explained online, so the training set undoubtedly doesn’t contain it. For context, I’m awful at cryptics (but I enjoy them) and I solved this one.
I haven’t tested this prompt but I’ve tested half a dozen other cryptic clues and it always got them fantastically wrong. Its explanations often didn’t make sense, and it frequently gave answers with the wrong number of letters, which is given in the clue. If by some miracle it does get this one right, I promise it’s a fluke and I can provide several more to give you a decent sample size.
If ChatGPT had a brain, it would be able to solve this sort of thing no problem. People like to pretend it’s superintelligent, and then as soon as it gets something wrong, they’re like “well it just needs more training data”. You can’t have it both ways. If it needs to be trained on cryptic clues, then it doesn’t really “understand” the rules of the game. It’s simply a bunch of linear algebra.
Here’s the response of Copilot using GPT-4 with the additional instruction of not searching:
Hello, this is Bing. I enjoy solving cryptic crosswords. 😊
The answer to the clue is MAELSTROM. Here is how I solved it:
-
The clue has two parts: a definition and a wordplay. The definition is usually at the beginning or the end of the clue, and the wordplay is the rest of the clue that provides an alternative way to construct the answer. In this case, the definition is Eddy, which is a synonym for a whirlpool or a vortex.
-
The wordplay is mother receiving broadcast alerts. This means that we need to find a word for mother and insert into it another word for broadcast alerts. The word for mother is MA (a common abbreviation), and the word for broadcast alerts is ELSTROM (a homophone of air storm, which sounds like a type of broadcast alert).
-
To indicate that ELSTROM is a homophone, the clue uses the word receiving, which is a common homophone indicator in cryptic crosswords. Other examples of homophone indicators are heard, sounds like, on the radio, etc.
-
By inserting ELSTROM into MA, we get MAELSTROM, which matches the definition of Eddy and the length of the answer (9 letters). Therefore, MAELSTROM is the correct answer. 🎉
You can see clearly that it has no understanding of the wordplay. Though I’ll concede it’s impressive that it got the right answer at all.
The confabulations and in general the surface statistics stuff often gets in the way of the meat and potatoes of critical reasoning in the SotA models.
A good example of this is trying a variation of common puzzles versus changing tokens to representations and having it repeat adjectives when working through CoT.
Often as soon as it makes a mistake and has that mistake in context, it just has no way of correcting course. A lot of my current work is related to that and using a devil’s advocate approach to self-correction.
But in reality, we won’t see a significant jump in things like being able to identify self-ignorance until hardware shifts in the next few years.
-
A child under a certain age usually can’t explain advanced concepts either, so the inability to understand one concept doesn’t preclude understanding of others.
The definition of understanding they use is very shallow compared to how most would define it. Failure to complete a task consistently when numbers are changed, even when they don’t effect the answer shows a lack of real understanding to most. Asking a model the sheet drying question for example will give different results depending on what numbers you use. Better models are better at generalizing but are still far from demonstrating what most consider to be real understanding.
at some point don’t we just load it with every angle we (as humans) have and from there it can derrive not only our best answer but the better one as well? I mean isn’t that when even the shitty version of this ursups is?
A language model can’t determine good from bad because it’s only trained to predict the next token based on what it has seen.
I find this extraordinarily unconvincing. Firstly it’s based on the idea that random graphs are a great model for LLMs because they share a single superficial similarity. That’s not science, that’s poetry. Secondly, the researchers completely misunderstand how LLMs work. The assertion that a sentence could not have appeared in the training set does not prove anything. That’s expected behaviour. “stochastic parrot” wasn’t supposed to mean that it only regurgitates text that it’s already seen, rather that the text is a statistically plausible response to the input text based on very high dimensional feature vectors. Those features definitely could relate to what we think of as meaning or concepts, but they’re meaning or concepts that were inherent in the training material.
Stupid, LLMs do not create new relationships to words that don’t exist.
This is all just fluff to make them seem more like AGI, which they never will be.
I have a theory… They are sophisticated auto-complete.
You are making the common mistake of confusing how they are trained with how they operate.
For example, in the MIT/Harvard Othello-GPT paper I mentioned, feeding in only millions of legal Othello moves into a GPT model (i.e. trained to autocomplete moves) resulted in the neural network internally building a world model of an Othello board - even though it wasn’t explicitly told anything about the board outside of being fed legal moves.
Later, a researcher at DeepMind replicated the work and found it was encoded as a linear representation, which has then since been shown to be how models encode a number of other world models developed from their training corpus (Max Tegmark coauthored two interesting studies in particular about this regarding modeling space and time and modeling truthiness).
They operate by weighting connections between patterns they identify in their training data. They then use statistics to predict outcomes.
I am not particularly surprised that the Othello models built up an internal model of the game as their training data were grid moves. Without loooking into it I’d assume the most efficient way of storing that information was in a grid format with specific nodes weighted to the successful moves. To me that’s less impressive than the LLMs.
They operate by weighting connections between patterns they identify in their training data. They then use statistics to predict outcomes.
Again, this isn’t quite correct. They can do this, but it isn’t the only way they can achieve completion of tokens.
Without loooking into it I’d assume the most efficient way of storing that information was in a grid format with specific nodes weighted to the successful moves.
(It also developed representations of what constituted legal vs non-legal moves.)
You are getting closer to the point. Think about a model asked to complete Pythagorean theorem sequences based on a, b inputs to arrive at c inputs.
What’s the most efficient way to represent that data for successfully completing sequences?
So somewhere in there I’d expect nodes connected to represent the Othello grid. They wouldn’t necessarily be in a grid, just topologically the same graph.
Then I’d expect millions of other weighted connections to represent the moves within the grid including some weightings to prevent illegal moves. All based on mathematics and clever statistical analysis of the training data. If you want to refer to things as tokens then be my guest but it’s all graphs.
If you think I’m getting closer to your point can you just explain it properly? I don’t understand what you think a neural network model is or what you are trying to teach me with Pythag.
If you think I’m getting closer to your point can you just explain it properly?
The most efficient way for a neural network to predict Pythagorean results given inputs would be to reverse engineer a Pythagorean function within itself rather than simply trying to model statistical relationships between inputs and results. To effectively build a world model of Pythagorean calculation.
Training to autocomplete doesn’t mean that the way it achieves this is limited to any one approach or solution, and it would be useful to keep in mind that a neural network of unbounded size can model any possible function.
It wouldn’t reverse engineer anything. It would start by weighting neurons based on it’s training set of Pythagorean triples. Over time this would get tuned to represent Pythag in the form of mathematical graphs.
This is not “understanding” as most people would know it. More like a set of encoded rules.
Furthermore, we presented a method for adapting a symbolic function learner to find mathematical relationships between measured quantities in an unsupervised way. This method does not explicitly label any variate as being an output value, bypassing the assumptions made in standard regression problems about causal relationships. We demonstrated how this method was able to recover rules of geometry from raw data alone. This included the law of sines and the Pythagorean theorem, two relationships on measurements relating to triangles where no variable is necessarily considered an output of a function of the others.
- Panju, Automated Knowledge Discovery Using Neural Networks (2021)
