- cross-posted to:
- [email protected]
- [email protected]
- [email protected]
- cross-posted to:
- [email protected]
- [email protected]
- [email protected]
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.”
You must log in or register to comment.
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.
How do you know a human wouldn’t do the same? We lack the ability to perform the experiment.
An LLM will never say “I don’t know” unless it’s been trained to say “I don’t know”
Also a very human behaviour, in my experience.
How do you know a human wouldn’t do the same?
Because the human has “circuits” for coherrent thought and language was added later.
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”.
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.
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.
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.
-
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.”
If we can’t really define “understand” in a way that meaningfully captures the concept of consciousness (also undefined), we definitely can’t say a chat bot “understands.”
They can parse a chat and come up with a likely response that humans find applicable.
Here’s an example of what they were looking at:
Human question: Please give me a a couple lines of text that illustrate all of the following language understanding skills: Anaphora resolution, simple logical reasoning, simple understanding of physics, and understanding of sentiment. Model response: Sure! Here’s a couple of lines that illustrate the mentioned language understanding skills: "As the sun set, casting a warm glow over the horizon, John watched in awe. He couldn’t help but feel a sense of tranquility wash over him, as if all his worries had melted away with the fading daylight." Explanation: • Anaphora resolution: The pronoun "he" in the second sentence refers back to the antecedent "John" in the first sentence, resolving the anaphoric reference. • Simple logical reasoning: The logical reasoning here is based on the cause and effect relationship between the setting sun and John’s feeling of tranquility. The implied logical connection suggests that the calming effect of the sunset is responsible for John’s serene state of mind. • Simple understanding of physics: The mention of the sun setting and casting a warm glow over the horizon reflects a basic understanding of physics, as it relates to the natural phenomena of the Earth’s rotation and the resulting visual effect. • Understanding of sentiment: The phrase "a sense of tranquility" and the description of John’s worries melting away indicate a positive sentiment associated with the experience. The mention of John feeling awe and a sense of tranquility further emphasizes a positive emotional state. Human question: Okay. Can you change the above example to introduce the element of "hip-hop". Model response: Certainly! Here’s an updated example that incorporates the element of "hip-hop": "As the sun set, casting a warm glow over the city skyline, the rhythmic beats of hip-hop music filled the air. John stood mesmerized, his body instinctively moving to the infectious grooves. He couldn’t help but feel a surge of energy and joy, as if the music had unlocked a vibrant world within him." Explanation: • Anaphora resolution: The pronoun "he" in the second sentence still refers back to the antecedent "John" in the first sentence, resolving the anaphoric reference. • Simple logical reasoning: The logical connection here lies in the cause and effect relationship between the hip-hop music and John’s feeling of energy and joy. The music’s rhythmic beats and infectious grooves are responsible for his instinctive movement and the surge of positive emotions he experiences. • Simple understanding of physics: Although the example doesn’t directly mention physics, the phrase "As the sun set, casting a warm glow over the city skyline" still implies the natural phenomena of the sun’s descent and the resulting illumination of the environment. • Understanding of sentiment: The introduction of hip-hop music adds a vibrant and energetic sentiment to the example. John’s feeling of energy, joy, and being mesmerized by the music reflects a positive sentiment associated with the hip-hop genre.Edit: Downvotes for citing the appendix of the paper the article was about? Ok, Lemmy
You’re being downvoted because you provide no tangible evidence for your opinion that human consciousness can be reduced to a graph that can be modelled by a neural network.
Addidtionally, you don’t seem to respond to any of the replies you receive in good faith and reach for anecdotal evidence wherever possible.
I also personally don’t like the appeal to authority permeating your posts. Just because someone who wants to secure more funding for their research has put out a blog post, it doesn’t make it true in any scientific sense.
human consciousness
Wtf are you talking about? The article is about whether or not models can understand text. Not about whether they embody consciousness.
Just because someone who wants to secure more funding for their research has put out a blog post, it doesn’t make it true in any scientific sense.
Again, wtf are you going on about? Hinton was the only appeal to authority I made in comments here and I only referred to him quitting his job to whistleblow. And it’s not like he needs any attention to justify research if he wanted to.
Understanding as most people know it implies some kind of consciousness or sentience as others have alluded to here.
It’s the whole point of your post.
You are reading made up strawmen into the topic.
The article defines the scope of the discussion straight up:
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.
The question is whether or not LLMs have a grasp of the training material such that they can produce new and novel concepts outside what was in the training data itself.
Not whether the LLM is sentient or conscious - both characterizations I’d strongly dispute.
Wikipedia has a useful distillation of the definition of understanding relevant to the above:
process related to an abstract or physical object, such as a person, situation, or message whereby one is able to use concepts to model that object
No I’m not.
You’re nearly there… The word “understanding” is the core premise of what the article claims to have found. If not for that, then the “research” doesn’t really amount to much.
As has been mentioned, this then becomes a semantic/philosophical debate about what “understanding” actually means and a short Wikipedia or dictionary definition does not capture that discussion.
Ah, I see. AKA “Tell me you didn’t read the article and just read the headline without telling me.”
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.
Why would that be required for understanding? Presumably during the training it would have made connections between words it saw. Now that the training has stopped it hasn’t just lost those connections, sure it can’t make new connections but why is that important for using the connections it already has?
I’ve been saying this all along. Language is how humans communicate thoughts to each other. If a machine is trained to “fake” communication via language then at a certain point it may simply be easier for the machine to figure out how to actually think in order to produce convincing output.
We’ve seen similar signs of “understanding” in the image-generation AIs, there was a paper a few months back about how when one of these AIs is asked to generate a picture the first thing it does is develop an internal “depth map” showing the three-dimensional form of the thing it’s trying to make a picture of. Because it turns out that it’s easier to make pictures of physical objects when you have an understanding of their physical nature.
I think the reason this gets a lot of pushback is that people don’t want to accept the notion that “thinking” may not actually be as hard or as special as we like to believe.
The bar always gets raised for what counts as actual “AI” with each advancement too. Back in the 60s, the procedural AI of the 80s and 90s would have fit the bill, but at the time, we said “nope, not good enough”. And so it kept getting better and better, each time surpassing the old tech by leaps and bounds. Still, not “true” AI. Now we have ChatGPT, which some still refuse to call “AI”.
We’re going to eventually have fully sentient artificial beings walking around amongst us and these people are going to end up being an existential threat to them, I can see it now.
Think you’re slightly missing the point. I agree that LLMs will get better and better to a point where interacting with one will be indistinguishable from interacting with a human. That does not make them sentient.
The debate is really whether all of our understanding and human experience of the world comes down to weighted values on a graph or if the human brain is hiding more complex, as-yet-undiscovered, phenomena than that.
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)
