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.’
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.
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)
New theory wrong.
