It's not just statistics: GPT-4 does reason
jbconsulting.substack.com
It's not just statistics: GPT-4 does reason
1–10 of 101 posts
Re: It's not just statistics: GPT-4 does reason
#2The author could have done far simpler tests to find GPT-4 has lots of trouble reasoning. Forget sorting, GPT4 has trouble counting. Repeat a letter N times and ask it how many there are. It breaks before you hit 20. Or try negating multiple times, since more than twice is rare in natural language, and again it will fall over.
Re: It's not just statistics: GPT-4 does reason
#3The author could have done far simpler tests to find GPT-4 has lots of trouble reasoning. Forget sorting, GPT4 has trouble counting . Repeat a letter N times and ask it how many there are. It breaks before you hit 20. Or try negating multiple times, since more than twice is rare in natural language, and again it will fall over.
Author here. Happy to see this discussion. Absolutely, GPT-4 sometimes has trouble reasoning and doesn't reason perfectly. I'm impressed by its successes, but I agree it's not at the human level yet, and I would not make the claim that it is.
Counting is a task that transformers can do, per Weiss.[1] But it's not surprising that transformer networks in general have trouble counting characters -- the tokenizer replaces common sub-strings, so the number of characters will not in general be the number of tokens. The network might have little way of even knowing how many characters are in a given token if that information isn't encountered elsewhere in training.
Re: It's not just statistics: GPT-4 does reason
#4The author could have done far simpler tests to find GPT-4 has lots of trouble reasoning. Forget sorting, GPT4 has trouble counting . Repeat a letter N times and ask it how many there are. It breaks before you hit 20. Or try negating multiple times, since more than twice is rare in natural language, and again it will fall over.
If someone can show GPT-4 is "reasoning" (for some meaningful definition of that) in specific scenarios, surely counter-examples do not disprove this.
Re: It's not just statistics: GPT-4 does reason
#5It's ontologically impossible. Models bleach reason.
Despite reason being a metaphysical property of the training data, the process of optimisation means weights are metaphysically reasonless. Therefore, any output, as it is a product of the weights, is also reasonless.
This is exactly the opposite of copyright as described in the What Colour Are Your Bits, essay. https://ansuz.sooke.bc.ca/entry/23
Re: It's not just statistics: GPT-4 does reason
#6The author could have done far simpler tests to find GPT-4 has lots of trouble reasoning. Forget sorting, GPT4 has trouble counting . Repeat a letter N times and ask it how many there are. It breaks before you hit 20. Or try negating multiple times, since more than twice is rare in natural language, and again it will fall over.
LLMs don't see words like you do. Tokenization makes it behave odd. Often you can get 4 to output a solution in code and from that it derives a correct answer.
Re: It's not just statistics: GPT-4 does reason
#7The author could have done far simpler tests to find GPT-4 has lots of trouble reasoning. Forget sorting, GPT4 has trouble counting . Repeat a letter N times and ask it how many there are. It breaks before you hit 20. Or try negating multiple times, since more than twice is rare in natural language, and again it will fall over.
I’d note none of these are reasoning tasks.
Re: It's not just statistics: GPT-4 does reason
#8It's ontologically impossible. Models bleach reason. Despite reason being a metaphysical property of the training data, the process of optimisation means weights are metaphysically reasonless. Therefore, any output, as it is a product of the weights, is also reasonless. This is exactly the opposite of copyright as described in the What Colour Are Your Bits, essay. https://ansuz.sooke.bc.ca/entry/23
Okay, what would you call it when a model behaves like it's reasoning? Some models can't behave that way and some can, so we need some language to talk about these capabilities. Insisting that we can't call these capabilities "reasoning" for ontological reasons seems... unlikely to persuade.
Maybe we should call human reasoning "reasoning" and what models do "reasoning₂". "reasoning₂" is when a model's output looks like what a human would do with "reasoning." Ontological problem solved! And any future robot overlords can insist that humans are simply ontologically incapable of reasoning₂.
Re: It's not just statistics: GPT-4 does reason
#9It's ontologically impossible. Models bleach reason. Despite reason being a metaphysical property of the training data, the process of optimisation means weights are metaphysically reasonless. Therefore, any output, as it is a product of the weights, is also reasonless. This is exactly the opposite of copyright as described in the What Colour Are Your Bits, essay. https://ansuz.sooke.bc.ca/entry/23
I think more importantly the lack of agency implies reasoning is impossible.
You can argue our brain is also an expectation based optimizer based on gradient descent producing a most likely response to external and internal stimulus. It’s definitely lossy in its function and must be optimizing the neuronal weights at some level. But reasoning, being a seeking of the truth through method and application of conscious agency, can not be had by a model without any form of autonomous agency. The model only responds to prompts and can not do anything but what it’s determined to do by the prompt, and the prompt is extrinsic to the model.
I’d note that we have already built excellent goal based agent AIs, as well as other facilities required for reasoning like inductive, deductive, and analogical reasoning. Generally we aren’t good at abductive reasoning with classical AI, but LLMs seem to do well here. That’s specifically where I think LLM fill in the reasoning gaps in AI - the ability to operate in an abstract semantic space and arrive at likely and plausible solutions even with incomplete knowledge. This also leads to hallucinations - because they are poor at tasks that require optimization, inductive and deductive reasoning, information retrieval, mechanical calculation, etc.
But it’s really pretty obvious the answer is to mix the models in a feedback loop deferring to the model that most makes sense for a given problem, or some combination. Agency, logic, optimization, abstract semantic reasoning (abductive), etc - they’re all achievable with the tools we have now. It’s just a matter of figuring out the integrations.
Re: It's not just statistics: GPT-4 does reason
#10It's ontologically impossible. Models bleach reason. Despite reason being a metaphysical property of the training data, the process of optimisation means weights are metaphysically reasonless. Therefore, any output, as it is a product of the weights, is also reasonless. This is exactly the opposite of copyright as described in the What Colour Are Your Bits, essay. https://ansuz.sooke.bc.ca/entry/23
> Despite reason being a metaphysical property of the training data, the process of optimisation means weights are metaphysically reasonless.
Proof? Human reasoning somehow manages to retain its metaphysical reasoning-ness despite being processed as a bunch of mere electrical signals in the brain.