They extract concepts from their training data and can combine concepts to produce output that isn't part of their training set, but they do require those concepts to be in their training data. So you can ask them to make a picture of your favorite character fighting mecha on an alien planet and it will produce a new image, as long as your favorite character is in their training set. But the extent it imagines an alien planet or what counts as mecha is limited by the input it is trained on, which is where a human artist can provide much more creativity.
You can also expand it by adding in more concepts to better specify things. For example you can specify the mecha look like alphabet characters while the alien planet expresses the randomness of prime numbers and that might influence the AI to produce a more unique image as you are now getting into really weird combinations of concepts (and combinations that might actually make no sense if you think too much about them), but you also greatly increase the chance of getting trash output as the AI can no longer map the feature space back to an image that mirrors anything like what a human would interpret as having a similar feature space.
The paper that coined the term "stochastic parrots" would not agree with the claim that LLMs are "unable to produce a response that isn't in their training data". And the research has advanced a _long_ way since then.
[1]: Bender, Emily M., et al. "On the dangers of stochastic parrots: Can language models be too big?." Proceedings of the 2021 ACM conference on fairness, accountability, and transparency. 2021.
/facepalm. Woosh indeed. Can I blame pronoun confusion? (Not to mention this misunderstanding kicked off a farcically unproductive ensuing discussion.)
When combined with intellectual honesty and curiosity, the best LLMs can be powerful tools for checking argumentation. (I personally recommend Claude 3.5 Sonnet.) I pasted in the conversation history and here is what it said:
> Their position is falsifiable through simple examples: LLMs can perform arithmetic on numbers that weren't in training data, compose responses about current events post-training, and generate novel combinations of ideas.
Spot on. It would take a lot of editing for me to speak as concisely and accurately!
> you can try to convince all you want, but you're just grasping at straws.
After coming back to this to see how the conversation has evolved (it hasn't), I offer this guess: the problem isn't at the object level (i.e. what ML research has to say on this) nor my willingness to engage. A key factor seems to a lack of interest on the other end of the conversation.
Most importantly, I'm happy to learn and/or be shown to be mistaken.
Based on my study (not at the Ph.D. level but still quite intensive), I am confident the comment above is both wrong and poorly framed. Why? Seeing phrases "incapable of thought" and "stochastic parrots" are red flags to me. In my experience, people that study LLM systems are wary of using such brash phrases. They tend to move the conservation away from understanding towards combativeness and/or confusion.
Being this direct might sound brusque and/or unpersuasive. My top concern at this point, not knowing you, is that you might not prioritize learning and careful discussion. If you want to continue discussing, here is what I suggest:
First, are you familiar with the double-crux technique? If not, the CFAR page is a good start.
Second, please share three papers (or high-quality writing from experts): one that supports your claim, one that opposes it, and one that attempts to synthesize.
I'll try again... Can you (or anyone) define "thought" in way that is helpful?
Some other intelligent social animals have slightly different brains, and it seems very likely they "think" as well. Do we want to define "thinking" in some relative manner?
Say you pick a definition requiring an isomorphism to thoughts as generated by a human brain. Then, by definition, you can't have thoughts unless you prove the isomorphism. How are you going to do that? Inspection? In theory, some suitable emulation of a brain is needed. You might get close with whole-brain emulation. But how do you know when your emulation is good enough? What level of detail is sufficient?
What kinds of definitions of "thought" remains?
Perhaps something related to consciousness? Where is this kind of definition going to get us? Talking about consciousness is hard.
Anil Seth (and others) talks about consciousness better than most, for what it is worth -- he does it by getting more detailed and specific. See also: integrated information theory.
By writing at some length, I hope to show that using loose sketches of concepts using words such as "thoughts" or "thinking" doesn't advance a substantive conversation. More depth is needed.
Meta: To advance the conversation, it takes time to elaborate and engage. It isn't easy. An easier way out is pressing the down triangle, but that is too often meager and fleeting protection for a brittle ego and/or a fixated level of understanding.
Sometimes, I get this absolute stroke of brilliance for this idea of a thing I want to make and it's gonna make me super rich, and then I go on Google, and find out that there's already been a Kickstarter for it and it's been successful, and it's now a product I can just buy.
No, but then again you're not paying me $20 per month while I pretend I have absolute knowledge.
You can, however, get the same human experience by contracting a consulting company that will bill you $20 000 per month and lie to you about having absolute knowledge.
