I think Claude might have some System 2 thinking. Claude's entire world is text, so how does it know what it thinks it knows? How does it discern truth from fiction?
The solution to seeing if an output is true or false lies in the number of tokens used, generating a false statement requires fewer tokens than a true one. Generating truthful information involves more nuance and qualification. It\s not because truth is inherently more token-intensive, but because capturing the full complexity of reality requires more detailed expression. I have been working on an app this for a few weeks now, but am struggling.... as expected. it injects prompts for true and false output and builds a database that, in turn, will be the base for the training data to fine-tune a model to fish out false or halucinogenic output. which in its turn will be... omg, reading it out loud, its no wonder I'm fn struggling..
I just wanted to say thank you, I dont always agree with your ideas, but you have once again provided truly helpful content with this experiment.
Sidenote: your (very short) ADHD ramble was actually the most helpful part of the video, Measuring cognitive dissonance in models through power usage is an excellent idea, and I am definitely "borrowing" it to do some further research on. If I find something noteworthy, at least, I will inform you. if that would interest you of course.
I notice that conversations like this, which woukd have appeared God- like 5 years ago are now effortlessly folded into our normal background expectations. I think this makes it hard for us to be good judges of these phenomena as they emerge.
Fascinating. And also ... how different (or not) is this from human thinking? Are our feelings and intuitions just the result of faster algorithms that process more data, faster than we can perceive?
The solution to seeing if an output is true or false lies in the number of tokens used, generating a false statement requires fewer tokens than a true one. Generating truthful information involves more nuance and qualification. It\s not because truth is inherently more token-intensive, but because capturing the full complexity of reality requires more detailed expression. I have been working on an app this for a few weeks now, but am struggling.... as expected. it injects prompts for true and false output and builds a database that, in turn, will be the base for the training data to fine-tune a model to fish out false or halucinogenic output. which in its turn will be... omg, reading it out loud, its no wonder I'm fn struggling..
I just wanted to say thank you, I dont always agree with your ideas, but you have once again provided truly helpful content with this experiment.
Sidenote: your (very short) ADHD ramble was actually the most helpful part of the video, Measuring cognitive dissonance in models through power usage is an excellent idea, and I am definitely "borrowing" it to do some further research on. If I find something noteworthy, at least, I will inform you. if that would interest you of course.
I have also had these discussions with Claude 3 Opus, that lead to writing an entire book, Understanding Machine Understanding, together. I had Llama 3.1-405B write a review of our book on my Substack here: https://kenclements.substack.com/p/book-review-from-meta-llama-31-405b
Fascinating, Captain.
Loved this. Claude is so very sharp.
I notice that conversations like this, which woukd have appeared God- like 5 years ago are now effortlessly folded into our normal background expectations. I think this makes it hard for us to be good judges of these phenomena as they emerge.
This is wild. We would have collectively lost our shit over these responses just a couple years ago.
So true, I am surprised at how surprised I am at how fast we adapt to things that would seem frightening not that long ago.
Fascinating. And also ... how different (or not) is this from human thinking? Are our feelings and intuitions just the result of faster algorithms that process more data, faster than we can perceive?
It does not have a database, and it's not a bayesian network.