NokiMo
robertskmiles
robertskmiles

patreon


Computerphile Video: ChatGPT

I made a new video with Computerphile, about ChatGPT!

https://www.youtube.com/watch?v=viJt_DXTfwA

Language Models, Simulacra, Reinforcement Learning from Human Feedback, Scaling and Inverse Scaling, Sycophancy and Deception and Instrumental Convergence!

Computerphile Video: ChatGPT

Comments

Thanks for your response. Much appreciated. I’m not sure I understand all of it, but after reading several times, I’m assuming the way I’m trying to keep all this simple, so I can explain it to my kids and some friends is not totally off. Thanks again. Simplifying without being completely wrong is not easy. Especially, when I barely know what I’m talking about. :) The little I’ve read from academic research about LLMs was enough to convince me that I should stick to Youtube tutorials. On finetuning, you might be interested in what David Shapiro says about it. It starts at 7:50 of his presentation. This was posted four months ago, so might be already obsolete. :). https://youtu.be/9qq6HTr7Ocw

To add to what Boris said: "Prompt engineering" vectors towards terms like "clarity of composition, relevance of words, completeness, consistency in style, simplicity of composition and variety of examples." It was something that I targeted in my first two weeks of working with LLMs (ChatGPT) and working to build my understanding of the strengths and weaknesses of particular prompts, and how to use that knowledge to achieve token economy. To support the other things he said: RE: 4 - Fine tuning is, as I understand it, training the AI with additional domain specific examples. Variety of examples can be very helpful in reducing the token requirements for achieving desirable output consistently. RE: 3 - Self reflection papers have been written on how the model is forward-only text prediction, and feedback back its text prediction into reflection can improve the overall quality of the final output. (e.g., Just because it is statistically probable that a human will make a logic mistake, does not mean that the logic mistake is desirable output. Self reflection encourages a model to self assess its own output for desirability and to refine it output based on training.) RE: 1 - A statistical prediction engine is not thinking. ChatGPT lacks several other attributes commonly associated with intelligence. It has no executive function. Projects like AutoGPT want to mitigate some of these deficiencies. Someone I do not respect suggested that ChatGPT is little more than a stochastic parrot. While the analogy has flaws, I find it amusing enough that I like to share the idiom.


Related Creators