2 Comments

Thanks for sharing - it is enlightening to understand that the system is itself driven by a giant prompt. As a guy who did assembler in the 90s I'm quite familiar with the "memory map" that you showed. In that case you needed technical skills and you knew exactly what you were doing (to the faulty comma) ... now it's plain English which is way easier but much less accurate (I guess as accurate as instructions you give to an employee - there is always room for misunderstanding or interpretation).

I did not know that it would be different when using the API. The API must have it's own restrictions too I suppose - I suspect that if you'd ask a list of deadly pathogens you would not get it, but you would get song lyrics or a cooking recipe? That's interesting :-)

Expand full comment

I gave up entirely on using the current gen. AI applications. I created a children's book series that teaches kids and their clever parents about AI in the most entertaining and funny way (www.romyandroby.ai). I tried to use image programs to illustrate it. I abandoned the idea and used the old pens and pencils before passing my drafts to a graphic designer. A friend of mine is an interior designer, and she claimed she can't use any of the AIs today. The same is true with writing poetry. I can look for a rhyme, but the generated text is dull, flat, and useless. Gen AI models use tons of energy to perform the tasks. They hallucinate because they fail to master distribution shift - like every deep learning application. In this context, I wonder whether the architecture based on deep learning is a path to take after all.

Expand full comment