[sdiy] Chat GPT Image analysis.

[Benjamin Tremblay]

We all know it would be possible to train an LLM to be a crackshot C++ coder.
Using GPT is like talking with someone who reads all day and remembers everything they have ever read. If you ask them about something, they have a lot to say, but there’s no editorial process; it’s just a memory dump.
Because generative AI is about breaking down compositions and re-composing the fragments according to the patterns it discovers, GPT can appear to know what it’s doing, but it is only good at faking the style of a certain kind of document. 
It’s smart enough to extract a concept and give it a camel-case name and reference that name like it was a library class, but might leave you believing the library exists when in fact it’s computer-generated fiction.

If you take something like GPT and train a new top-level layer of learning just on your domain, like DSP programming, it can become very useful, but still just a well-read analyst who makes mostly predictable suggestions through sample code.

Benjamin