For sure, I asked chatGPT to generate code for a certain embedded device to interface with a specific pH sensor and the result was totally bananas.

There is no reason to believe that niche and bespoke requirements will be swallowed up in an efficient, reliable, repeatable package in the very near future. It can be taught, and will be, but I don't think it'll be too snappy.

> immensely pump their productivity, more with less and all

The more I use chatGPT, the more I think this is the case. There are too many loose ends to tie, too many semantics to clarify and realign, and so on. It's immensely powerful today, but perhaps we sometimes forget how much we've learned and how far we've come.

I felt a little doomed on my first go with chatGPT, but now the shock has worn off. I'm optimistic that it'll be a useful tool. The more I play with it, the better I get at prompting it, the more I think alright – I can actually see this being useful one day. It might actually be a useful way to rubber duck problems, generate stubs, review rudimentary implementations for improvements, etc.

Given its current limitations and how deep they are entrenched by limitations of its design, I suspect it will require a significant breakthrough to overcome them and truly replace people.

Try GPT3 with the code specific model maybe ?

Asking chatGPT for code is like asking a cook to solve a math equation

With better known programming problems it actually does a great job.

This is where my mind has been as well - I feel we're heading towards a future that will benefit small business - but I don't want to be overly optimistic.