Because those "dumb" assistants were designed and programmed by humans to solve specific goals. The new "smart" chatbots just say whatever they're going to say based on their training data (which is just scraped wholesale, and is too enormous to be meaningfully curated) so they can only have their behavior adjusted very indirectly.
I continue to be amazed that as powerful as these language models are, the only thing people seem to want to use them for is "predict the most plausible output token that follows a given input", instead of as a human-friendly input/output stage for a more rigorously controllable system. We have mountains of evidence that LLMs on their own (at least in their current state) can't reliably do things that involve logical reasoning, so why continue trying to force a round peg into a square hole?
I’ve asked ChatGPT write over a dozen Python scripts where it had to have an understanding of both the Python language and the AWS SDK (boto3). It got it right 99% of the time and I know it just didn’t copy and paste my exact requirements from something it found on the web.
I would ask it to make slight changes and it would.
There is no reason with just a little human curation it couldn’t delegate certain answers to third party APIS like the dumb assistants do.
However LLMs are good at logical reasoning. It can solve many word problems and I am repeatedly amazed how well it can spit out code if it knows the domain well based on vague requirements.
Or another simple word problem I gave it.
“I have a credit card with a $250 annual fee. I get 4 membership reward points for every dollar I spend on groceries. A membership reward point is worth 1.4 cents. How much would I need to spend on groceries to break even?”
It answered that correctly and told me how it derived the answer. There are so many concepts it would have to intuitively understand to solve that problem.
I purposefully mixed up dollars and cents and used the term “break even” and didn’t say “over the year” when I referred to “how much would I need to spend”
All of the “dumb” assistants can recognize certain questions and then call APIs where they can get accurate up to date information.