Related terms. Even though an answer generated by an LLM is most likely wrong, and definitely can't be taken at face value, the words and phrases used in that answer can be exactly what you need to create a search query that you wouldn't be able to otherwise.
Also, thesaurus isn't a good tool for exploring an unknown problem domain. It gives synonyms, not related terms. LLMs let you input a layman description of your problem, and get an answer that's using correct domain terms and phrases (even if using them incorrectly).
I imagine Google will add such a function. They probably tried already - I've heard that current search is already powered by ML models to a degree.
A Z80 routine can call outside of its 16-bit address space with 'CALL.IL'. This pushes the 16-bit return address onto the 16-bit stack, switches to 24-bit mode, and pushes the magic 'return to 16-bit mode' number to the 24-bit stack. However, it's not clear from the manuals or datasheets what happens if you use the prefixed 'CALL.IL' opcode sequence when the 'MADL' bit is reset.
I asked ChatGPT, because this is something that Google searching hasn't yielded answers for. It had this to say:
"The MADL bit (short for Memory Access During Interrupts Low) is a flag in the Interrupt Control Register that determines whether or not interrupt service routines (ISRs) can access low memory (addresses 0000h-3FFFh) during interrupts."
Plus some more stuff building on that, on CALL.IL being about ISRs, and about low memory. All of it is completely, fundamentally wrong. I did a handful of rounds of trying to steer it to a more correct answer but it continued to get additional basic facts wrong and would lean back to earlier incorrect facts as others conflicted with its answers.
I asked it another question I have, this time about the UART on the CPU. There is a Receive Buffer Register (UARTx_RBR) that contains the head of the receive FIFO. The documentation does not make it clear what is in the RBR if the FIFO is empty, so I asked ChatGPT. It told me a very plausible answer, the one I suspect myself, which is that it'll keep returning the same value until new data is available. But then it went on to tell me this is called receiver overrun, and described how an overrun occurs, including noting that it happens when the FIFO is full. And we went round in circles on this for a little while.
ChatGPT is a major step forwards in our post-truth existence: its answers are an amalgam of the most frequently repeated views on a topic, not those with stronger reasoning or more effective evidence to support. If there is little or no source data on a topic (as would be the case with my very specific questions on a rarely used processor) LLMs are (presently?) unable to detect that they are responding to a topic with limited contextual information and tailor their responses accordingly, and instead confidently provide utter nonsense.
I trust ChatGPT to do things that LLMs are good at, though: if I give it some bullet points and some style guidance it can give me written paragraphs. If I ask it to rephrase a well known song in the style of some modern artist I'll get something back that's pretty plausible. It can give me some starting points for learning more about some well known topic, even.
I would definitely not trust it _at all_ to give me something factual like part numbers of uncommon ICs, because LLMs cannot distinguish between fact and fiction, not in what they ingest, and not in what they produce.
And that’s the worrying thing.
And of human communication. Look at most of what main stream online and offline media are blasting and people consuming.
Do people watching reality shows really care about truth? Or authoritative answers? And that’s probably most people for most things.
In this context, this shows that LLM cannot be used for Search of novel and technical stuff the way you are doing it. It still has to be fine-tuned for a market who wants to know more about your kind of stuff.
Or complete bullshit. As long as the problem of AI hallucinations remain unsolved, I can't trust AI like ChatGPT - at least Google will tell you if it has no idea what you are talking about.
That's interesting, what features of LLM/ChatGPT architecture are likely to drive this?
The point is they don't know the answer, they just come up with something.
It would be best if all ChatGPT replies started with "I really don't know the answer, but some people have at some point written something like this: "...". I don't remember who my sources are, but trust me.
Thing is, many people - probably the majority - work on just that. They aren't looking for answers to challenging issues like your question where 'Google searching hasn't yielded answers for', they are looking for answers to questions where google and stack overflow does have thousands of results for similar, potentially related scenarios, and want something to summarize or filter it into an usable answer, and ChatGPT provides that option. When the official documentation provides all the information in a poor format so you can't just search for the answer, ChatGPT can extract an answer from it. Not "some starting points for learning more about some well known topic" as you say, but rather some "digested, complete, specific result from a well known topic to avoid having to learn learning anything more than strictly necessary for the outcome".