I imagine an AI creator tool that allows you to specify your desired outcome and provides suggestions or allows you to drag and drop a chain of actions into your macro.
It could even show articles directly from the Keyboard Maestro app or guide you to check certain links and suggestions from others.
For this to work, the AI tool would need to study existing macros and suggestions from others to learn. It’s just a thought, but who knows, maybe it will become a reality one day. 
Maybe, but for now, the free version of ChatGPT is at least aware of Keyboard Maestro and can do some basic stuff...
-rob.
I looked at some AI support in Keyboard Maestro, but I am hesitant to incorporate any of the existing offerings because it seems likely they will either be pulled or become prohibitively expensive, like several other APIs have over the last few years (Twitter, Reddit, etc).
Thanks, I think an AppleScript AI assistant could be useful if implemented.
These LLMs are not AI – just other people's IP chopped up into distributionally profiled tokens, and recombined in ways that sound distributionally plausible, but are in fact utterly unhinged from anything outside language. 'Hallucination' is not a glitch, it's the fundamental architecture of the thing.
The output mess is not in any sense understood by the system that emits it, and in the case of languages where the volume of online material is too small to cook very delicate distributional profiles, the usability is particularly low, and often takes more time to sort out than it's worth. If a lot of low-grade (low-understanding) AppleScript gets generated that way, and is itself redigested into the models, the wells will acquire a heady (and growing) perfume of stagnation.
There is one "AI" that you can rely on for being free, which is Siri. (Siri isn't the best AI in the world, but it could be getting much better later this year.) There is currently a way in the macOS GUI to send a text request to Siri, it's called "Type to Siri" which is available in the system preferences. Doing this allows a macOS user to type a message to Siri and get a typewritten message back. This would be something that KM could rely on, if there was an API for it. My guess is that new APIs for this are coming in June 2024 WWDC.
I could be wrong, but SiriKit appears to be available only for mobile OSs at this time, not macOS. (This may change at the next WWDC.) And it's limited to apps that meet six categories, and it would be a stretch to call KM one of those categories (but there's no harm in applying.) However there might be a workaround. If KM can read the notification messages that appear in macOS, then it could use type to Siri to send a message and read the notification to get the result from Siri.
As it is right now, it is possible to use Type to Siri to send it a question, (either by hand or using KM's Insert Text action) such as "how much is five plus two and a half" and you can get one or more than one result in a pop-up window.
If one was desperate, one could use KM's OCR action to read the result back into a KM macro.
Besides math, it also can respond to general questions, like:
"Who is the prime minister of Canada?"
"Turn on my bedroom lights"
I have a macro that can retrieve spoken content within my house and so I could actually get a KM macro to interact with Siri in this manner. I won't be releasing this macro because it's not 100% reliable. But it would be really great if KM could address either of these techniques directly:
Interesting analysis of the structure and statistics of wanting to "believe" that there are signs of intelligence (beyond purloined IP) in Large Language Model outputs:
While your post isn't entirely untrue, I find it a bit uncharitable. I'm often amazed by the quality of LLM output, even in domains like programming and math where it's made me much more productive. It's also good for coming up with new ideas and discovering new things, something that search engines used to be good at but no longer are. Sometimes the output is not so good, but on the whole I've really been impressed with modern AI. 
Well played. Also this isn’t a race, look for tools that give us a useful edge. I don’t need another way to get ChatGPT to spit back copyrighted content. There are already hundreds of ways to do that. (Few of them are useful(.
I would love if something existed to make automating small tasks easier. I suspect this is a few years away. I’m patient and will pay for the upgrade when it arrives.
Those kinds of "codified" domains are exactly where LLMs should shine, especially when they have a large corpus to draw from. They're less good in similar domains with a smaller body of work, as @griffman's ChatGPT output, above, shows -- the answer isn't wrong, but it isn't right either!
What they can do is certainly impressive. Whether they continue to improve as they have will, to an extent, depend on whether they can filter out the cruft when they ingest more data -- cruft that's only increasing as they generate data themselves.
Meanwhile, Keep it Shot (AI labeling screenshots and PDFs) https://keepitshot.com and Cotypist (AI offering writing suggestions as you write) https://cotypist.app are pretty well implemented current AI offerings I'm enjoying that might scratch the itch a bit.
I hope not!
