A well-known app starts an experiment. I’ll explain it in my own words. This could be inaccurate. But it works roughly like this.
Instead of integrating an AI into the app, the app is integrated into the AI. The AI gets all the necessary information / specifications, etc. about this app to make the AI a super user.
Under ideal circumstances, and applied to DT, this means that the AI can do everything.
AI, create me an smart group/rule that does this and that.
AI, can I change this and that so and so in the settings? Yes? Then do it.
I’m interested in your opinion on this. Would that be a useful AI addition for DT?
While not a crazy idea, it’s certainly one that gives AI far more control than I would trust. It also would not be something that would run locally in a performant way.
And in the end, you also learn nothing about the application itself. So what do you do when AI isn’t working correctly (or at all)?
This is all hypothetical, but to create a complicated smart group with many conditions, the AI does not need personal data from a DT database. And once the smart group has been created, the AI is no longer needed.
That’s true, but I can’t fix my own vacuum cleaner either, and I don’t want to learn to do it if I don’t have to.
What everyone is already doing when the AI doesn’t work.
To create a smart group with many conditions, you need to have a clear idea that you can communicate. If you have that you can create the smart group yourself. If you don’t have a clear idea, what are you going to tell the AI?
That was a random example. And I don’t think you can do everything yourself if you have a precise idea. And even then, the AI would probably be faster.
As I said, this is all hypothetical. I would tell the AI the same thing I would tell you. I would try to explain what I want. If you don’t understand, you would ask. That’s exactly what the AI would do.
In any case, the current AI possibilities are now being used, for example, to write summaries of something. The AI summarizes what the “client” has never read. I’d rather have a smart group built for me. That seems much more sensible to me.
I tried to make the point that one has to be able to describe the desired result as exactly as possible. It’s less about
but more about “if I can’t describe it”.
Which is something else entirely. The software goes over existing text and runs its probabilistic magic on it. If you’re lucky, it doesn’t hallucinate anything.
Having something new created, is (I think) more demanding because the requirements must be clear.
Recently, someone here described how they needed several iterations with an AI just to create a trivial smart rule. So yes – it works. But it’s not necessarily faster than a human who knows what they want and what they are doing.
Do not forget, in this “experiment” the AI is much smarter than in “real life”. The developers would share every detail about their app. Ideally, a DT specialist would emerge.
Sure, if I can’t describe it at all, then …
But a normally intelligent person can always describe something, even if it’s vague.
And then the specially trained AI would ask, just like a human would … maybe.
The point is that AI doesn’t know what I want to summarize. Users often don’t want “summaries” at all. They want to know what is most important to them in a text. The AI knows that even less. And the users can’t control it because they haven’t read the text.
The thing is that people can’t simply describe every aspect of the app they wrote. We’ve had that in another thread: VW tried to grab their mechanics’ knowledge and experience to create an expert system that should facilitate car repairs.
Didn’t work, because the programmers didn’t know what to ask and the mechanics didn’t know how to put their experience into words.
That’s why I suggested to have the AI read the app code. Then it would just know every detail.
If the AI can do everything, than there is no need for smart groups/rules or the app you would just have to ask big brother and everything will be fine
There might not be a reason to have the human operator
At some point, either you or the AI is going to have to test the rule against real data. At that point, either the AI runs into the limitations of your local hardware, or your data leaks out to the AI’s “cloud.”
Indeed - I have a working version of an integration of MCP with Keyboard Maestro. You can use Claude Desktop or Cursor to execute any Keyboard Maestro macro, including but not limited to Devonthink Applescripts.
I do not recomend leaving it to itself i.e. Hal or The Entity. But for specific focused tasks which you define it can be quite helpful.
Using GenAI without knowing a few key things renders it useless:
How does it work?
What is it good at?
What is it not very good at?
What are you trying to achieve? and wlll AI help?
I’m going to link to two recent blog posts of my own. (Not an attempt to sell).
I link to the blog posts because I think they illustrate the strenghts and weaknesses of these tools. In addition, you might adapt my rough guidelines to you own use:
But I guess OP may already know much about DT as a user and what is his intention to add/change/create within DT as it relates to his ideas how to organise things in DT. (or at least it is far from “car repairs” example).
Then precisely describing what is needed may be very easy and left all hard, detailed work to AI.
This is the exact area where AI absolutely excels.
For example. Sometimes I write bash scripts but it is so cryptic!!! It requires you be either seasoned expert in bash or put in A LOT of effort.
Nowadays I describe AI what I need and what extra qualities I require (log everything very nicely, check every possible boundary conditions to make the script bulletproof) and it WILL magically do anything for me. I can verify the script and assess quality of it, BTW learning something about it having fun and invest minimal mental energy.
Eventually I get quality I could never achieve myself because it is so much other stuff in life to cover. Never had time and energy to write yet another script. Now I can write many of them effortlessly.
ach, @chrillek , I’m sorry.
Possibly you meant that AI would have to have access to source code of DT to understand what’s possible?
Maybe yes, maybe no.
I think someone could map DT entities and operations into kind of a facade.
Then an AI agent is needed as bridge between AI model and DT and then you would be able to say “based on all available operations in DT you have here … please configure DT to achieve this …”.
Using AI in a particular app is obvious to everyone and has been discussed frequently in this forum and elsewhere.
Thinking about the reverse is much more difficult and was also new to me. What if you integrate an app into the AI?
The developer who is doing this experiment with his app says this evolution is inevitable. I can’t judge that. He will hardly become superfluous because someone will have to legally develop the app further. The AI can’t do that. But the AI can help the user to use what is already there.
I’m not doing either for the time being. But without knowing all the consequences, I like the second option better … maybe
@FrankT , inevitable indeed!
Many try these days as this is hot as Web was once. Competition grows.
Example app:
This is an app then can connect and automate two sides:
Query any AI models for answers
And actually call “actions” in configured applications to basically DO things in the apps.
Maybe it is not obvious from their landing page but the paragraph about “Connect to Make & Zapier” says that is capable of executing tasks in other apps and services.
And here is how to configure an action:
).
