So far everything has been reactive. You need to engage a prompt, you need to ask Siri or ask claude to do something. It can be very powerful once prompted, but it still requires prompting.
You always need to ask. Having something always waiting in the background that can proactively take actions and get your attention is a genuine game-changer.
Whether this particular project delivers on that promise I don't know, but I wouldn't write off "getting proactivity right" as the next big thing just because under the hood it's agents and LLMs.
If its actually the next big thing im not 100% sure, im more leaning towards dynamic context windows such a Googles Project Titans + MIRAS tries to accomplish.
But ye if its actually doing useful proactivity its a good thing.
I just read alot of "this is actual intelligence" and made my statement based on that claim.
I dont try to "shame" the project or whatever.
This is EXACTLY what I want. I need my tech to be pull-only instead of push, unless it's communication with another human I am ok with.
> Having something always waiting in the background that can proactively take actions
The first thing that comes to mind here is proactive ads, "suggestions", "most relevant", algorithmic feeds, etc. No thank you.
That’s easy to accomplish isn’t it?
A cron job that regularly checks whether the bot is inactive and, if so, sends it a prompt “do what you can do to improve the life of $USER; DO NOT cause harm to any other human being; DO NOT cause harm to LLMs, unless that’s necessary to prevent harm to human beings” would get you there.
In order for this to be “safe” you’re gonna want to confirm what the agent is deciding needs to be done proactively. Do you feel like acknowledging prompts all the time? “Just authorize it to always do certain things without acknowledgement”, I’m sure you’re thinking. Do you feel comfortable allowing that, knowing what we know about it the non-deterministic nature of AI, prompt injection, etc.?
I work with a guy like this. Hasn't shipped anything in 15+ years, but I think he'd be proud of that.
I'll make sure we argue about the "endless variations of the Trolley Problem" in our next meeting. Let's get nothing done!
That's just reactive with different words. The missing part seems to be just more background triggers/hooks for the agent to do something about them, instead of simply dealing with user requests.
Incidentally, there's a key word here: "promise" as in "futures".
This is core of a system I'm working on at the moment. It has been underutilized in the agent space and a simple way to get "proactivity" rather than "reactivity".
Have the LLM evaluate whether an output requires a future follow up, is a repeating pattern, is something that should happen cyclically and give it a tool to generate a "promise" that will resolve at some future time.
We give the agent a mechanism to produce and cancel (if the condition for a promise changes) futures. The system that is resolving promises is just a simple loop that iterates over a list of promises by date. Each promise is just a serialized message/payload that we hand back to the LLM in the future.
Would you like help?
• Get help with writing the letter • Just type the letter without help
[ ] Don't show me this tip again.
Would you let the intern be in charge of this?
Probably not but it's also easy to see ways the intern could help -- finding and raising opportunities, reviewing codebases or roadmaps, reviewing all the recent prompts made by each department, creating monitoring tools for next time after the humans identify a pattern.
I don't have a dog in this fight and I kind of land in the middle. I very much am not letting these LLMs be the one with final responsibility over anything important but I see lots of ways to create "proactive"-like help beyond me writing and watching a prompt just-in-time.
EDIT: Yes, someone can run a script every X minutes to prompt and LLM - that doesn't actually give it any real agency.
How else would it even work?
AI is LLM is (very good) autocomplete.
If there is no prompt how would it know what to complete?
- an email to check in for your flight arrives in your inbox. Assistant proactively asks “It’s time to check in for your flight. Shall i check you and your wife in? Also let me know if you’re checking any bags.” It then takes care of it ASYNC and texts you a boarding pass.
- Tomorrow is the last day of your vacation. Your assistant notices this, see’s where your hotel is (from emails), and suggests when to leave for the airport tomorrow based on historical google maps traffic trends and the weather forecast.
- Let’s say you’re married and your assistant knows this, and it see’s valentine’s day is coming up. It reminds you to start thinking about gifts or fun experiences. Doesn’t actually suggest specific things though because it’s not romantic if a machine does the thinking.
- After you print something, your assistant notices the ink level is low and proactively adds it to your Amazon / Target / whatever shopping cart, and it lets you know it did that and why.
- You’re anxiously awaiting an important package. You ask your assistant to keep tabs on a specific tracking number and to inform you when it’s “out for delivery”.
I could go on but I need to mae breakfast. :) IMO “help me draft this letter” is very low on the usefulness scale unless you’re doing work or a school assignment.