Hacker News Reader

Comment by no_wizard | Hacker News Reader

What I want is for it to surface information to me, not me have to query it.

Where is that AI? For example, if I usually eat between 2-4 PM, and I'm in the middle of time square, start suggesting places to eat based on my transaction history, or location history of restaurants I frequent. Something like that would be useful.

If I have to ask, I might as well look at my phone most of the time. It'd likely be faster in most cases.

I don't need something like that, where it must be queried to be useful, like asking it to read back my text messages, but I sure would love it if when my wife messaged me, it was smart enough to play the message unprompted if my headphones are already connected and active

View on HN

Not saying it's your exact use case, however, in the "saved info" section of gemini I have a prompt about the llm letting me *know what's on it's mind " along with some other details to where when I am just "chatting" it has brought up relevant books to our previous discussion / projects. Alongside local events (bay to breakers, roots game memorial day weekend, some single events, etc within that first" hello" of our conversations and brought some news to the foreground that was relevant to me, although I wouldn't necessarily seek out that info. It's been so handy to bring relevant info into my hands in an actionable amount of time. Plain Jane gemini didn't offer those amenities but I was able to build them out.

How do you propose you would train it to know what you want? Besides a versioned system prompt that you (or the AI) would have to continually adjust.

If it gathers enough data on you, it can theoretically figure that out. Siri "Suggestions" have been around for years, and if you go to the same place frequently on certain days/times (e.g., your workplace, a friend you visit every Saturday, places you often go to lunch) or use the same apps at similar days/times (e.g., pulling up Callsheet in the evening when you're watching TV shows or movies to do TMDB lookups), those suggestions will show up. All of those examples are real ones I've experienced. The quality is certainly variable, but it's decent.

(Of course, it's "non-LLM" AI, which isn't particularly fashionable right now, but if we really want smarter AI agents we need to stop treating all problems as solvable with large language models.)

I don't think this leaves out initial setup. Another source of information: habit observation. If I do something around the same time every day, over and over again, it would be nice if it simply helped me along unless I interrupt explicitly. It should fine tune itself to observations it makes about my behavior and patterns, as opposed to me interjecting constantly

The constant need to query for information, rather than have useful information contextually pushed to me, fundamentally limits its utility once the novelty wears off. Without a sufficient complexity threshold (and this assumes accurate information and trust) its more work to query for things than it is to simply do them.

From a consumer perspective, thats not great.

If you had a butler or PA who's with you all the time, they would know what genre of food you like based on your restaurant visits and your ravings about what you liked. The imaginary AI would have your location history for restaurant visits, your Instagram feed for pictures of foods/review of restaurants, your chat history to see what you've raved about. It would also have big data from other people who have seemingly similar tastes to you, to recommend you the next place to eat.

Obviously since we're in lage-stage capitalism and everything is designed to extract profit out of you, we can't give commercial systems all our private data...

Reply on news.ycombinator.com