Getting chatGPT To Remember You
The challenge
A great assistant learns your tastes so it can get more and more accurate as it learns more things about you. But chatGPT doesn’t know who you are from session to session. On the other hand, Bing + chatGPT can read websites, meaning you can give it a personal dossier to base its results on.
Example one: what movie should I watch?
In this test, I asked chatGPT and Bing+chatGPT the same question. chatGPT can’t read websites so it gives me some standard suggestions. But Bing can see the note I’ve written about Sophia Coppola and give me recommendations based off it. So let’s keep adding things to the dossier.
Example two: where should I eat?
chatGPT would have done a lot better if I had said “near me in Tulsa” but it wouldn’t have known this rhetorical person was a vegetarian unless I specifically spelled it out in every search. Bing spots both details and does a great job with the result.
Example three: a present + activity for my daughter
I’ve mentioned that Sally is 9, likes Star Wars, and uses a wheelchair, so when I ask for book and activity advice, I’m told about a great book and something called “wheelchair hiking” that is apparently popular in Tusla. Home run.
Example four: what should I read?
This was an interesting one. When I asked about books, it gave me general advice. But when I asked it to focus on this page only, it got more relevent:
When I asked to learn more about the recommendations, it explicitly talked about reviews and popularity, which is applicable to anyone. But it implicitly could tell that I wanted to hear about relationships because of the phrasing I used, and the examples I used. NANA is about two women, both named Nana, who end up becoming roommates. Notice that 2 of the 3 recommendations are about sisters who are close! Pretty cool.
Example five: now you know me better, any other movies?
Bing + chatGPT messed up here because On the Rocks is also her movie. But other than that error, these are good insights and suggestions. Really good. Another fantastic answer. (Before Sunrise is amazing, btw)
Example six: what manga should I read?
After 11 data points, I can just say “based on what you know about me…” and get a really good answer. Note that it’s not just spotting examples of manga I mentioned, it’s specifically noting that I said I like memoirs and real-life experiences, and it’s tying that into the result.
Note: I’ve made up this dossier so it’s not exactly my tastes and preferences. But I’ve never heard of Penguin Loves Mev before now and it actually might be right up my alley! :)
This is exciting
I was on the Windows Phone design team when we dreamed up Cortana. My very small contribution (I wasn’t formally on that team, but the people on the team were my friends/mentees) was talking about making Siri less of a black box. Those modest insights are what one day turned into Cortana Notebook.
There is no doubt in my mind that AI software will have a notebook or dossier attached to it in some way. In this example I’ve hacked it together with a little webpage, but in the future it’ll be a standard part of the software that everything is run through.
You shouldn’t have to add “wheelchair friendly” or “married” or “allergic to fish” or “lives in Tulsa” or “already saw Avatar” or “doesn’t listen to top 40” ever again. You should be able to explain your tastes once, and only once, and know that the recommendations are always tailor-made to you.
This also means apps are dead. But that’s another post.