Hello to the 19 subscribers of this newsletter. Last week’s issue received 46 views, and 3 new people signed up.
In this issue
Deconstructing personal AI assistant service
Some interesting things I found online
📝 What would it take to create a sustainable personal AI assistant service?
One thought exercise I like is brainstorming how to execute products that I’m most undoubtedly sure will be a commodity in the future but are not available just yet for some reason. They have to exist; it’s just so apparent that you don’t even question it.
There is one product that I can’t stop thinking about since its launch in 2017 and its unfortunate demise in 2019. I can’t stop thinking what it’d take to execute this idea, how to make this happen. I considered starting a company based on this idea twice over a two-year period. But decided against because of not understanding how to execute it well enough and be profitable at the same time. This idea excites me so much, but it seems incredibly challenging to execute well.
Today I want to share some of my thoughts on this.
In November 2017 new product called Fin launched on Product Hunt and their demo video just blew my mind:
🤯 Holy. Shit. If what I just saw is for real, this is HUGE!
Here is the basic idea:
Fin is your personal AI assistant.
But it’s not fully AI; it’s hybrid.
AI helps carry out simple tasks by itself (schedule a meeting, send calendar invite to my family members, order tomatoes, etc.)
An actual human carries out more complicated tasks.
But AI helps this person to perform certain tasks faster and better.
For example, a customer asks to book “that Greek restaurant he went to last April.” A human would need up to 10-20 minutes to scroll through the chat history and find out which exact restaurant he has in mind. But good AI can find that instantly.
The idea is that over time AI learns more and more about your preferences and tastes and creates a massive structured database of your likes.
Sounds challenging but doable in the long run. Let’s deconstruct some examples to figure out how to perform them.
Example 1.
The customer sends a link. The customer follows with a voice message:
Buy ingredients for this recipe, schedule delivery when I’ll be home this Friday evening, and add the delivery to my calendar.
This sounds easy if you have a decent NLP.
Go to the link.
Parse recipe from the webpage (most good recipe managing software can do this).
List ingredients from the recipe.
Open grocery delivery app (if the app doesn’t have an API, this is where the human comes in; AI would handle him a list of ingredients and the task itself “order groceries”).
Human assistant schedules delivery.
AI adds the delivery date to the customer’s calendar.
Sounds manageable (again, if you have good NLP).
Example 2.
Customer sends a voice message:
Could you please find the best spa center in the city and find if they have gift certificates. If they do, let me know the prices.
This one is more difficult without human involvement.
Search specific internal knowledge base for spa centers (I assume having pre-defined sources of truth for different categories would help a lot).
List 5-10 with the highest rating and send to a human.
The human decides what is “the best” (does some additional research) and sends the choice to the AI.
AI goes to its website and looks for “gift certificates.” If finds – sends to the customers.
If the website doesn’t have any, then AI asks a human for help to call this place.
Human calls the spa center and asks the real person about gift certificates.
AI parses the data from the call, picks gift certificate prices, and sends it to the customer.
I bet if you show it to someone and ask, “Would this be real in 20 years?” almost always, you’d hear a resounding “Yes!”.
But why it’s not available now? And why did this Fin project failed?
Here are four major challenges I see.
Challenge 1 – Technology
This is the obvious and the easiest to fix, in my opinion. We are still not there just yet in terms of NLP quality, but it looks like you just need more data. GPT-4 or GPT-5 might provide the level of quality we’d need here.
Challenge 2 – Discovery
This is what most voice assistants still struggle with. If your assistant can help you with 200-300 mini-tasks but is useless with the other 500, how do you explain to the customer what exactly he can do? The line is tough to define.
Challenge 3 – Pricing
How do you consider pricing in a hybrid model? Some tasks can cost practically nothing (those done entirely by AI), but some can cost a fortune. And the worst part is that the customer can’t know beforehand how much that would cost.
Fin had probably the worst pricing strategy I’ve ever seen. You had to pay a base subscription fee + $1 per effective minute. What the fuck is “effective minute”? How many “effective minutes” would this task take? No one knew, and it caused a lot of uncertainty.
Here is one review on Fin:
I've tried it three times, haven't been able to get value out of it. Last time I got a $160 bill for finding and booking an in-home dinner chef for a trip my fiance was taking with friends, for 8 emails and 20 mins on the phone - way too expensive for normal use.
Challenge 4 – Privacy
Real personal assistants usually have access to your work email, calendar, and credit cards. Without that, you’d get way less value. I’m okay to share these with a real person I’ve built trust with over at least a few months. But how do you share that access with a network of human assistants? They can change every single day, and you don’t even know who these people are.
You could store these on your website in a secured fashion without providing access to human assistants. But how do you buy plane tickets from an external website? I don’t see how this can be done.
Six months after the launch of Fin, Google launched a product called Duplex.
Duplex was a voice assistant that could make phone calls for you. In our Example 2 above, step 6 could be done without turning to a human for help.
This seemed extremely promising at the moment back in 2018, but look at what Google is selling under that product now – drive website visits and get more phone calls for your business.
Fin shut down somewhere around 2019, but I still can’t stop thinking about the world where AI assistant of this level exists. I hope someone somewhere will crack it up and solve this extremely challenging task.
👓 Things I've been reading/watching/enjoying
The end of the office
Seth Godin returns to the history of modern office and why it has got to come to an end
As social creatures, many people very much need a place to go, a community to be part of, a sense of belonging and meaning. But it’s not at all clear that the 1957 office building is the best way to solve those problems.
How to rest well
A skill you probably haven’t considered for your resume. But the one needed for creative, sustainable life.
Playlist Cover Maker
Last week I decided to make my Spotify playlists easier to find and better-looking. This tool is free and created covers I’m super happy with.
👨🏻💻 Behind the Scenes – Preparing this issue
That’s it for today. Thanks for reading. Until next week 👋🏻