š„ Modern parenthood, invisible load and AI .... 5 lessons after a year of putting AI to work
We've been handed fire - question is: how to harness it safely and well, without burning the place down.
What a difference a year makes. A year ago I wrote with the blind optimism of someone handed a magical elixir. Someone who believed that this new tech of LLMs could simply and uniformly be the answer to the ever-present invisible load: Just drink it and āpoof!ā - problems solved.
Today I write to you with the informed, measured optimism of someone who realizes that the magical elixir is actually fire: something bigger and more transformative thanĀ imagined, but also something that is much harder to wield reliably and safely.
šš½ Btw, this month Iām personally helping families tackle this invisible load for their unique situations. 4 weeks to lighten the load - if you want to give it a shot: https://lnkd.in/gtrUPdED
āØ Anti-tech tech
Over the past year, weāve designed, built, and tested our AI-powered partner for families. In doing so we have learned a tremendous amount - not only what works for a couple dozen families at a high cost, but can work for thousands and millions in a really affordable way.
After all this work, I realized that what we are building is actually anti-tech tech.
Cutting edge tech built for balance to all of the noise. Instead of productivity software to do more in less time, a forcefield that filters and allows you to focus on what matters most.
Milo is two halves to make a powerful whole:
something simple that can pull all the info into one place and keep everyone informed (solvable with good olā fashioned specialized software)
a partner to make sense of it all and help with prob solving and the joy bitsā this has been the stubborn piece that eluded solving - until Al/LLMs.
But in the process of building, weāve gained some incredibly important insights - some on the technical side, others on the product and business model, but all in service of what it means to build technology, made for the way we human.
šš½ 5 things weāve learned:
š While AI feels like magic, it is not, in fact, magic. AI (large language models specifically) have long been a promise with choppy progress. Some big breakthroughs in recent years all building on top of each other have given the world something that feels very different, powerful and yes, magic.
But beyond ChatGPT, there are few products that have been able to really harness this power at scale and in a way that feels reliably helpful. The challenge lies in picking problems that LLMs are good at solving in a āprobabilistic wayā. Meaning that you can work with the user to give likely options but itās not frustrating if it gets it wrong. Itās why ChatGPT is so great on fun poems, first drafts of letters, brainstorming and why there is whole separate machinery when you ask it for facts or recipes or trip itineraries for actual places.
The trouble is, for the past 15 -20 years, weāve been trained to expect ādeterministicā services. Things like Instacart and Uber - press a button, get your groceries or a taxi. The whole value proposition depends on the contract of āask for a thing, get that exact thingā and removing any uncertainty.
LLMs (large language models) are on the other end. They are probability machines - spitting out the āmost probableā thing, given a request or instructions. It makes them super helpful in solving problems with many options but the one thing they canāt give you is certainty. Especially not at scale.
Theyāre great as the other half of deterministic software because they can do what rule based software canāt.
But to date, beyond ChatGPT, the only other product that both solves a problem and really works at scale is Github āCopilotā. Itās a use case and a UI that allows LLMs to predict the answer without the noose of certainty. You ask a thing, Copilot suggests the most probable solution. If itās exactly as you need it (rare but possible), you can take as is. More likely, itās a very good starting point that you can edit from there (hence the ācopilotā). And if itās way off? You just decline the suggestions, improve the prompt or do it yourself.
At the core - the contract is not of certainty and reliability.
Thatās a massive mindset shift. There are lots of places I want to give parents the ācertaintyā that something will be done - immediately, easily and perfectly.
But thatās not how this works - especially not in these early innings of the tech.
So weāve had to pick. Fo the places we need certainty (like when is pizza day or add this soccer email to our calendar) - weāre building deterministic/certainty tools. And where we need smart suggestions in a messy āit dependsā world that I can review/edit (figure out pick-ups for this week or what conflicts do we have) - weāre building a probabilistic copilot and partner.š³ Sustainable, scalable solutions matter, especially in AI. Getting a new product or solution to work is hard, not only because youāre going against the laws of physics that would prefer a system to stay as it were or the status quo, but because you also need to do it in a way that also obeys the laws of capitalism.
Iāll often be asked: why is this a paid product? Well lots of reasons - the first being: where there is value created, I think itās fair to ask for value to be rewarded. It holds our team to the highest bar of being sure weāre building things that matter and actually drive value.
The second is because there are only so many ways to make money and I have been clear from the start that compromising on selling data, trust or attention is not how we want to do things. āFreeā products have to make money some way and it typically ends up being via selling data, affiliates or ads. Instead, we choose to keep things simple by asking users to pay - like you would for Airpods or Nikes.
And lastly: while software might be cheap, AI is not. Each request and action has real cost to it. While we expect it to get cheaper and weāre constantly optimizing, we also need to use the state of the art models and capabilities. The last 2 decades have also been the era of many, free things. I think for lots of reasons weāre entering a new era of being intentional on a fewer number of high value things that we also pay for.
Over the past year my thinking has evolved on what is the value being created, what is the cost of serving at the highest levels and how can we offer both - a base, free product that can help anyone with the basics and then a fully powered AI partner that earns its keep and then some.šļø Garbage in, garbage out. LLMs are trained on data. Loads of it. But what data, and what theyāre trained to focused their attention on, matters. These models are effectively a representation of the world as it was handed to them in data. Most like GPT4 have been trained on the data from the internet, and then fine-tuned/ aligned towards certain goals.
As we started, we figured that parenting was like a general life use case. Turns out that parenting (and care, more broadly) is much closer to a specialized knowledge space like medicine or architecture - it has complex info/logic (childcare, pickup/dropoffs, picky eaters), unique lingo, invisible dependencies (bday party invite = reminder to buy bday present) that isnāt obvious if youāre not a parent.
Using generalized models work to a certain degree but after that, it takes a lot of work to make it understand that meals usually mean 5pm not 7pm or that kids go through food phases or if thereās soccer the meal needs to be quick and portable.
Building the Family AI means teaching, painstakingly, what it is to care. The art and skills codified into words and logic and actions. Itās an act of love in itself. To see every usually invisible action, to value it, and to put it to work to help.
A year into it all I know is that weāre still so early but also that AI is here to stay in big and important ways. And we need to design and dictate what those ways are. We need to painstakingly teach one of our most human things - how to care, how to family, how to community and to build representations that honor and do justice to the complexity and beauty of the work of our homes.š Safety, privacy and security is all in the details. After a year+ of being in the literal weeds of building hereās the only thing I am certain of: there are many things to be concerned about but the real things we should be worried about we canāt actually see yet. Weāre so early that most of the scary things being talked about have more to do with sci-fi movies than real things you should be worried about.
Itās why I think even more than what is being built, I think consumers need to know and judge whoās behind it. What is their world view and their values? Because those are the guideposts that are going to be making the small calls on the ground that end up deciding the big things.
AI is an amplification engine. From the speed to the computing, it is a tool of our creation that allows us to amplify ā¦ really anything. So what are those things and who is making the calls on how?
Everyone is going to say that data privacy and security is of āutmost importanceā - but is the product āfreeā? How do they make money? āAI safety is a priorityā - but safety is a relative term - whatās the line of harm? What does safe mean to them vs. what it means to me.
Iāve pushed to frontlines of these issues and questions not only because I feel the responsibility for Milo and our families. But because I have 9 and 12 year old girls at home, headed into some of the most complicated years (with or without tech). How is my familyās data dealt with? How can I make it as secure and as private as possible? How can I add the guardrails needed - erring on the side of prudence.
Overall I know this: more than the headlines, these battles are and will be fought in the daily little calls. And for my money, Iām judging the whoās behind the thing, what they believe/ are optimizing for and how theyāre making their money.
5. āØ Magic lies in the practical, the understandable and the dependable. Perhaps the biggest lesson Iāve learned: what magic actually is - at least for now and for families.
Itās the way of a long time relationship or marriage - learning that love isnāt in the big flashy moments of a grand gesture, but in the thousand little acts of service over long days and hard moments. In quietly and constantly showing up everyday and doing the work.
In the same way, what weāre able to do should *absolutely* feel like magic. But not the big flashy kind - ask a thing and *poof!* - itās done. But rather have magic lie in the constancy and transparency and sheer dependability.This kind of magic is painstakingly stitched together in a careful orchestration of software and AI and human guidance. And it asks you to be a partner to it. To not bear the entire burden of a thing but ratherā¦ share the load.
Magic also lies in teaching. In helping people understand exactly what they hold in their hands. Today, many use ChatGPT - usually with wonder and delight or trepidation and skepticism. But itās been hard to understand what it actually is and how it works in the details. The truth is, once you learn that you start falling somewhere in the middle - seeing exactly what should be cause for marvel and what for caution.
š Still just the beginning.
As much as weāve learned and built, this post really just marks a beginning. An invitation to join the journey.
Because in such high stakes and friction filled space, the job isnāt simply to build a solution, itās also to do no harm. To not suggest it can do a thing if it can't do it, day in and day out.
Parenthood needs magic of the dependable kind because at the end of the day, isnāt that what the magic of parenthood is?
And itās more than time we got magic back into our lives, but the kind we can rely on.