In the mid-1970s, Americans hosted or attended an average of 7.1 dinner parties a year.

Explains why in Mad Men, people were always glammed up over glasses of wine.

By the late 1990s, dinner party attendance nearly halved, according to survey data from Bowling Alone.

It’s gotten even more dire. In the last year, 43% of people said they had neither hosted nor attended a dinner party!

Sure, sure, this all makes sense. I mean, life today is busy! And restaurants are convenient! And a bunch of us live with our parents. Also, do you know how much work it is to have people over? Even coming up with a date is a chore, not to mention figuring out who to invite, what to cook (there is many a gluten / shrimp / nut allergy among us!), food prep, post-party clean-up, and of course the low-grade anxiety of making sure everyone is entertained.

At the same time, we keep sighing, “We should get together!” or “It’s been waaaaay too long!” when we see each other. One wonders: shouldn’t this be solved by now?

Sonia Baschez went viral for her comment recently that today’s consumer tech seems geared towards “husband-y tasks” — i.e., one-off random jobs like booking flights or buying tickets.

But what about realizing that it’s time to buy a gift? Or planning the weekly shopping for the daily meals? These “wife-y tasks” deal with the subtle art of noticing, deciding, and keeping track.

The good news is that the capabilities are starting to arrive! AI Agents are increasingly competent at suggesting plans, remembering details, and (with the right permissions) taking actions on our behalf.

But today’s AI is woefully lacking when it comes to furthering our social lives. In this essay, I describe the 3 jobs a truly social AI would do well.

The Looking Glass discusses how to build and live better. Subscribe to follow along.

Social Task #1: Help us get together

Real-talk: I hate hate hate organizing and planning logistics. I hate the nitty-gritty origami work of calendar-matching and invite sending and party prepping. I hate it so much that I married someone who excels at this so I wouldn’t have to.

But perhaps AI can help me rewrite my identity?

This is the utilitarian part of social, the Make it easier to get together, and it’s the first echelon that personal agents can actually tackle.

My college reunion is just right around the corner, and over a recent dinner with my besties we reminisced on all the folks who had lived with us freshman year. What were they like now? Wouldn’t it be great to see them all?

I searched the catalog of mini-reunions for my freshman dorm, hoping somebody else would have taken up the mantle of coordinating. Nope. No such luck.

So it came down to this: make it happen, or give up the dream. A year ago I would have surrendered. This year, I dispatched the agents.

And boy were they up to the task. They scoured the reunion yearbook to figure out who was coming to this thing. They collected the emails. When emails were missing, they did reconnaissance across personal websites. They published the Partiful and pushed the reminders. They read itineraries and mapped locations and got approvals for permissions. They delivered me a to-do list.

More importantly, I am now convinced that YES I CAN HOST AN ACTUAL DINNER PARTY!

Of course, hosting at that scale is still a one-off event, but even for more casual get-togethers, AI agents earn their ilk. I’ve been saying for years how awesome it would be to have my agent show up in my group messages and help coordinate a get-together or a date night.

Just last week, Instinct announced early-access group chats, which work even if friends don’t have Instinct accounts. Its proposed uses include planning trips, coordinating carpools, and buying and splitting tickets. The best UIs are often invisible because they don’t require us to change any existing patterns, and I’m sure we’ll see this feature across many other consumer agents going forward in the places we already communicate, whether chat, email or Slack.

Social utility would be even more powerful with agent-to-agent coordination, where each agent can look at their user’s calendars, report on their preferences, and determine the best path to a fantastic Friday night.

It’s already happening! See example below.

Meta’s Muse also considers hosting a major use case, with the ability to remember friends’ dietary restrictions, suggest a dinner-party menu, and turn saved recipes into a grocery list.

My mini-reunion would have taken at least 5–10 hours of planning in the olden days. Now, I’m able to make it happen in 30–60.

If that doesn’t bode well for the future of dinner parties, I don’t know what will.

Social Task 2: Help us become closer

Utility and time savings are great and all, but what about helping me connect more deeply with the people I already know?

I caught up last week with a good friend I hadn’t seen in a few months and as we were exchanging our big updates (a lot of life can happen in 3 months!) I was reminded of how much I missed chatting with her. Yes, our schedules are busy with kids and work and travel and whatnot, but I also felt a twinge of Why didn’t I reach out more? Couldn’t I have texted hi, or shared something interesting / funny / outrageous?

This is what I want my agent to do: be on the hunt like a mouse to a cheddar trail for daily opportunities to strengthen my relationships.

For example, I just took my kids to the pumpkin patch. Could my agent notice the photos in my camera roll and suggest we send them to the grandparents and extended family?

Nudge me to share with my mom!

I just finished a Kindle book that my bookish buddies recommended. Could my agent remind me to share back what I thought?

I am scheduled to go to San Jose later this week for a conference. Could my agent hit me with a message that says “A, B and C live in San Jose close to your conference location. Want to ping them for a coffee afterwards?”

I’ve been overthinking an interaction with a colleague and wondering if his sharp words imply he thinks I’m an idiot and therefore I should mount an ironclad defense. Could my agent lightly advise me to go for a walk and possibly re-read some of my old essays on the topic?

Social apps used to be about connecting more meaningfully with friends and family. But then they got overrun by algorithms and short-form videos and an endless stream of enticing content from people all around the world whom we will never meet face-to-face.

I want a return to social social. I want an agent to realize that it’s been too long since I’ve gone on a walk with X, Y or Z and suggest that I reach out about doing a dish hike. I want my agent to know the important people in my life and gently direct my attention towards investing into them. I want my agent to catch me when I’m neglecting friends, and brainstorm with me on ways of bringing them together, and talk me off a ledge whenever I’m overreacting.

Research shows that people feel greater well-being when they receive a personal message from a close friend. Reading broadcasts and getting a “like” did not have the same impact.

I am not saying that every article of communication should suddenly become AI-generated (in fact experiments on AI-suggested replies show that people evaluated partners more negatively when they suspected AI use). There is a world of difference between aiding sincere intent and normalizing one-click platitudes!

Aristotle described a good life as “sharing in discussion and thought.” I think about that whenever I send my mom a photo, or text a friend about a memory. These are small ways of inviting someone into my day. I want AI to help me do more of that.

iii. Social Task 3: Help us find each other

Whenever I visit Shanghai, my childhood home, one of my favorite things to do is to stroll the big green public parks where the retirees do tai chi together, play chess, chase grandkids, and occasionally broker their children’s marriages.

This type of match-making involves the mothers talking up the wonderful qualities of their (usually) son: He has a great job! He makes this amount every month! He has excellent vitality and a great temper!

Let us examine the goods!

The moms here play the traditional role of agent: a third party who represents and negotiates one’s interests. Pre-AI, we already had talent agents and sports agents, recruiting agents and author’s agents.

Imagine now the AI version that knows far more about your interests, what you consider a stimulating conversation, and the kinds of people you want to meet. Imagine them on a mission to pitch your great virtues in front of your best prospects, whether for dating, work, friendship or simply a fabulous afternoon. Imagine them helping you debrief on whether that person actually likes you, or whether this deal is the best that you can get.

I see a distinction between personal agents and mediator agents, and both have their place in helping us expand our social networks. Whereas a personal agent represents solely you and what you want, a mediator tries to find the ideal solution for multiple parties. Your personal dating agent may hand you a list of people that it deems optimal given your list of “must-haves,” while a matchmaking agent might know that you and X would really, really get along even if X isn’t what you think you want. I imagine mediator agents like good hostesses who know both sides and can smooth over the awkwardness of new introductions.

It’s interesting to think of AI agents gossiping about us (as of course human agents do!) In the best case, they make our connections to new people feel warm and inevitable (“X has told me so many great things about you!”). They’ll share tactful connective details that make it easier to get conversations started (“You are both people I have gone very deep with on Taylor Swift song lyrics!”). They could even follow up like a good conspiratorial friend (“I heard the conversation was stilted at first but then ended amazing, how’d you manage that?!”)

Have you ever imagined how many potential best friends are waiting for you out there, on your street, in your neighborhood, at your workplace or even across the vast expanses of this Earth? People who happen to have a soft spot for niche 80s animes like you do, or obsess over cat-and-mouse detective shows, or are 3w4 enneagrams, or have the exact same infuriating dynamic with their husband’s running shoe collection.

And yet, there are so many little hurdles that get in the way of those hypothetical friendships: shortage of time, lack of reason, calcification of past disappointments, abundance of subtly chafing edges surrounding any new social interaction.

David Whyte wrote in his book Consolations that “the ultimate touchstone of friendship is not improvement, neither of the other nor of the self, the ultimate touchstone is witness,” the privilege of seeing and being seen.

Agents can build us the bridges; whether we walk across them to be the witnesses through each other’s ups and downs is something we’ll have to discover together.

I hope those matchmaking moms in the park will one day soon be doing tai chi and chess and running after their grandkids.

Subscribe for more essays on technology, building, and the things that make us human.

More from my AI essays

• Everyday life:
• What will we use our minds for?
• What will make personal AI go big?
• AI for No Good Reason
• How to Stop Complaining About Your Software
• How life was different during the early Internet
• Work and organizations:
• You Cannot Mandate an AI Transformation
• You’re definitely going to be a manager now
• The Thing You Are Expert at Will Be Your Career Downfall
• The Death of Product Development as We Know it
• Design and taste:
• Conversational Interfaces
• When AI Has Better Taste Than You
• The AI Quality Coup
• Designing in an AI World
• Being human:
• I am an idiot
• Eye, Hand, AI: a letter to my daughter
• Our Souls Need Proof of Work
• What AI Won’t Change

In 1867, a young journalist named Henry Morton Stanley met a man whose body he couldn’t help describing: “broad shoulders, well-formed chest and limbs, and a face strikingly handsome.”

According to Stanley, “Whether on foot or on horseback, he was one of the most perfect types of physical manhood I ever saw.”

That man was Wild Bill Hickok, whose career included being a wagon master, scout, and guide. His day-to-day involved the moving of people and objects across vast territories.

Wild Bill Hickok in buckskins; alas I could not find a shirtless one

Now consider another famous specimen of physical manhood: Arnold Schwarzenegger. How did he earn that distinction?

By going to the gym. A lot.

In a 1991 interview, he said he trained five hours a day when competing and said, “I love it.”

Arnold Schwarzenegger training at Gold’s Gym in 1975.

In both these examples, moving a lot of weight around was how one got that sweet, sweet body.

But in the first case, Wild Bill did it to earn his livelihood.

Arnold, and many generations of gym-goers after him, did it because we want to be fit and healthy and hot, and maybe because we also happen to love it.

History doesn’t repeat, but maybe it rhymes.

i. the evolution of jobs

A long time ago, physical exertion came bundled with living. A farmer who finishes his day doesn’t think, “shoot, I need to lift some more heavy things.”

But slowly, technology took away our need for physical labor. One study of American jobs estimated that those requiring moderate physical activity fell from 48% in 1960 to 20% in 2008. Now I’m not saying that lifting hay bales resulted in some golden age of physical health. But as jobs and transport became more sedentary, moving your body required more intentional choices.

Meanwhile, work expanded to stretch humans in a different way. Only 10% of Americans had professional or managerial jobs in 1900, but that figure grew to 40% in 2017. In last year’s employment numbers, that’s more than 100 million Americans!

For many of us, the workday delivers a steady stream of intellectual challenges. How do I explain this complicated idea? Why aren’t the numbers adding up? What’s the best way to solve this customer’s problems? How do I create a more efficient solution?

Enter now the era of AI.

These days, AI is being used to do more and more hard human tasks, like writing code, or drafting emails, or analyzing data. It’s even solving our hardest math problems!

When machines took over physical tasks, many new mental jobs sprang up. If machines take those over too, what then?

ii. we still like doing hard things

Consider the marathon.

The first occurred thousands of years ago, when a messenger ran all the way from the city of Marathon to Athens to deliver news of victory. He blurted out the message and then fell over dead, exhausted by the effort.

Today, we have cars and trains, bikes and boats, all perfectly good ways to go 26.2 miles without losing a toenail. Yet 1.3 million people applied for the 2027 London Marathon, more than twice the number who applied for the 2024 race!

That’s a lot of people hoping for the opportunity to do something hard!

We used to saddle up for war to defeat neighboring invaders. Now we lace up our trainers to enter the stadium. We used to cheer on our military; now we cheer for our World Cup team. Competition is alive and well; even in a world with abundant goods, beating someone who is trying to beat you is a compelling game. We like to see ourselves improve: run farther than last month, make that trick shot, lift heavier. And these games are even more fun with others working and watching.

We have long passed the era when physical fitness mattered for the sake of hunting antelopes or defending property.

For many of us, we do it because we can and because we want to.

iii. thinking for the fun of it

Look at the games people end up playing in retirement: bridge, mahjong, chess. The end of a career doesn’t mean the end of deep thinking.

With bridge, you have to remember which cards have been played, calculate what others are holding, and coordinate with your partner. In chess, you try to imagine what your opponent will do before they do it.

For 30 years, my grandmother played mah-jong like it was her full-time job. She’d dutifully leave every morning at 9am to join her ladies, and return before dinner smirking because she’d won a few extra bucks. She remained sharp as a tack until she passed.

My Ah-bu teaching my daughter to play mah-jong

Chess I find especially interesting because computers already play it better than humans. After Deep Blue beat Kasparov, many bemoaned the end of a beautiful human game. But this has proved to be patently false: chess is flourishing. Chess.com reported 250 million accounts in February 2026, up from 100 million in December 2022. The knowledge that a machine can make a better move hasn’t eliminated the joy of finding your own.

Can these activities help us stay sharper? There is some encouraging evidence that researchers who tested training in memory, reasoning, and processing speed in older adults showed benefits at the ten-year follow-up.

Of course, we still have a lot to learn about the specifics of mental practice. In the physical world, we have developed an entire science of training: Zone 2 versus Zone 5, maximizing volume versus load, compound versus isolated movements. We have established gyms and trainers, athletes and influencers, exercise and nutrition programs.

Could we do the same for the mind? Imagine stopping by a mental gym after lunch. One room has chess, bridge and mahjong; another StarCraft, Civilization and Age of Empires. There is a puzzle emporium. There is a philosophy debate club. There are computers for learning anything, and labs for making anything. There are intro classes and tournaments that people take way too seriously.

Mental gyms of the future. Hey, one can dream!

People hire coaches to help them train. They hit the mental gym with their buddies and measure their progress and get excited about every tiny improvement.

The Peter Attias of the future will write bestselling books about mental longevity and how to keep your memory, creativity, and critical thinking sharp as you age.

We already have all the ingredients today in the form of clubs, community halls, schools, libraries, and lecture halls. What will grow is the recognition that all this effort needs intentionality.

Of course, retiring is not the same as losing a job, and worrying about how you’re going to pay rent does not create an atmosphere of leisure. There are many questions yet about the manner in which the AI revolution will give people freedom in terms of money and time.

But for those of us already finding that AI can reduce our grindwork, the question that starts to pulse louder and louder is: what do we do with more space?

iv. what do you want to get good at?

What’s 17 × 24?

If you’re like me, you might start reflexively calculating in your head that 7x4 = 28 or 20x25 = 500, before another voice interrupts with this is hard and presents an image of a calculator.

And lo and behold, the calculator stoically delivers the answer: 408.

Is this a good thing?

I ask myself this whenever AI delivers me answers without the mental wrestlings of yore. It pulls me to stay high-level, to repeatedly give feedback on results rather than dig into the details of the process. My attention feels like dandelion seeds scattered to the winds, each branching into new territory rather than deepening its roots.

I recently had a chat with Joe Liemandt, co-founder of the buzzy K-12 Alpha School. Alpha’s model consists of a two-hour learning block in the morning with a personalized AI that teaches kids all the academic subjects they need (at twice the speed of traditional schools, so they claim). The rest of the day is spent in real-world, practical “workshops” where kids do creative projects (many of them using AI).

Joe called out the paradox: AI is being used to sharpen kids’ brains, but in the next hour it turns into a tool that shortcuts the path to their goals.

Thus the most important question becomes the one we answer in our heads: what do we want to get good at?

If I want to get better at fast calculations, then I should resist the calculator.

If I want to get better at completing work I don’t care about, then the calculator is exactly what I need.

There is no universally right or wrong answer here. The only way to succeed is to better know ourselves.

Some recent diary entries

For paid subscribers, my personal diary entries that inspired this essay.

Read more

Muse launched earlier this month, and it’s the first consumer agent I’ve used that makes me think people might actually change their habits for it.

I’m not the only one who thinks so! The X-o-sphere is ablaze with (grudgingly?) positive reviews. In his New York Times review, Eli Tan called Muse “the most useful AI app I had ever used” (though he also questioned how many people would trust it with their personal data). The app has shot to #1 on the App Store charts, and early indicators show stickiness among the early adopter crowd (though really, as any good analyst knows, it’s way too early to trust week 1 retention numbers).

I want to break down what makes a consumer agent appealing to more than techies, and where Muse gets closer than ChatGPT or Claude. (I haven’t used Instinct yet, so if anyone wants to provide me an invite, I may be tempted to do a similar type of critique!)

I’m grounding this in my week-and-a-half exploration, which had me going down vintage DVD rabbit holes, emailing an obscure director, and happily giving Meta the keys to my personal email.

Keep in mind that while my use cases are definitively not average, the principles they illustrate are generalized from decades of working on consumer tech.

Let’s dive into the four principles of consumer app design…

Principle 1: Make it approachable (aka the “mom test”)

Could you hand this app to your mom and have her figure out the mechanics of how to use it without sitting her down with a whiteboard and having her learn a bunch of new concepts like “AI models,” “Cloud”, “VM”, or “Plugins?”

Approachability simply boils down to: does this experience match my previously existing mental model for how something works?

The first big “mom test” design innovation came from ChatGPT with the chat interface, which I previously wrote about. Everyone knows how to text already!

Muse adds a few more touches to increase its approachability and build upon most folks’ existing mental model of how a helpful human works.

The first is personalization of the agent in the form of this yeti-like creature (I call mine Moggy). It looks friendly and sincere. It is soft and cuddly, not at all suggesting I am a bot. Whenever you give it a task, it immediately starts plugging away with what looks to be great satisfaction and diligence.

A second nicety is that whenever I type anything, it immediately likes or emojis my response, just like an attentive human would.

The third is its length and language. No load-bearing arguments here; the tone is short and friendly, entirely appropriate for a text message.

Lastly, there is only one “main” conversation, which I thought was an interesting design choice. Though you can have ‘side conversations,’ the main thread is the default view and labeled as “main.”

Typically you have only one thread with a friend even as it encompasses many different topics, so this mirrors what we are already used to. (There are some costs to that choice, which I’ll get to later, but my hunch is that this was done in the spirit of approachability.)

Principle 2: Make me feel I can trust it (aka the “spare key test”)

Imagine a stranger asking, “Hey, can I have a spare key to your apartment?” Hell no, get away from me, you creep!

Now imagine the same stranger asking, “Hey, can I have a spare key to your apartment?” after you have booked his plant-watering services because you’re going to be traveling for the next month. Of course, here you go.

People are willing to hand over access when the context is appropriate, and when they have opted into getting something valuable in return.

Of course, I’d still evaluate whether the plant-waterer seems trustworthy (does he look legit? Friendly? Knowledgeable about plants?), so all the things we talked about in Principle 1 also apply (good design leads to more trust, as Patrick Collison has explained).

Muse, to its credit, asks for nothing up front and seems to follow the rule of “ask only for something sensitive when the user has something to gain.”

A great example of this was when I asked it to create an account for me on a new website so I could purchase an item.

The website sent over an email verification code. Muse then asked me, “do you want to type in the verification code yourself (by going through the browser link), or connect your Gmail account so I can do it for you?” This was a perfectly executed ask for the keys to my private inbox!

The same thing happened with my credit card. By the time Muse had come back to me with the full product, amount, and receipt, saying ‘buy’ and linking up with Link (which I had used before on many a merchant site) was easy-peasy.

3. Give me outsized value (aka the “I’ve been putting this off” test)

Most Silicon Valley techies, like economists, looooove efficiency. So when something purports to do X but like, 10 times faster, that’s highly appealing. So they tend to believe EVERYONE in the world is as psyched about efficiency as they are.

Alas, this just isn’t the case. If you pitch me faster online shopping, I’m not impressed. Going to Amazon for most things is already fairly efficient, and digital window shopping is actually fun!

That said, there are unsexy tasks, the ones you grumble about and put off until the last minute, and this is where a good agent earns its stripes.

My personal example deals with international treasure hunting.

For years, I’ve coveted a lost childhood relic: a set of cassette tapes of an obscure 80s anime called Triton of the Sea, dubbed in Mandarin. It was one of my few possessions when I immigrated from Shanghai to the US at the age of five. I remember lying awake at night, listening to it over and over again, achingly homesick while trying to hold on to this last remnant of my lost world.

As you might imagine, Google search could not help me with this truly obscure request. To make progress, I knew I had to slash my way through the brambles of Chinese auction sites and forums, all without knowing how to read the language. The tools were there (I could Google Translate my way across this endeavor), but I could not muster up the energy needed for the hunt.

So, you can imagine my utter delight when I gave this quest to Muse. In its first few hours, it had no good news for me. But a few days later, I got a notification: a seller from Shanghai had just listed it on some Chinese auction site, and it looks to be in Mandarin. Shall we go for it?

Damn right we should! Muse then discovered that this site only ships to Chinese residents, but then found me a workaround: I could buy it through a third-party middleman that would purchase and then ship it to me. I didn’t even know that was a thing!

At this point, I was feeling exceedingly impressed. The one hiccup was that the middleman site had some captchas Muse couldn’t figure out, and the ‘go into the browser yourself and fix it’ step, while manageable, was annoying. But the promise of the prize was worth it! I’m eagerly awaiting my dose of nostalgia.

There are many, many other “put it off”-type tasks in life, and each of us has our own pet collections: Filing for refunds. Filling out paper forms. Corresponding across languages. Booking an appointment. Researching the best option.

When an agent does something you can do but faster, that might just get a shrug or even a frown.

But when it does what you hate doing seamlessly, that’s a true aha moment that turns skeptics into fans.

4. Give me a sense of possibility (aka the “bored in line” test)

The reason Instagram, TikTok, X, YouTube and Facebook became so crazy prevalent was because they killed it on the test of: does this entertain you if you’re bored waiting in line?

They killed it precisely because they made it take so little effort to avoid an uncomfortable moment. These winning consumer products also follow the rule of “No dead ends,” because a natural stopping point where the user has no more good, easy options to do something interesting is a moment where they shut off the app.

Contrast this with ChatGPT or Google’s page today: blank! Which means, the user needs to supply the intent. This is a high bar when you find yourself with a few unexpected waiting moments!

Muse’s solution here is two-fold.

The first is that they put in a tab with a feed of “content you may be interested in” which seems to be stories generated every half day or so, seeded by the contents of the conversation I’m having with my agent. It’s not surprising to see this from the company who put algorithmic feeds on the map, but this particular decision feels odd. This Muse feed is much less relevant than the ones I already have in FB or IG. If Meta wants a good “interesting content” feed, why not ask it to import what it knows about me from those apps?

But stepping back, do we really want a feed in my personal AI app in the first place? If this feed gets amazing over time, wouldn’t we just be shifting time spent from one place to some other less optimized surface? (To be clear, one of Meta’s massive advantages is its distribution network. So we should absolutely expect to see cross-references across its apps in the future, a la seeing your fave celeb in an outfit that slays and having the option to send it to your Muse to find you similar in your price range.)

On the other hand, Muse does a brilliant job with Ideas, which has its own tab and also gets dripped into the main chat in the form of suggestive offers from your agent (“Want me to draft a response to this email?”) Ideas is an effective antidote to the Blank Slate Problem I wrote about a year ago.

Like the feed, it gets more and more personalized over time, though even their non-personalized version is engaging. (One of the options is: “Want me to see if there is free money available to you that you can collect?” Don’t mind if I do!) I appreciate also how it separates into categories like finance and health and shopping to give you a sense of its range.

The personalization is surprisingly great. I asked Muse to help me figure out how to acquire a fantastic foreign movie I watched on an international flight last year, and one of its ideas was “hey, the director recently just gave an interview about this movie, shall I translate it for you?” Why yes, thank you! Another idea was to help me plan a viewing party for this movie. Cool idea I hadn’t even considered!

Of course, recommendations live and die by their relevance and I did get a few irrelevant ones (“Facebook just emailed an ALERT of someone trying to log in unsuccessfully! Was it you? Better change your password!” Calm down, Moggy. This happens like every other day.) But I can imagine this only gets better with time.

How can Muse go further?

1. The biggest challenge in marketing a “personal assistant” is that most people (the vast majority of whom haven’t had a human personal assistant before) aren’t aware of all the things they could do with one. There’s a running joke in Silicon Valley that all tech demos seem to breathlessly feature tasks like “plan my vacation” or “book me a rezzie at my favorite restaurant,” which are NOT high-frequency actions and read more like vitamins than painkillers.
2. Towards that end, I’d love to see Ideas closer to the default experience. If the hard question for a noob is “What am I supposed to do with this?”, the marketing and the very first touch point should try to address that. “Saving money” and “Help me get what I want for cheap” are good places to start!
   
   I wouldn’t be surprised if suggestions became the primary landing place over time, with chat becoming the place you go after choosing something you want to do.
3. In the best-case scenario, when one is using Muse in a highly engaged way, keeping track of multiple threads becomes unwieldy. It feels like one of those active threads where as you’re responding to one topic your friend is simultaneously talking about another, and you can’t easily catch up without doing a lot of scrolling up and down to find the old thread and catch the responses.
   
   Muse has side chats, but the name ‘side chats’ felt vague and the manner of using them non-obvious. Ongoing work should be easier to keep track of. Personally, I’d like to see the current goals tab combine with side threads / activity log and show a list of ‘active work’ which lets you branch off into separate chats for follow-ups on that work thread.
   
   The test here would be whether people with multiple tasks going on can return after a day away and pick up the right thread without rereading everything.
4. Integrating with what I’m already engaging with on Instagram or Facebook to improve my recommendations. (Please ask me first if I’d like that, and sell me the benefits!)
5. Gifts. Making cool stuff as personalized gifts for others is an underutilized bucket of possibility! In fact, AI for fun and leisure as a whole is undermarketed, which I wrote about last week.
6. Social / multi-player. This one is such a rich topic that I’ll probably write its own follow-up essay.
7. An actual Moggy desktop companion. Bonus points if plush. Maybe you can sell crocs and a mullet as extras.

Meta Connect had a few fun announcements (voice and video and a Moggy keychain). There are rumors OpenAI will debut a consumer agent (perhaps with social?) shortly. Let’s get out our popcorn and see how it all unfolds.

All in all, very excited to see the future of this arena. A week and a half in, Muse has me waiting for a set of childhood tapes and thinking about what else I’ve been putting off.

Disclaimer: This is my first review after a week and a half of use. I had no advance knowledge or early access. I do know some of the folks on the design team behind this (hi Luke Woods! Alex Cornell!), though I haven’t spoken with them about this project.

Bonus for paid subscribers: a closer look at the design details I liked and disliked

Here is, in no specific order, a list of other random design details I appreciated:

Read more

Out of all the lives that have been lived throughout history, if you drew one at random, who would you be?

This was the premise of The Veil of History. In just a few clicks, you can determine where and when you would have lived, and whether you would have been a prince or peon. (Spoiler alert: you were probably a peon with a short life.)

Ethan Mollick made it with Claude Code, and it was one of the many cool projects I’ve stumbled upon made far more accessible with AI.

In my recent essay on AI transformation, I made the point that AI transformation really starts to happen when you change your personal relationship with AI. Instead of bemoaning its capacity to take your job or your creative talents, why not think of it as your own personal Chorebot2000 and give it all the digital-world tasks you hate: filling out fiddly forms, following up, finding some needle in giant, unsorted piles of hay?

But what comes after that? Once we’ve expanded our day as a result with less Chorework, what do we fill it with that is exciting, creative, and high-agency?

Last week I wrote about one answer: building your own hyperpersonalized software.

Merriam-Webster gives one definition of a tool as “a means to an end.”

We spend a lot of time comparing and optimizing our means. Claude or Codex? Instinct or Muse? Kimi or Qwen?

Today, I want to spend more time talking about the ends: how can we use AI tools to get more out of leisure?

Leisure, to many, is not a “good” reason for AI, especially when the stakes seem to center around whether you’ll be a part of the permanent underclass or who will win in global geopolitics or whether there is a 10% chance the human race will be destroyed.

But I can think of no better reason for leisure than the idea of agency itself: the drive to create that comes from a personal sense of freedom and choice.

I break it down into four buckets:

1. Hobbies++
2. The pot at the end of the what-if rainbow
3. Future doses of delight
4. One-of-a-kind gifts

Let’s dive into each!

New here? Subscribe for future essays on the art of building!

Bucket 1: Hobbies++

The summer after my freshman year, a fellow intern introduced me to Catan, the board game where you gather resources and compete to trade and build settlements. I became absolutely obsessed. Apparently I am not the only one!

Meng To used Astra to make a playable version with animated pieces and computer opponents so you can get that ore and wheat even when your friends are unavailable!

What about understanding the inner workings of things, like what makes a smooth muscle car go?

Dilum Sanjaya asked Astra for an interactive model of an eight-cylinder engine. You can see the pistons moving, peek closer, and change the speed. So much richer than reading a block of text and seeing a static picture with labels!

I love seeing people take their hobbies to the next level with AI. You start with something you already like, whether a game or a topic of interest, and then you explore it further, ideally pushing on its creative boundaries.

For example, my childhood hobby was drawing. Now, I was proud of my skill, so replicating my illustrations with AI wouldn’t have excited me. But what if my drawings could move, or morph into playable characters in a game?

Here’s another example that inspired the fashionista in me: cdngdev had AI identify all the clothes from his camera roll, inventory them into an online closet, and give him more styling options. It’s like a whole new wardrobe without having to buy a single new item!

I also loved this Pokémon Go for real cats by Sebastian Seidel, mashing together three real-world hobbies: felines, exploring a neighborhood, and collecting. Each cat you find becomes a collectible complete with a name, rarity, and stats. Turn your walks into side quests!

Bucket 2: The pot at the end of the what-if rainbow

What if AI agents existed in the 1970s, but you had to call them on a rotary phone and see their results on a mechanical split-flap display?

This is exactly what Karen X. Cheng wondered, and now we have an answer! It’s pretty delightful.

I would never have thought of it, but after seeing it I started wondering: okay, so what other unobvious things could we connect together?

Here’s another fun one: what if growing your website traffic was like growing a village? DataEmpire explores that question, with visitors becoming villagers who chop trees and build houses. Isn’t this way more fun than a dashboard?

There is a certain deep pleasure that comes from seeing a weird idea come to life. I remember this feeling when I was a kid with a box of legos (yes, this was the Jurassic period before legos became inseparable from “sets”).

The game was to explore the what-ifs: What if that three-story building was also a car? What if a group of pirate-knights had a pool party? What if your cat lived in a cat-shaped dwelling?

This is the same feeling that building with AI gives today, a curious tickle that eventually expands into a notion: well, why not try?

I caught a whiff of this feeling when rpnickson fed Astra nine photos of his studio and asked for a little 3D world he could explore.

Instantly, it made me wonder: what if we did the above with a childhood home, or a place that only exists in a drawing?

And then, the what-ifs kept coming:

What if you could order ramen with the press of a button?

What if plucking a certain ukulele chord could transform your living room into a disco party?

What if the cat really could live in a cat-shaped dwelling?

What-if possibilities are endless, and perfect for a weekend afternoon.

Remember: the aim isn’t about getting some specific result (although the pot at the end may very well be filled with something shiny).

It’s about the adventure of simply going on a little jaunt with curiosity.

Bucket 3: Future doses of delight

You know who you should try to coax a smile out of?

Future you.

That’s right, make yourself a gift that keeps on giving!

For example, someone built this system that listens for birds outside their apartment. It identifies the species, and then turns those birds into an artsy illustration within a frame. How cool is that?!

Imagine going from your office to the kitchen and stopping to muse at the symphony of birds that have alighted around you. What a delight!

I have my own build story to share here too. I’ve always been a sucker for “on this day” type of posts, whether enjoying those bite-sized Apple videos made from past photos or reminiscing on a throwback Facebook post. But I kept thinking, “these could be so much richer!” Decades of being a compulsive Dear Diary girl and the One Who Shows Up To a Party with a Fancy Camera have left me with a lot of material.

So I built myself a daily loop that goes through all my old photos and videos (compiled from both my fancy camera and my phone) as well as my past journal entries, and compiles a story from the same month in my life in the past. Every morning, it WhatsApps it to my husband and me and we both love reading them and getting transported back to another past moment in our lives.

Bucket 4: One-of-a-kind gifts

One of my friends is an avid painter and I’ve been lucky to be gifted some of her watercolors to display in my office. Meanwhile, my kids’ drawings decorate the walls leading to our bedroom. None of this is professional work, but I cherish them because of the people who made them and I know I’ve got the only one of it in the world.

In the era of AI, there is so much blue sky for creative expression in the form of gifts!

Case in point, my son just turned 10, and asked us for a Geometry Dash birthday party. If you aren’t familiar with the game, it’s a ridiculously unforgiving single-player web platformer where you are a cube trying to avoid triangles and other dangerous shapes while jumping to the beat of catchy EDM songs.

My husband Mike, who goes ALL OUT for party planning, decided that he did not want to host a party where a bunch of kids brought their iPads and played this game solo. Alas, Geometry Dash does not offer multiplayer, so Mike decided to build his own multiplayer version personalized for our son. You can play as various cubes in the theme of my son’s other favorite hobbies: Pokémon, Minecraft, Rubik’s Cube. His levels also come with songs composed with the help of AI featuring a ton of insider family jokes (Healthy Gaming is one of my faves!)

I saw another example of personalized gifts when a father whose son loves Mario built a game starring the son, his siblings, and his cousins. You can choose which family member to play!

Instead of gifting an Amazon buy that thousands of other people already have, what if the gift is the experience of making something that only you can make for only that person?

This week, make something just for fun

If you’re still reading, I hope you’re feeling inspired! Maybe inspired enough to try the below?

Start with a person you love, or an activity you enjoy, or just an odd idea you keep coming back to. Then, complete this sentence: “It would make me happy if _____.”

Make it specific enough that you can visualize the results.

Then make a single request of AI: It would make me happy if ___________. I’m imagining it could look like ____________. Can you help me do that?

Hit enter and see what you get back. Tell the AI where it got your vision right or wrong. Keep iterating. Maybe include family members in your exploration.

You may discover magic in the process itself. You may find yourself with even more ideas after the first one. You may inspire someone else to make their imagination tangible.

Whatever it is, I hope it feels like play!

Free subscribers get every essay. Paid subscribers also get bonus and more personal content. Either way, thank you for reading!

38 ideas I’ve been inspired to make

One more thing, for paid subscribers: my own list of 38 things I have been inspired to make, sorted into the same four buckets.

Some I have completed, some are in progress, and some are still on my list (which keeps growing ever longer!).

Read more

i.

In 1975, Andy Warhol made the observation that “a Coke is a Coke, and no amount of money can get you a better Coke than the one the bum on the corner is drinking.”

Marc Andreessen referenced this in 2023 and threw in an extra line: “Same for the browser, the smartphone, the chatbot!”

That’s kind of cool in some ways, right? The richest and most powerful people in the world are opening the same Slack as we do. They’ve got the same smart phones. Their Duolingo owl sends them the same guilt-tripping Your streak is in peril! notifications.

Sure, with more money, you can buy a private chef, a private jet, a private yacht. But when you sit down at your laptop, you’re using roughly the same software as every other fresh-faced college intern.

That was before. Going forward, I no longer believe this holds.

ii.

Let me back up a bit.

If you have multiple young kids, like I do, you will be familiar with the famous challenge known as the “Mommy, who do you love more?”-barrage. After being repeatedly humbled by this judo move, I have learned that the best method is to sidestep the question entirely and actually answer a different one: tell each what I love about them specifically. It seems to work: after all, we humans love what’s unique to us. In 1959 a psychologist named Neville Moray played two streams of speech into people’s ears. They were told to just focus on one of the streams, but a third of them still heard their own name in the stream they were supposed to be ignoring. Dale Carnegie once wrote that a person’s name is, to that person, the sweetest sound in any language.

Personalization has been commercialized ever since. Coke, if you can believe it, had a period where they printed your first name onto their bottles. Nike lets you pick the color of your own swoosh; Louis Vuitton advertises more than 200 million monogram combinations.

And it’s not just names and colors; entertainment is now personalized to your tastes. In the 1980s, a single show could still command over 70% of the TVs in use. Today, streaming is almost half of all TV time and sliced into the thousands of niches that make up the long tail. A good three-fourths of YouTube watching starts from a personalized recommendation.

AI knows us better than ever. That trend is only continuing.

iii.

And yet, most software today, with the exception of personalized social, feels rather uniform. Duolingo, with its 58 million daily users, personalizes how hard the next sentence is, but not what those sentences are about. My seven-year-old loves the app but let me tell you, he is never going to tell someone “My, your kitchen looks so lovely!” in any language.

Where software played with personalization was largely in the settings menu. We got dark mode, color themes, maybe a font size slider here and there. But settings operated like a rule book. Somebody decided in advance the small handful of things one would be allowed to change. Designer Jared Spool once asked several hundred people to send in their Microsoft Word settings files and discovered fewer than 5% had changed a single setting. And if what you wanted wasn’t in there, you’d send a support ticket or beg for it on the forums.

This model made sense because software used to be expensive to make, so it had to be made for a lot of people at once. Clay Shirky wrote in 2004 that building an application for a few dozen users was “an absurd target population.” Twenty-one years later, it’s time to rethink what’s possible.

Subscribe to The Looking Glass for essays on building things, and on what happens to design when everyone can build.

iv.

Here is the story that changed my mind.

I run a lot of AI agents at once, usually across multiple projects both personal and professional. I work in the command line for the flexibility and power of it. My typical set-up is eight terminal windows open on my home machine, and I arrange them by hand like a poor game of Tetris. While on the go, I used an app called Mosh to control my terminal windows remotely, but the entire setup was six clicks deep. Meanwhile my husband was using the remote control version of Claude and Codex, and I jealously watched as he attached screenshots and two-tapped between his agent tasks. I was gearing up to trade my homemade setup for the more polished third-party version, despite having to give up some of the flexibility of my setup.

Then a thought crept into my mind: could I build the thing I was envying?

Half an hour later, I got a mostly-working version that answered my question with a resounding yes.

My app: two-taps to switch agents (real interface, not-real data)

With the core questions answered, I spent the next few days elatedly designing for myself, adding in little moments of delight that would make sense to very few others.

I wanted to know how many subagents were working at a time by looking at the number of sparkles. I reinstated Claude’s status verbs with my own twist. I wanted more clarity on what subagents were doing. I called this app Mog Squad. If you know, you know, kupo.

one session at work: the General delegating, three workers doing their thang.

Two days later, satisfied with my mobile app, I turned my attention to the desktop version.

So many sweet, sweet threads of work.

After this, I thought to myself, I never want to use third-party software again. I don’t mean it boastfully. I don’t think my version is better than the official apps. But it works better for me.

For one thing, it lets me use both Claude and Codex, and neither official app does that. It’s hyper-optimized for my workflow, with my eight agents arranged in a grid. (Why eight? Because it’s the ideal number for me. Fewer, and I get the itch to start something new. More, and I start losing track of ongoing projects.) It has none of the bloat from features I don’t use. Because I created it, I know exactly how every button and drawer and keyboard shortcut works. When I find a little friction point, I remove it. When I get a new idea, I implement it. The feedback loop could not be more efficient. Because I am my best customer, I care immensely about making myself happy. I will never complain about my own software!

Economists even have a name for this feeling, the IKEA effect: we love the things we assemble ourselves more than the ones that come pre-packaged for us.

v.

I’ll share one more example of hyperpersonalization.

My kids love Duolingo, so over the summer I built them a Duolingo-inspired math and reading app.

Now, like many kids, my kids are not especially motivated to answer math and reading questions. And I didn’t want to resort to state-of-the-art gamification techniques like streaks or “you got a 2x boost for the next 15 minutes!!!” So I tried a different tack: the stories would be about their lives: their real friends, their real life news, and their real hobbies. My daughter’s stories are about fashion and theater. My son gets Minecraft, Geometry Dash, and a large dose of Pokemon. When we take a trip, those stories show up in next week’s lessons. And just as I used to sneak veggies into their meals, I make sure the stories carry a sprinkle or two of our family’s values in what the characters choose to do.

A story about an arcade trip (names changed to protect the innocent; the raccoon is real tho)

There is a “feedback” button at the end of every lesson so the kids can tell me if it was too easy or too hard, or if they want a particular story (my daughter has used it more than once to request a story in which she meets and gets to sing with Idina Menzel.)

One request was not about the stories at all but about the rewards. They wanted something to collect. So I made holographic cards of their favorite stuffies. The rarer the card, the shinier it is when you tilt it!

Rooney the raccoon is now immortalized in virtual trading card form.

There is some research beneath this: in 2013, a study gave 145 ninth graders algebra problems. Half were standard and half were rewritten around each student’s own interests. The personalized group solved them faster and more accurately, with the gain being largest for the students who were struggling. Best of all, it stayed after the personalization was taken away!

This makes sense to me: even as an adult, I pay more attention and do better work on the things I’m actually interested in.

vi.

So what did I learn from these experiments?

Hyperpersonalization is the next great trend. And there are two major reasons for it:

The first reason is workflow utility: anything you do fifty times a day is a place where friction compounds. Two clicks instead of one. A menu with 20 items rather than 2 that your eyes have to process. Two screens you need to flip through to remember the context. When I wrote last week about how to pull off an AI transformation, I made the point that workflow improvements are great early wins. And workflows are deeply personal, influenced by our unique brains (are you an audio or visual learner? Do you like tl;dr or details?), lifestyles (are you constantly on the go, or do you sit in front of a computer all day?), and the constellation of tools at our disposal. Even where best practices exist for a single tool, there are hairy seams between tools because vendors are not typically incentivized to make sure every possible combination of tools works beautifully. The only people who can describe a perfect workflow are the ones living it.

The second reason is expression. I have a closet full of shirts that perform the same utility in keeping me warm, but I’ll still have favorites that I think express myself better. Software has tended towards utility: most software (especially enterprise software) is bland, most UIs converge (can you even tell the difference between the main chat screens of ChatGPT or Codex or Gemini?), and I get it, bland is safe, bland is least-likely-to-offend when you’re creating for millions. Turn your agents into a flock of winged cat-creatures and a lot of people will reasonably be like WTF? But hey, it puts a smile on my face. And I’d hope your own software tickles you too.

I think this is where interface design goes next. When I wrote about conversational interfaces last year and talked about the golden opportunity of personalization, I was thinking from the perspective of a company learning you well enough to customize what it shows you. I now think the inverse is also true: enabling you to tune the thing for yourself, whether that’s UI or content.

A lot of people will not do this, at least for some time, and that’s fine. Just like not everyone cares about fashion or eking out that last ounce of optimization, not everyone will care about perfecting their workflows or expressing themselves through software.

But enough will. And no software is safe from the personal remixes to come.

vii.

I think about software now in levels.

At the left are the use cases. This is where most software lives today, because use cases are tangible. An app that creates photo collages or helps you tune your guitar or lets you watch movies solves for very specific problems that need only a yes-no answer: Does this solve a problem I care about?

One level up from that is a platform builder like Shopify, Webflow, Roblox or Lovable, that helps you create a personalized version of a certain type of thing: whether an online store, a website, a game, or an app. Here, you need some imagination to picture how it might look and feel, and what functionality is important. But there are still guardrails around how much you can build or change.

One level beyond that is the anything builder: a coding LLM with a full set of tools and permissions to create whatever your mind conjures up.

Each step from left to right requires more imagination and delivers more freedom. The far right is where few live today. But slowly but surely, everything to the left will start shifting more and more to the right. That’s why the Replits and Lovables of the world are growing the way they are. If operating systems are smart, they’ll get in on this. It’s hard to imagine that an AI-first operating system of the future will look like a grid of apps.

If you build software today, I’d love for you to consider four things:

1. Make your interface malleable. A UI should be something a person can take apart and put back together in the way that best suits them.
2. Let preferences reach past the rule book. A descriptive vision is worth more than fifty toggles.
3. Let feedback change the product. Give people a chance to tell you what they really want, and immediately make it so.
4. Treat the seams as part of the experience. Nobody lives in just one tool, so a better workflow is a more integrated experience across the tools a person already uses.

viii.

Cokes remain popular today. They are still cheaper and more efficient than making your own drink. Perhaps one day, when personal robots can prepare ambrosia to our specific tastes, this will no longer be true. Until then, we’ve got software to play with. Happy remixing.

Still here? Subscribe to The Looking Glass and the next essay will be in your inbox with your unique and personal name on it.

For paid subscribers: the rough steps I used to create my two apps

Not a full tutorial, but hopefully gives you a sense of how I got started.

Read more

We have heard it so often that it now occupies the space between meme and cult mantra: AI is going to change everything!

Markets are pricing in the cheap, cheap cost of building (goodbye, SaaS giants!). CXOs (chief executives of every stripe) are wringing their hands and sprouting AI investments like dandelions in a field. Experiments are running amok. Productivity has 10xed, or is it 1.1xed, or is it just spending that is x-ing at all?

Being an employee at the center of an “AI transformation” feels like whiplash crossed with Wile E. Coyote chasing the Road Runner. Impossible to keep up. Claude is the best thing since sliced bread! No wait, Codex is king! No wait, the future is open source! And my security team still has not greenlit a frontier model. (True story I heard last week.)

These stories have played in my head for a few years, exploding into a Times Square-level cacophony this year. My company, Sundial.ai, was formed before ChatGPT. Our sundial.so email addresses are relics of ye olden days. We had to cross the chasm into an AI-native way of working ourselves.

Working with our customers, I have spent hundreds of hours with everyone from the savviest post-AI-boom startups to decades-old Fortune 500 companies. I am just as riveted by how adoption happens as by what the technology enables. Even if AI improves zilch for the next decade, we are still in the Stone Age of using what already exists. The bottleneck is our human capacity to adapt.

What follows is the clearest path I know. Build personal belief. Turn belief into shared wins. Turn wins into reliable systems. Then redesign the organization around what has become possible.

This is for leaders and for the individual contributors (ICs) on the front lines. Both have real jobs to do, and they are not the same job.

A note before we start. This is long, because it is the whole path rather than one idea about it. You do not have to read it in order. If nothing has started at your company, Part 1 is where you are. If you have a couple of wins and no idea how to make them add up, skip to Part 3. Every part ends with what a manager can do this week and what an IC can do this week, and if you only have five minutes, read those.

What is in here

• Part 1: Discover positive belief. Step 1: automate something you hate.
• Part 2: Get some tangible wins. Steps 2 to 4: automate a small, unloved team process, goal on the outcome, make the win loud.
• Part 3: Turn wins into systems. Steps 5 to 7: map the work, earn trust with evals, refactor what repeats.
• Part 4: Recompose for abundance. Steps 8 to 10: tell an abundant story, find the new bottleneck, change what the organization rewards.

Subscribe to The Looking Glass for essays on building things, running teams, and what AI is doing to both.

Part 1: Discover positive belief

I see this pattern over and over. A CEO reads that competitors are cutting costs with AI and blasts the company with a directive: “Become AI-first or get left behind.” Everyone scrambles, because fear makes for great ultimatums. By the time it reaches middle management, the directive has congealed into a mandate with key performance indicators attached. Now the manager’s job is to look like the team is transforming. So they buy the licenses, schedule the trainings, and put “AI” on the roadmap.

It does not work. Fear makes people optimize for appearances instead of spelunking the true edges of transformation. The prize is not “yay, we cut 30% of costs” or “hooray, we are not getting left behind.” You will never win the popular vote on those. The prize is a way of working that is more creative, more impactful and, dare we say it, fun.

That is the hypothesis worth chasing. CEO, vice president, middle manager, IC: you have to stay open to the possibility that AI gives people more room to create, decide, connect, and do work they care about. You do not have to believe it yet. Skepticism is fine. Let the possibility marinate like sweet, savory teriyaki readying for the grill. You only need to loosen your guard enough to find out.

Step 1: Automate something you hate

The first experiment matters. Most people sit down to “try AI” and ask ChatGPT or Claude to do a part of their job they like and are good at. I used to lob over prompts such as “design me an app that…” or “write me an essay that…”

Two predictable reactions:

A) “HAH, this thing sucks. It is nowhere near as good as me.”

B) “SHIT, this result is actually passable. What does this mean for my future?”

Neither one helped. Do not begin a diet by giving up the foods you love.

Instead, stop doing what you hate.

I am the opposite of a detailed person. Remembering where I need to follow up is my version of the third level of hell. One of the first things I built was a little system that pings me when I have left something unanswered too long. Then I thought, you know what else I do not love? Drafting the follow-up message. So that went next. Then the forms. I would rather do burpees than fill out long, fiddly forms.

Piece by piece, the things I am not into started taking less of my time: contracts, balance sheets, research, to-do lists, digital dumpster diving. My excitement grew in proportion to my relief, and my ideas got bolder.

Aim AI at the chores instead of the craft and the whole thing flips. It stops being the thing coming for what you are good at. It becomes the thing clearing junk out of the way so you can do more of it.

You cannot argue anyone into that flip. They have to feel it.

I was at an executive roundtable where the conversation about AI was bleak. Everyone was being squeezed to do three times more with half the time or people. Then someone asked, “What has been a good use of AI in your life?” One vice president after another told the same story: “I hate writing status updates. Now AI does most of it and gives me back my Friday afternoon.” The mood in the room shifted.

A leader cannot mandate enthusiasm. A leader can arrange for relief.

What to do

If you manage a team

• Automate one thing you hate this week. Show your team the result.
• Ask each person to name one task they dread. Give them time and cover to make it disappear.

If you are an individual contributor

• Pick the most tedious part of your week. Work or home, either counts.
• If you do not know where to start, ask the AI itself: “I hate doing [task]. Give me a step-by-step plan to automate it.”

Easy starting points: choosing a restaurant for date night, loading a school calendar, summarizing competitor news, turning a grocery plan into a shopping list, drafting a follow-up.

Part 2: Get some tangible wins

Once a few people have felt that relief, turn the feeling into a win on real work. Do the same thing again, one level up.

Step 2: Automate a small, unloved team process

The magic words are team, small, and unloved.

Team, because personal automation convinces one person and a shared process convinces a group. Small, because you want a win in days, not months. Unloved, for the same reason as Part 1.

Meetings are a fertile place to look. They generate the same little pile of work every time. Beforehand, someone researches the people, digs up old context, pulls the latest numbers. Afterward, someone writes notes, updates a system of record, chases follow-ups. And nobody ever gets to whether the hour was worth it.

These jobs also have clear triggers. The customer call starts in three hours. The metrics review happens every Tuesday. The Zoom transcript just arrived. Triggers are what make a workflow suited to an agent: an AI system that runs several steps on its own instead of waiting for a person to prompt it at every step.

Here are three examples.

1. Prepare for a customer meeting

At Sundial, every customer call used to begin with the same scavenger hunt. Who is this person? What is happening at the company? What happened the last time we spoke? Someone searched LinkedIn, asked ChatGPT or Claude about the company, dug through the customer relationship management system (CRM), and reconciled it all before the call. Nobody enjoyed updating the CRM and Slack afterward either, so the context was missing again next time.

One of our teammates built an agent named Dwight, after the character from The Office. Yes, it had Dwight’s personality. Why not get a few chuckles out of it?

A few hours before each meeting, Dwight read the person’s profile, researched the company, pulled our old CRM notes, and delivered a sharp brief. The reaction was immediate: “Wow, this is great. It saved us so much time!”

Nobody needed a speech about AI transformation. People started imagining what Dwight could do next. Prepare a sales deck? Run prospecting? One useful agent created both relief and appetite.

2. Prepare for a weekly metrics review

At many companies, one or two people lose a day or two every week getting ready for a metrics meeting. They pull data from several tables, reconcile numbers that do not match, investigate what changed. An analysis agent can do most of that first pass. If activation drops, it checks whether the decline came from one product, one geography, or one channel.

People still verify the analysis and decide what to do. But the meeting opens with “What should we do?” instead of half an hour arguing about whose spreadsheet is current.

3. Give immediate meeting feedback

I want to give feedback after meetings, and I would love to receive it. Alas, I remember far less often than I intend to. So we built an agent that reads each Zoom transcript and privately tells the organizer how much everyone spoke, where the conversation meandered, and what might have made it better.

The feedback is about 85% useful. I sometimes disagree with it. But an outside view while the meeting is still fresh has helped me calibrate.

Do not start with a board meeting. Do not automate your one-on-one unless both people hate it. Do not renovate the kitchen while hosting Thanksgiving. Start with the junk drawer: a status report, a metrics review, some other low-drama weekly grind where stumbling is cheap and a win will be celebrated.

Let the most curious, AI-positive person lead. If five people attend a meeting, only one needs to reinvent the workflow, and it may be a junior IC or someone adjacent to the process. The manager supplies time, access, and safety. The IC brings the copy-paste, the edge cases, the workarounds, and the unreliable sources of truth that decide whether the thing actually works.

Step 3: Goal on the outcome, not the process

A win is a result, not a good performance of transformation.

You want to say, “We reclaimed eight hours a week,” or “We caught twice as many issues before release,” or “Customers get an answer in ten minutes instead of two days.”

The win is not “we used AI,” “we adopted a new workflow,” or, lord help us, “we maximized tokens.” Those are inputs dressing up for Halloween as outcomes. People care that the work got faster, cheaper, or better. The AI is incidental.

Saving time is a wonderful early outcome. Easy to feel, easy to measure, and it gives people back the one thing they never have enough of.

Step 4: Make the win loud

When a win happens, say so in public. Share the useful facts: here is the process nobody liked, here is what we tried, here is what failed, here is the result.

At Sundial, we made every question people asked Dwight public, so anyone could see what teammates were trying and steal the good ideas.

Each example gives the next person permission, inspiration, and a shortcut. Bank two or three wins and the mood shifts from “Is this even a thing?” to “What could we do next?”

What to do

If you manage a team

• Pick one low-stakes weekly ritual: a status update, a metrics review.
• Give one curious person time, access, and permission to rebuild it. Goal them on hours, dollars, or quality.
• Publish the result, failures included.

If you are an individual contributor

• Volunteer to rebuild one repetitive team task.
• Write down the old baseline first, so you can prove what changed.
• Advertise the outcome. Publish the recipe so the next person can steal it.

Still with me? Subscribe to The Looking Glass and the next one will find you.

Part 3: Turn wins into systems

A few wins prove a better way of working is possible. What you do not want at the end of the rainbow is a pile of one-offs: a bot that runs only on one person’s laptop, a team prompt approaching the length of a Harry Potter novel, three agents synthesizing three different sources of truth.

A company that uses AI has a few good workflows. An AI-native company can find, build, test, and scale a new one on purpose, again and again.

OpenAI’s enterprise research reaches a similar conclusion: “access alone may not be enough to scale AI.” The companies furthest ahead connect AI to their own context and tools, then turn one person’s working workflow into everybody’s.

Step 5: Map the work, not the job titles

Because AI sounds like a human, people assume it can take a human job. But asking AI to “do sales” is nonsense. What is sales, exactly? Preparing for calls, building trust, negotiating a contract, updating the CRM, talking the product team into fixing a customer’s complaint. A job is too broad to build or evaluate.

Workflows are the useful unit. “One day before a customer meeting, assemble our past interactions and relevant company news into a brief” is something you can understand, test, and improve.

Most workflows contain some combination of four kinds of work:

• Research: Gather context, watch for changes, investigate a question. Competitive analyses, meeting briefs, metric investigations.
• Create: Turn an idea into an artifact. Software, prototypes, designs, test suites, campaigns, lesson plans, policies.
• Align: Move decisions and context through people and systems. Refreshing the CRM, routing tickets, collecting approvals, replenishing inventory.
• Improve: Judge the work so the next attempt is better. Quality checks, coaching, meeting feedback, reflection, critique.

Every department has all four. To map one workflow, answer eight questions:

1. What triggers the work, and how often does it happen?
2. What information must be gathered, and can the system get at it?
3. Which steps are the same every time?
4. Where does the work require judgment, taste, or context?
5. What gets produced, and who relies on it?
6. How can someone tell whether the result is good?
7. If it goes wrong, how visible is the error and how reversible is the damage?
8. What outcome matters, and who owns it?

The best early candidates are frequent or time-consuming, guided by a repeatable playbook, easy to verify. Start there. Save the rare, judgment-heavy, high-stakes ones for later.

Managers and ICs have to find these together. I sometimes see a manager arrive with a checklist of “things to automate” and hand it out like homework. But the work a manager can see, the meetings and reports and deliverables, is a fraction of what keeps the place running.

ICs see the machine underneath. They know where people copy and paste between systems, which edge cases break the happy path, which workaround everyone depends on, and which source of truth is not truthful at all. Managers can compare opportunities across the team, move resources, set the risk tolerance. Neither view is enough alone.

Step 6: Earn trust with evals

You earn trust in an agent through evaluations, usually shortened to evals. An eval is a repeatable test of whether the AI did the job right.

Think of it as the agent’s driving test. Before you hand over the keys, define what good driving looks like and test it on roads you already know.

Some evals have right-or-wrong answers. A data agent might face a set of “golden queries,” questions whose correct data and calculations have already been verified. If it cannot reliably answer those, it is not ready to answer them for the chief financial officer.

Sundial’s writeup on what it actually takes to trust AI with your data puts numbers on this. Fifty real user questions, given to a state-of-the-art model sitting on clean warehouse tables with a semantic layer, scored slightly better than 80%. The same questions, once the missing institutional context was supplied, jumped to 98%. What changed was not the model.

Other work requires a rubric. A recruiter might score candidates on required experience, evidence of impact, and role-specific skills. Past candidate packets with known scores become the test set. You are not testing whether the agent writes like a recruiter. You are testing whether it finds the right evidence and applies the same standard every time.

Build the eval alongside the workflow:

1. Collect real past examples: routine cases, edge cases, known failures.
2. Define the pass criteria before you look at the agent’s answer.
3. Run every change against the same set: new model, new instructions, new tools, new context.
4. Pilot with human review and log the important misses.
5. Add each new mistake to the eval set so it cannot come back unnoticed.
6. Give the test set an owner who keeps it current and watches for regression, meaning a new version got worse at something the old one handled.

The eval is also how an agent earns a wider circle of agency. I think of it in three levels:

• Level 1: Advisor. You consult the AI, but you do the work. It suggests the points for a report you then write.
• Level 2: Deputy. The AI does the work under your authority, but a person reviews every action before it commits. It drafts the report, or preps a CRM update for approval.
• Level 3: Regent. The AI acts within a defined domain and reports what it did. Its authority is borrowed, bounded, monitored, and reversible.

Promotion requires evidence. An Advisor becomes a Deputy once it passes an offline eval set and people know its common failure modes. A Deputy becomes a Regent after a supervised live pilot, with monitoring and a fallback when something goes wrong.

The evidence required rises with the risk. An internal meeting brief is easy to inspect and cheap to correct. Sending a customer message, approving a discount, moving money, evaluating an employee: those reach farther and do more damage.

Trust stays local. An agent that has earned Regent status for sales briefs has earned it for sales briefs, and nothing else.

Step 7: Refactor what repeats

During the experiments in Parts 1 and 2, inconsistency is fine. Different people, different models. Two teams can solve the same problem two ways. Proof matters more than tidiness.

Eventually successful experiments start depending on the same foundations. Sales, finance, and support may all need the number of new customers each week. If each agent defines “new customer” differently, headaches ensue. Three workflows may all need calendar access, CRM history, and the current company strategy. If every team builds its own plumbing, the systems drift.

Think of it like refactoring a growing codebase. The first workflow is an experiment. The second reveals a pattern. By the third, the repeated piece deserves shared infrastructure.

At a bigger company, a small central AI team should hunt for those repetitions. The job is plumbing:

• Provide approved models, consistent data connections, logging, and eval tools.
• Notice when teams have rebuilt the same component, and replace the copies with one reliable version.
• Own the foundation. Let local teams own what they build on it.

Local teams pick the problem, design the workflow, own the outcome. A central group automating other people’s work from a distance will miss the details and get resented for it. The center’s job is to make the local team’s good idea faster to build and harder to break.

You do not need a perfect architecture up front. Let teams experiment, evaluate what works, refactor what repeats. Run that loop reliably and you are ready for Part 4.

What to do

If you manage a team

• Ask the team to map one or two workflows with the eight questions. Do not arrive with the answers.
• Name each workflow’s agency level. No promotion without an eval set and an owner.
• Keep a running list of agents, data sources, and shared parts. When one shows up three times, make it shared.

If you are an individual contributor

• Map a painful workflow your manager cannot see, edge cases and hidden workarounds included.
• Build the eval set as you build the workflow, from past examples whose answers you already know.
• Share the reusable pieces. Keep owning the outcome.

Part 4: Recompose for abundance

By now you have believers. You have wins that scaled. You have workflows growing in number and in ambition.

Now the hard part. Everything so far has been additive. From here you question the old shape of the work and decide what the organization should become.

Step 8: Tell an abundant story

Change is scary, and no amount of magical technology changes that. When people cannot see the future, they imagine the worst. The default AI story inside most companies is subtraction: “We will do exactly what we do now, with fewer people.”

That is the absolute lamest future imaginable.

You need a bigger story: given what we value and what we are good at, what is suddenly within reach that was unthinkable before?

One of my favorite recent examples is Midjourney, an early leader in AI image generation. When it described its next chapter, it did not promise a better, cheaper image model. It described a push into medical imaging: affordable body-scanning machines, custom wellness spaces. Whether or not that bet lands, I love the shape of the ambition. The falling cost of one capability became permission to chase a much bigger problem.

When a resource gets dramatically cheaper, people find far more uses for it. Economists call this the rebound effect, or, in its strongest form, Jevons paradox. An abundant story asks what your company will make, serve, or discover once intelligence and execution get cheap.

This matters emotionally, too. When a team grows, people feel a loss as they hand pieces of their job to new colleagues. Molly Graham calls this “giving away your Legos.” The loss feels worthwhile when you believe those people will build something great with them, and a bigger castle is waiting for you.

Handing work to AI is the same. It becomes growth only when there are larger problems worth turning your eye and hand toward.

I have heard some of the best people in their craft describe the moment the game changed underneath them. If you were one of the rare few who could do something almost nobody else could, watching AI hand that skill to everyone is jarring. There is grief in it, and a real “What now?”

Then many of those same people say, in the next breath, “What is amazing is that I can go horizontal now.” The best designer in the room can make a damn good marketing video or build the app they always imagined. Deep skill gains range. Taste and judgment travel from one domain to another, even as AI closes the gap on taste itself and agency becomes the part that stays ours.

Telling that story is the leader’s job. The story is the strategy, not a communications wrapper around it. Everyone can add to the plot, but leaders make the opportunity concrete enough that people see where they fit. “You will do more strategic work” is not a story. Name the customers you can finally serve, the product you can finally build, the stubborn problem the team can finally take on.

Step 9: Find the new bottleneck and recompose

As AI handles more work, a funny pattern emerges. You celebrate the win, pat yourself on the back, and a few weeks later everyone is grumbling about a new bottleneck. How quickly we humans adapt.

That is a good sign. Relieve a constraint and it moves. Your job is to keep finding where the work piles up next, and redesign around that.

I recently asked a senior leader at one of the biggest companies in the world where the bottleneck had moved after an aggressive push into AI workflows. He suspected human decision-making. The company had a handful of senior decision-makers, and too much work was flowing through them like cars into an understaffed tollbooth.

His next questions were different from the ones he began with. How do we help these people make better decisions faster? Which decisions no longer need to reach them at all?

Sometimes an agent can gather the context, compare options, and recommend. Sometimes you codify the principles so people closer to the work can decide. Sometimes the approval does not need to exist at all.

The person is rarely the whole bottleneck. More often it is the structure around that person. A leader at a Fortune 500 company described the problem this way: “We are very consensus-driven. You need people across different organizations to say yes, which means you need to know whom to talk to and what they will care about.” I have seen routine decisions take five weeks because they cross four approval layers and, during the summer, one approver is always on vacation.

Start by asking the people closest to the work where it waits. When their answers cluster, audit the workflow with real data:

1. Track the full life of the work, from request to result.
2. Separate doing time from waiting time.
3. Mark every handoff, approval, and repeated explanation.
4. Name the constraint: missing context, scarce judgment, unclear ownership, too many approvers, or not enough capacity.

Then redesign around it:

• If people wait for context, have an agent assemble and maintain it.
• If judgment is scarce, codify the common cases and save human attention for exceptions.
• If ownership is unclear, name who decides and what principles govern it.
• If there are too many approvers, remove the gates and move authority closer to the work.
• If capacity is still the constraint, decide whether AI expands it or whether roles and staffing have to change.

Notice what just happened. You started out redesigning workflows and ended up redesigning where the humans sit.

A founder told me that every new feature at her company had to pass through a central design team, which decided where in the app it belonged. She thought about hiring more designers. Instead the company carved out a self-serve experimental area where product teams could launch on their own, with the expectation that losers get retired.

Now only the clear winners graduate into the core product. Teams move independently, the designers stopped being treated as blockers, and the company runs far more experiments.

That is recomposition. She did not add capacity to the old machine. She built a different machine.

Recomposition changes who owns decisions, which skills sit together, and how many hands the work passes through. I have written about what this looks like inside product teams, where the engineer, PM and designer triad gives way to prototype-and-prune. Not a faster horse. An automobile.

I will not wave away the hard part. A recomposed team may need fewer people in certain roles. A veteran engineering leader at a startup described the vertigo to me: “LLMs let us create in a way where I do not need nearly as many engineers, data analysts, or data engineers for the foreseeable future. That is shocking to me. I know I still need people, but not in the ways I used to.”

Rather than retreat into a generic debate about whether a discipline is “dead,” go back to the abundant future. What valuable work was out of reach? Which customers are still underserved? Which experiments never got run? Where can a talented person’s judgment travel next?

If the work is there, retraining and redeploying people toward it gives you momentum and them a humane path. If it is not there, leaders owe people the truth. You cannot sell recomposition as empowerment when the real goal is subtraction. Tell people what is changing, what you do not know yet, and what choices they have.

Step 10: Change what the organization rewards

Once you have evidence of how the work and the teams should change, change the rules in public.

People have heard “we are doing AI” a hundred times. Talk is cheap. The truest expression of what an organization values is whom it hires, rewards, and promotes.

Change the hiring bar. Look at the people behind your strongest experiments. What made them effective? Maybe they thrived under uncertainty, took initiative, spotted patterns across systems, went horizontal without losing the depth of their craft. Turn those observed behaviors into the new bar. Do not invent a list of fashionable AI traits off a conference slide.

It also forces your interviewers to define what good AI-native work looks like. Nothing sharpens your thinking like having to spot the behavior in a stranger.

Change the feedback and promotion bar. This one carries further, because everyone already inside is watching. Describe how an excellent team member now approaches a problem. How do they use AI without outsourcing judgment? What do they do when handed a tool instead of headcount? How do they make their colleagues better? What outcomes can they now own?

Performance and promotion still follow outcomes. “Used AI” is not an achievement. Make the new expectations clear enough that people can choose how to grow into them.

What to do

If you manage a team

• Name what this team can now attempt that was out of reach before.
• Audit one important workflow from request to result. Measure working time against waiting time. Redesign the biggest constraint.
• Rewrite one visible rubric, hiring or promotion, around behaviors that already produced strong outcomes.
• Tell people plainly how roles may change and what real choices they have.

If you are an individual contributor

• Say out loud where the work waits: the context explained for the fifth time, the decision that always escalates.
• Propose a different design, not a faster version of the current one.
• Spend the capacity AI gives back on a bigger or adjacent problem you actually want.
• Make your new range visible through outcomes. Help define what excellent looks like in your craft now.

The Master doesn’t talk, he acts.

When his work is done,

the people say,

“Amazing: we did it, all by ourselves!”

Lao Tzu, Tao Te Ching, ch. 17, translated by Stephen Mitchell

Enjoyed this? Subscribe to The Looking Glass, and send it to whoever in your company has to make this real.

Bonus for paid subscribers: the principles underneath all of it

The ten steps tell you what to do. Underneath them are a handful of principles that decide whether the whole thing feels human, or feels like something done to people.

This guide originally ended with them. I cut them, because arriving after ten steps they read as a fifth act nobody asked for. They are also the ones I would tattoo on the wall, so they live down here instead: honor human agency, be honest, apply flexible rigor, and believe in the endless ingenuity of the human spirit.

Read more

I write a letter series for my daughter, who is now 11. This is one I gave her a few months ago.

Dear L,

Subscribe to the Looking Glass for regular essays on AI and the art of building.

When I was in college I read a wonderful book called “The History of Love” and it made me want to become a novelist. (As yet an unfulfilled dream, for reasons you will shortly see!)

In the book, there is a quote from Wittgenstein that I’ve never forgotten: “When the eye sees something beautiful, the hand wants to draw it.” The idea of the Eye and the Hand has been with me ever since.

A few years later, after I got my first jobs in engineering and then design, I came across another gem of a quote, this time by Ira Glass:

“Nobody tells this to people who are beginners, I wish someone told me. All of us who do creative work, we get into it because we have good taste. But there is this gap. For the first couple years you make stuff, it’s just not that good. It’s trying to be good, it has potential, but it’s not. But your taste, the thing that got you into the game, is still killer. And your taste is why your work disappoints you.”

What I love about these two quotes is that they draw a clear distinction between the Eye and the Hand, and for makers, this distinction is crucially important.

Your eye can be thought of as determining your “taste” (which is sort of a confusing analogy, because shouldn’t taste have to do with your mouth? But stay with me here). What “taste” means, in the creative realm, is the ability to distinguish the truly excellent from the merely good.

For example, you may say of a friend “they have great taste in music” if every time they select a track, you find yourself bopping along and yelling, “What band is this?!” You may say of an architect, “I love their taste,” if every time you walk into one of their works it takes your breath away (remember that time we explored Gaudí in Spain?). You may pause in wonder at a piece of fashion or food or even software if something about it strikes you as having a special quality that sets it above all the numerous other examples you have encountered.

With things that feel “tasteful,” you get the sense that the creator poured a lot of care into the creation, that they possessed both a keen eye and a deep obsession for quality.

We must remember that taste is not simply uniqueness, it is uniqueness with mastery. A master is someone who has studied their craft for years, and whose natural curiosity has led them to deep explorations of what precisely makes this thing great.

For example, I might think I’m making a unique piece of art by tossing spaghetti and pastel paint splatter and confetti onto a corkboard. But I doubt anyone would call my work tasteful. Why? Well for one thing I haven’t really studied art in depth; I haven’t dug deeply into the abstract expressionism and pop art movements, haven’t pulled on the threads of Pollock or Warhol or Rauschenberg. I haven’t even tried making more than one of these creations! If I had, I’d see that my initial idea isn’t really all that unique, nor my outcomes that remarkable.

So in some ways, the more you learn about a craft, the more humbled you become about how difficult it really is to make something stand out. And yet, exploring the breadth of what has come before is what often trains the eye to see excellence and to understand why it is excellent.

A strong eye doesn’t just mean you can spot excellence, it also means you can understand the shape of excellence. A master musician doesn’t just respond “it sounds better” when you ask them “why do you like this song?” They’ll dissect for you what makes it so special. They’ll introduce you to a whole range of similar songs because they’ve gone deep in understanding musical lineage and the inspirations behind their favorites. They genuinely love this stuff, that’s why they’ve gone so deep.

Now let’s move onto the Hand. While the eye can appreciate, the hand is what makes. And most people’s hands cannot make things as well as their eyes can see them.

When I was learning to draw as a kid, I spent many years feeling frustrated that my work wasn’t as beautiful as the pieces that inspired me. I’d try to copy some of my favorites only to be disappointed at my attempts — my lines were shaky, my body proportions out of whack, my poses static and dull.

This was the same frustration I felt starting out as a writer. I loved myself a poetic turn of phrase, but my own sentences felt clunky and repetitive. I was prone to clichés and spindly sentences. (Hence why I have not yet penned a Great American novel!)

Thank goodness Ira Glass anticipated all this and dispensed some sound advice:

“A lot of people never get past [the frustration], they quit. Most people I know who do interesting, creative work went through years of this. We know our work doesn’t have this special thing that we want it to have. We all go through this. And if you are just starting out or you are still in this phase, you gotta know it’s normal and the most important thing you can do is do a lot of work. Put yourself on a deadline so that every week you will finish one story. It is only by going through a volume of work that you will close that gap, and your work will be as good as your ambitions.”

I took his words to heart, and I saw the results firsthand. After enough years of practice, my hand started to catch up to my eye. I started drawing things I was proud of, which I displayed on the Internet during my high school and college years. I started writing essays that didn’t make me cringe upon re-reading. (note: my college application did not pass this bar.)

Now we get to an interesting conundrum: while the Eye can see things that the Hand can’t yet make, the Hand can never consistently create better than the Eye can see.

To develop our ability to make increasingly more excellent things, we have to train both our Hand and our Eye!

This gets to some fascinating questions today about AI. What happens when technology gets so good that you don’t have to write your own essay or draw your own art to make something amazing?

Given everything we have discussed so far, it should be clear that having a good eye is still important. Otherwise, if you can’t tell if AI is creating mediocrity or brilliance, you’ll probably continue churning out mediocrity (albeit at a faster rate than before), and in the process cede your creative expression to the AI.

When it comes to the Hand, however, this is where AI has the advantage. Already AI creates far better drawings than most of us (me included, despite my years of practice.) It also writes better than most people (although for now, I like to think my hard-earned skills put me a notch above. But maybe that’s just my ego talking… What say you, L?) Certainly AI does most things better than a beginner. Does this mean AI can replace our Hand?

I think yes and no. The ‘yes’ comes from short-term practicality. If AI draws better than me, and my goal is to produce a watchable cartoon series as quickly and as cheaply as possible, I’d be silly not to use AI. Hand-animating all that is simply going to get me a worse (and more expensive) result.

I’m a big believer that our souls need proof of work to sing. There are, in my experience, no shortcuts to excellence. One must love the hard work’s journey, not merely its fruits. So L, one question I keep thinking about is: what happens to the girl who never sews the crooked seam, because AI gives her a finished outfit immediately? What happens if she never learns to overcome the struggle? I fear that girl may never quite grow her Eye. When given two AI-created outfits, one singular, and one quite mediocre, I fear she won’t be able to tell which is which. That is the danger these tools pose, my love. Not that the machine is a bad Hand, but that in being so polished it decreases our capacity to obsess over the details, and in doing so to learn and grow.

My hope for you is that in the era of AI, the shape of your work may look different but will be no less hard. I imagine that instead of perfecting the drawing of a character by hand, perhaps you will use AI to explore many different characters and animation styles. Still, you will find it necessary to put in your reps playing with the characters, testing them in a scene, experimenting and failing and succeeding until you understand what makes them truly come alive and touch our hearts.

As you grow up in an AI world, L, I hope you will remember these fundamentals. Your Eye needs training. So does your Hand. Even as it can wield ever more powerful tools, no new technology will ever replace our human journey of being curious, exploring new edges, and caring enough to push onward towards new frontiers of excellence.

You are such a force of creativity, L. I can’t wait to see how your Eye and your Hand grow to make magic in the decades ahead!

With tasty love,

M

Subscribe to The Looking Glass for monthly essays on technology, culture and the art of building.

Dear readers,

You know those people who cannot bear to throw anything away? Every menu from a tasty meal in an exotic location, every abstract painting created by a toddler hand, every moth-eaten old t-shirt from a team event decades ago. You go to their house and their shelves are piled with knick-knacks, their garage is a U-STOR-IT in disguise, their bedrooms hide enormous cabinets overflowing with STUFF.

This is me, except for digital stuff.

Yes, I am a digital hoarder of the S-Tier variety. Across a dozen or more computer upgrades since I was a teen, even before the era of broadband and cloud, I have faithfully and diligently copied all my files from one place to another: every high school and college essay, images of hot dudes from the internet, chat logs with my friends, past websites, in the hopes that one day, when I’m old and bored, I’ll have nothing better to do than relive the nostalgic glories of my Internet past.

Well, my friends, that day has come, ushered in by the herald known as “Claude.”

With the power of AI, all those dusty folders in my digital attic’s digital attic can now be evaluated and sorted and worked upon.

And work we have. I’ve been using AI for a variety of purposes (more to share in upcoming essays), but this issue is dedicated to 3 weekend restoration projects that took me on a trip down memory lane back into the Internet of my youth.

The Looking Glass by Julie Zhuo posts essays once or twice a month on the art of growing products, careers and ourselves in the era of AI.

Share

AOL Instant Messenger and my 2004 chat logs

It started, as these things do, with a directory I’d been carrying around in Dropbox for 20+ years: Memories/Old files/AIM Logs/, which was conceived of on a random Tuesday 3:43am in the clutches of a deep spell of self-preservation and future nostalgia. (This small step, dear reader, is how I started my journey into becoming a digital hoarder, S tier status.)

I dumped all my AIM logs into 1,556 .htm files from 2002–2007, each one a logged AIM conversation with someone from high school or college. My screen name in most of them was JouLeZoo. (Like any self-respecting teen, I split myself across four identities.)

Recently, while trawling through my copious digital archives, I saw this little folder and became curious to read the chats again. But double-clicking a .htm file in 2026 gets you red Times New Roman on white, no styling, no way to navigate between dates. So I tasked my trusty AI helpers to help me build a little web app to flip through them. I figured: ten minutes, a static index, done.

Little was I to know, I was about to embark on the great art of digital restoration.

The process

v1 was a sidebar of buddies and a clicker for opening the chat. Worked! But looked like every dev tool from 2010. I decided to make it more “old school” and what’s when it got fun.

The XP/AIM aesthetic came back hard: Tahoma, yellow buddy list, blue Luna title bars with the red ✕, rounded windows. We added the iconic Windows XP “Bliss” wallpaper as inline SVG with the blue sky, clouds, rolling green hills embedded right in the CSS.

A fake XP taskbar materialized at the bottom with a Start button and a clock that ticks. Anyone over 30 should get a small, specific dopamine hit from this view.

A few rounds of feedback (the buddy list was painted yellow when it should’ve been white; dates read like 2004–11–20 instead of November 20, 2004). Then I remembered I had a separate stash of .doc files I’d manually saved because some conversations were too funny to lose. Dropbox’s “conflicted copy” bug had dutifully duplicated them nine times. We used macOS’s textutil to convert each one, deduped by md5, and slotted them in alongside the regular logs. The original red/blue speaker colors survived the trip from 2002.

Then came sharing. I wanted my friends to be able to laugh at this blast from the past. But even though this was dozens of years ago I wanted to maintain the privacy of each person’s conversation. So: one deployment, Gmail login, an access.js mapping each email → the buddies they’re allowed to see, server-side filtering of every request. From a friend’s POV the experience gets transposed: their screen name becomes the “account” and my screen names become their buddies, so it reads like AIM in 2005 from their seat.

I can also “act as” any friend to verify what they see; the auth bar turns red while I’m impersonating, which saved my dignity at least three times during testing.

Then I tried it on my phone. Complete disaster: fixed-300px grid, overflow: hidden on the body, and a missing <meta name=”viewport”> tag meaning Safari was pretending the page was 980px wide and shrinking everything. Gotta make it mobile-friendly in this day and age! After the fixes, the XP taskbar at the bottom doubles as a window-switcher where you can tap to flip between Buddy List and IM.

The Era

These AIM conversations captured a specific moment in time: the last gasp of communication that happened on the home computer. Texting existed but cost ten cents a pop, so it wasn’t really a thing. The iPhone wouldn’t launch until June 2007.

Facebook had just been invented at Harvard in February 2004 and was still gated to specific college networks. MySpace, Friendster, and LiveJournal were the social internet. Gmail was an invite-only beta (the Google IPO was August 2004.) We pirated music on Kazaa and LimeWire and Bittorrent and sent each other songs via AIM file transfer.

For most of the people in my buddy list, the buddy list itself was the social network: a literal list of friends, in a window on your computer, where you could watch each of them come home from class by the green icon turning on. The “away message” was the status update before that was a phrase. The profile was the bio. AIM hit its all-time peak around 2006 (somewhere around 36 million American users) and was about to fall off a cliff once everyone got smartphones and being reachable became an ambient flat-state condition rather than an event. These logs are from the last era when “ARE YOU THERE” was a normal question to type at 1 AM.

What it actually looked like in there

• The away message was the social network (Jan 2005): “wow, you are FAST with the away message reading”
• Disambiguation that lands in 2026 (Jan 2005): “wow, and i mean ‘WOW’ as in ‘IMPRESSIVE’, not ‘WORLD OF WARCRAFT IN ALL CAPS’”
• The eternal Googler problem (Feb 2005): “it’s just SO HARD to find reasonable results for a search like ‘google caltrain shuttle’ since EVERYTHING ON THE WEB talks about google”
• Peak intern energy: “i was planning on making a new headphone purchase for my internship, since ‘desk headphones’ are very different from ‘exercise headphones’”
• Proto-parasocial creator-fan email, before that was a word (Dec 2005): “i emailed her once just to say i really enjoyed her blog… and she emailed back a ‘thanks! i’m always surprised people actually read my blog’”
• Recruitment, 2005 style — broadcast over away message: <X’s> away message read “WANTED: <Y> to Microsoft”
• Tribal coding tastes (June 2007): “i have a linux box at google, but it’s not letting me run php5, and google tech support has no idea why (as you can imagine, php is practically a swear word at google — we don’t use it ever)”
• The moment JSON started eating XML’s lunch (July 2007): “Is there any reason we don’t use more JSON around here for AJAX?”
• The junior-dev cope (Nov 2004): “despite the fact that i’ve been doing my tasks at a snails-pace that to me seems realllllly slow, i’m actually Perfectly On Task! On Track! On Time!!”
• The specific cruelties of Frontend Hell (spring 2007): “in ie6 they calculate the width of the content area incorrectly” / “isn’t there something weird where input elements can’t be styled in safari?” / “i hate how firefox doubles the padding of all inputs”
• Peak finals-week energy (Jan 2006): “efficiency-robot-from-now-on-that-is-not-distracted-by-anything-and-does-not-think-about-anything-but-work-not-even-chocolate-ramen”
• The week poker night went digital: “so poker night has been replaced by VIRTUAL poker night, which basically consists of World Of Warcraft”
• A productive Saturday plan: “how’s about i bring over my desktop again (i.e. ‘move in’) and play WoW at your place… while you be productive and get work done!” (Reader, I married him).
• Apple-watcher snark on iPod shuffle launch day (Jan 2005): “in case your iPod mini was just TOO BIG and you thought the limited user interface of the iPod (and on/off switch is sooo passé) was too complicated…”
• The universal library crush (May 2005): “actually, there’s this short cute asian girl in the library i was hoping to see there tonight”
• 2005 China take that has aged like burnt bacon: “the Yahoo and Google of China may well end up being… Yahoo and Google”
• And of course: “1337!!! 1337Z0RX!!!”

the life of a designer / front-end engineer in 2006

If you have a Dropbox folder somewhere with the contents of your early twenties in it, I highly recommend asking your favorite LLM to help you go back in.

My Website From the Age of Pure Internet Joy

A few weeks ago I spent an afternoon getting back online my website Tyro Tinge. My best friend Marie and I made it in high school between the ages of 16 and 18. Actually, we made eight different versions, because I could think of no better way to spend evenings, weekends and holidays than repeatedly redesigning this thing. (Yes, such is the life of an Early Internet nerd.)

Tyro Tinge was ostensibly the home of our digital illustration collection, which featured fan art and original characters done in Photoshop and Painter and Prismacolors, alongside a message board, print shop, an epic fantasy Marie was writing years before she became a famous author. There was even a tutorial section we named gyro minge, because it seemed important to us at the time that every part of the site had to rhyme with -inge.

To understand restoration, we have to remember what the Early Internet was like in that era.

This was the year 2000, before Facebook (2004), YouTube (2005), Twitter, Tumblr, or the iPhone. The biggest sites were AOL and Yahoo (actual portals, where the homepage was a curated directory of human-edited links plus a weather widget plus a stock ticker). People got news from those portals, or from blogs they’d bookmarked, or from Slashdot if they were nerds; there was no algorithm pushing news to you.

People like us made personal, bespoke websites using GeoCities and Angelfire, in HTML we typed ourselves, in tables nested five layers deep if we wanted fancy layouts. We got music from Napster, then Kazaa, then Limewire, then BitTorrent, ripping MP3s and playing them in Winamp (”it really whips the llama’s ass!”), burning the special mixes onto CD-Rs for the car.

Back then, you announced your existence with a hit counter and a “Best Viewed In Internet Explorer 6” badge. You talked to your friends on AIM with away messages that were almost always cryptic song lyrics. Blogging was DIY: if you wanted one, you downloaded CGI software like Greymatter or Movable Type and figured out how to install it on your shared-hosting account the way you figured out everything back then, which is by reading someone else’s FAQ or ‘how-to’ and a whole lot of trial and error.

I still had my old files from this website, faithfully copied from computer to computer. Restoring this was a perfect task for today’s vibe coding era.

First, the images: the originals were optimized for 1024×768 monitors and 56k modems, which means on a 4K screen they looked fuzzy or tiny. So I ran 1,572 of them through Real-ESRGAN, one AI model for the digital pieces, a different one for the scanned colored-pencil drawings (since the first model kept erasing the paper grain that reminded you these were real pieces made with my oh-so-prized Prismacolor collection because that shit was quality and expensive!)

Second, the markup: twenty-plus years of web-standards drift had broken a lot of things, from fancy Internet Explorer-only CSS that left whole pages rendering as blank canvases to hard-coded subdomain URLs that no longer resolved. There were framesets whose relative paths assumed a directory structure that didn’t survive the flattening and a 2002-era CGI message board whose admin password was sitting plaintext in a passwd.txt. AI shines in the thankless task of translating between standards and APIs, so I was grateful to my AI helpers for making this part a breeze.

Third, the eras switcher: I wanted to compress time and be able to access every past version of the site, so I created a floating toolbar with a dropdown that lets me easily visit each version. You can pop from v1: Forest & Fairies (June 2000) to v8: Nothing Gold Can Stay (~2002) in a single click and watch what nearly three years of obsessive teenage redesigning actually looks like.

The embarrassing part is seeing what I sounded like back then. Some excerpts:

Welcome, all ye happy travellers, to the quirky site of Tyro Tinge~ ^_^ Once here you will find utter randomness and chaos galore, as you can already tell from the oddball titles of the site’s sections.

Tyro Tinge is best viewed in Internet Explorer v6.0+, in 1024x768 screen resolution. Other formats may result in abnormally organized tables and cluttery strangeness.

we are busy students who like to procrastinate on assignments long enough to keep ourselves in a perpetual state of stress

Please do not take it personally…we’re just trying to accommodate our utter laaaaaziness~ @_@;;;

The design choices are also deeply cringe. Striped or over-the-top backgrounds on absolutely everything. JavaScript hover-scroll arrows whose speed you couldn’t change. Line-heights cranked down to illegibility because cramped felt socool back then. Sparkles. God, so, so so many sparkles, on everything, like I was some kind of next-gen Charles Dickens being paid-by-the-sparkle. As someone who was later inducted into the Dieter Rams way of design, I can only shrug and blame it on the poor taste of youth.

At the same time, I’m proud of that girl. She was an early fangirl before Tumblr made fangirls a real subculture. She had a blog in 2002, decades before traditional media died. She built a message board so her readers could write back. She invented a language of rhyming nonsense for the joy of how the words sounded.

What I want most to remember about that era is this: all my free time was spent on creating. My parents rolled their eyes and asked why I didn’t spend my time doing more worthwhile, college-resume-polishing work. There wasn’t any achievement in building websites back then; the Internet was still a fairly niche hobby in my suburban community.

That’s why I get nostalgic looking back at Tyro Tinge. At that point in my life, there was no audience to optimize for, no metric to hit, no outcome to build towards. Just a fascination with art and fandom and an absolutely magical new technology, the pouring of hundreds and hundreds of hours into making something for sheer and simple joy of creation.

(fluff)Friends during the nascent days of social gaming

Some weekends ago I cracked open a Photoshop file from 2007, dusted off some PHP, translated a bunch of ActionScript, and brought (fluff)Friends back to life as a gift for my husband.

Mike and I built (fluff)Friends one weekend in 2007 on the brand-new Facebook developer platform. It blew up practically overnight.

To explain why a tiny app where you adopt a cartoon panda instantly created millions of fans, you have to remember what the Facebook platform was back then.

You could write a few hundred lines of PHP and have real people, lots of them, mostly young folks, trying your thing by lunchtime.

I was a designer-engineer at Facebook back then (this wasn’t yet a meme) and I worked on the platform’s initial launch. We had an internal hackathon to build some stuff on the platform. Someone built a quiz app, someone build rock paper scissors, and I built a pet adoption app full of cute creatures I used to doodle for fun.

Remember that this was pre-Candy Crush, pre-Zynga, pre-mobile-anything. The iPhone had launched five months earlier, the App Store didn’t exist yet, and developer platforms as a concept was novel. (Apple was, at the time, telling people to write web apps for their new phone)

The Facebook platform took off. Tens of millions of people who had never installed a piece of software in their lives were suddenly doing so, daily, in their browser, because it was so easy to do so.

(fluff)Friends was our entry. Built in PHP, MySQL, a folder of Flash files, we launched it into the world and it caught fire.

Here’s how it worked: you’d adopt a creature — a panda, a flying giraffe, a turtle, a tofu cube with a tiny smile, a monkey called a Munkii. You’d name it. You’d give it a saying. (“pet me.” “i miss my mom.” “i contain multitudes.”) You’d share your link, and your friends would come pet it. The more love it got, the better its mood. That was v1.

A few weeks later we added munny, a virtual currency you earned by petting friends and spent it on habitats, clothing, and decorative items for your fluff. Then, a few weeks after that came the upgrade that changed everything: the FluffEditor. It was a drag-and-drop canvas where you could place habitats, props, and accessories around your creature, built in actionscript. Suddenly the game wasn’t just about adopting a thing. It was about expressing yourself with your pet.

People got unbelievably creative. They built comic book scenes. They built shrines. They built dioramas of their actual relationships. I spent every weekend after that drawing new props in Photoshop (this was pre-Figma, pre-Sketch) and then sitting in front of the app on Monday morning watching what people made with them.

The Facebook team had favorites too, and we printed our fluffs on t-shirts. For years afterwards my friend Jared would tell me, unprompted, how much he loved his Tottle shirt.

While it was a hackathon project for me, Mike took it much further. He quit his job and ran it full time. He made plush stuffies of the creatures and sold them. He built a racing game spinoff that became its own hit. He shipped subscription-based virtual goods years before the rest of the games industry would discover it.

The fluffs eventually went to sleep as the Facebook platform evolved (though I still have a few of those plushies sitting on my shelves.) Until now, when Claude and I spent a weekend re-rendering every creature at 4x with fresh transparent backgrounds. We translated the ActionScript animation logic into modern TypeScript. We rebuilt the Facebook 2007 chrome from memory and a few screenshots — sage frame, white canvas, that old shade of blue. We swapped MySQL for SQLite-on-the-edge, cookie sessions for Google sign-in, and a college-apartment hosting bill for the modern stack.

The creatures look exactly like they did in 2007, only sharper. Petting still gives munny. Munny still buys habitats. You can tell your fluff what to say. You can share your profile and friends can come pet it.

None of this would survive a roadmap review these days. And yet.

I think a lot about the difference between things we make to scale and things we make to make. (fluff)Friends was made to make. Mike and I sat on the floor of my apartment with a Flash IDE open and giggled at the creatures we were drawing. A few weeks later, millions of people were giggling too. We didn’t do market analysis or have a roadmap. We had a weekend and a feeling.

Every new platform has its early adopters, and every early adopter remembers the fun of the raw, wild days. It’s a feeling I’m still chasing.

Subscribe to The Looking Glass for essays on the art of growing products, careers and ourselves in the era of AI.

You did it. Congrats! All your hard work and lucky stars and ancestors’ sacrifices have led you to this day: your stock is about to be liquid. And you are about to be rich.

I use the word “rich” quite loosely. It means different things to different people. I knew folks who felt that way with a few hundred thousand, and folks who only felt that way with 100 million. It always comes down to expectations. If you ended up with way more than you expected at this point in your life, hopefully you are feeling delighted.

You have probably already heard the research studies about money and happiness, but here’s a quick primer: In 1978, psychologists studied Illinois lottery winners and found that within about a year they were no happier than their neighbors living ordinary lives down the street. Worse, they took measurably less pleasure in everyday things: ham and eggs, pleasant conversations, “wow you look nice today” compliments. Four decades later, economists ran a bigger and better version in Sweden, following more than 2,500 winners for years. The Swedish winners did get a lasting lift in life satisfaction (aka they generally felt better about their financial circumstances, which makes sense). But their day-to-day happiness and mental health barely moved.

What does this tell us? Money buys relief and comfort. But it does not buy joy. That part is still your job.

I was in your shoes once, pre-IPO at the hot company of my generation. We had a hackathon the night before we rang the bell. Some of us felt like life was about to change. Others of us felt it was just another ordinary day — a nice milestone, sure, but with so many already crossed on the path to hypergrowth, what’s yet another?

In the years since, I have watched friends and colleagues mature into their new fortunes the way one does into a new pair of jeans. At first you’re excited because your ass is gonna look great, but then you realize it feels kinda stiff and awkward. After enough time, though, you sink into it. It becomes another detail in the shaping of your life.

Some of my former colleagues slipped away from tech like aquarium fish flushed into the vast blue ocean, off to find their true homes. It wasn’t always who I expected — for example, wildly talented engineers would decide they didn’t care to ship impact anymore. I saw folks become chefs, hoteliers, artists, therapists, writers, teachers, parents. This group always fascinated me, because it felt like I was watching some kind of Cinderella transformation. Who would you bloom into, if you had that gift of freedom?

Others pursued the leisured lifestyles of the upper crust: weekly jaunts to exotic places in search of Michelin stars (gotta catch them all!), a sleek new home with tasteful furniture (goodbye, old IKEA!), experts to assist them with said tasteful furniture selection or how to get into the hottest clubs, VIP upgrades to Disneyland and Coachella, five-hundred-dollar hoodies from Japan.

Still others continued on the same track, remaining in tech. Some became founders or VCs, and some continued their march at the same company or at a younger imitation, riding the highs and lows for years and years and years, still chasing the thrill of the climb.

Watching all of this, and living much of it myself, here are the biggest things I wish someone had told me then:

1. Money changes your life, but you get to control how much it changes your identity.

I saw this play out in all three groups: among the fish, the leisure class, and the lifers alike. I know people for whom the money simply let them check off life goals like paying off their parents’ home or having enough for retirement. They upgraded the car and maybe the house. And then, largely, they stopped. They didn’t change how they carried themselves day to day. They didn’t change who they hung out with. They didn’t take up the game of more. What I noticed about these people is that they could still connect with almost anyone.

On the other hand, the more that access to whatever money can buy becomes part of who you are, the more you become a character to others. You are now that person who eats at fancy restaurants all the time. The one who brings in a full zoo and carnival and running choo-choo train for their 3-year-old’s birthday party. Who shows up with rare watches and cars and vintage ports. You may not mean for it to, you may have done absolutely nothing undeserved beyond enjoying what you love, but you become cloaked by all the shiny things your money can buy. And for the people who don’t have what you have, this sparkling cloak can be so distracting that they can no longer really see you. You become harder to connect with.

The money will change your life. Whether it changes your self is still up to you.

2. Decide what game you’re playing

There is a game, and that game is called more. Once the money lands, a question will be waiting for you that many people neglect to think deeply about: am I still playing?

One founder told me, after he already had a nine-figure net worth, that past a certain point the money is just for scorekeeping. It mattered to him because he made it matter: that number was a measure of what he was able to accomplish and how well he ranked against the competition. If you like that game, go for it. If you know it’s a game, by all means, play.

The people I worry about are the ones who didn’t like the game but kept playing because everyone around them was playing it. I saw this most in the group that stayed, still marching or sprinting on because it was all they knew, running up a score because they were afraid to stop and ask: What would happen if I stopped? Who would I be if not a player who wins?

You’d still be you, and that’s quite enough. And I suspect there may be something even more majestic on the other side.

3. The life of leisure does not equate to a life of joy.

This is the one nobody believes from the outset, and it’s the lifestyle of the second group (the luxurious consumers) who taught it to me. The life of leisure is the thing we all grew up wishing for. It looks really good on Instagram, and it sounds really good in the stories you tell friends about your fabulous adventure-seeking, jet-setting life. As with all such things, be careful what you wish for because you might just get it.

I have watched people go through deep bouts of depression because they had everything they thought they wanted only to realize that it didn’t bring peace; instead, it brought more wanting, higher expectations and less tolerance.

What I think the aquarium fish group understood (the artists and therapists and teachers) was that they weren’t swimming away from work. They were swimming toward meaning. Their work created struggle, but that struggle is the flip side of satisfaction.

And yes, money buys financial security. But if you were the kind of person who was anxious about financial security before the money, you will probably find yourself anxious about something else after it. Anxiety is rarely because of the specific thing; it’s a part of who you are, until you figure out how to train your mind to break its old patterns.

The hardest lesson of a post-economic life is that there are no more excuses for why you’re not happy. You now have everything society taught you to want: security, freedom, status.

If you aren’t happy, the only place left to look is within yourself. The great gift of money is that it strips away every external explanation. What’s left is you, and that is both terrifying and maybe the start of a really cool internal journey.

What money is really good for

So if the jeans eventually just feel like jeans, and the leisure doesn’t fulfill, and the scoreboard is optional, what should you actually do with the money?

First, find a financial advisor. Once you have money it’s far easier to earn more, and there are a lot of arcane pitfalls having to do with taxes and investing strategies and who and what you can give to that is worth understanding.

As for how to best use it: the best answer I’ve found is this: create memories with other people. The happiness research agrees on this point across many different countries and cultures: people who spend money on others report being happier than people who spend it solely on themselves, and the experiences that stick with us are the ones we share.

There are grand ways to do this, of course. But there are also many small, unfussy ways to go about it. Maybe it means you throw more dinner parties, because you no longer have to think about the cost of the catering. Maybe you become the person who says, “I’ve got an extra ticket, do you want to join?” Maybe you fund the family reunion that would otherwise never get organized. Maybe you spend your extra leisure time on contributing to your community. None of this will trend on Instagram. All of it, in my experience, is the highest-yield investment available to your happiness.

The lottery studies were half right. The money won’t make you happier. But spent deliberately, it can buy the table, and fill the chairs around it.

One last story

Many years after our IPO, a group of former colleagues put together an informal poll of the old crew. One question asked: When you think about your life today, are you happier than you were before the IPO?

It was our own version of the lottery studies, except this time we knew everyone in the dataset. The results struck me. Roughly half said I’m about the same in level of happiness. Around 30% said they were happier. But 20% (one in five people) who had won the exact lottery you are about to win said no, they were more miserable than before.

It’s my hope that among your cohort, many more end up in the happy bucket.

Bonus: two specific regrets on how I spent my money (for subscribers only)

Read more

The man, the legend: David Fischer

David Fischer ran revenue at Meta for over a decade.

Not a region, not a vertical, but the whole shebang: hundreds of billions of dollars in advertising revenue, thousands of salespeople, relationships with virtually every major brand and agency on this planet.

Before that, he was at the U.S. Treasury. Today, he is a partner at 01a advisors.

I knew next to nothing about sales when I started Sundial. I assumed two things: a) the product should speak for itself and b) good sales was the cherry on top: the razzle-dazzle and the polish.

Of course, I was horribly mistaken (see here: I am an idiot.) Through many a conversation with David and his team (hi Liat and Katie), and through being founder-shoved into daily conversations with customers and prospects, I’ve come to appreciate the great skill needed for the subtle art of sales.

You see, the consistent line that David draws between what good sellers do and what excellent sellers do has nothing to do with charisma and polish.

Below are seven takeaways from our conversation that explain why some salespeople and sales organizations produce 10x the results from the same opportunities.

1. Good salespeople pitch the product. Excellent salespeople can do the customer’s job with the product.

Good salespeople know their product well. They can walk through features, answer questions, demo the workflow.

Excellent salespeople have a higher standard.

David told me that at Meta, every single person in his sales organization, himself included, had to go through rigorous product training and testing every six months.

But even acing the test wasn’t the bar; the true North Star is this: can you sit on the client’s side of the table and use the product the way they do?

Not in explain to me what ad manager does but rather, can you actually run the campaign the way the media buyer would? Would you be able to hit their goals? Would you be able to anticipate where things might break down?

The reason this matters comes down to the most foundational building block of sales: trust. The moment you fumble not knowing your product, your entire relationship cracks.

I loved the simplicity of the litmus test:

If you can’t do the thing the client wants to do yourself, using your own product, you don’t know it well enough.

2. Good salespeople handle the account. Excellent salespeople join the customer’s team.

Good salespeople are responsive. When a client has a problem, they take the call. They route the ticket. They follow up.

Excellent salespeople go beyond responsive: they treat the client’s problem as their own problem.

David gave me an example from Meta. Sometimes a client would accidentally enter the wrong budget and run up spend way beyond what they intended. Technically, that was the client’s mistake.

But the best sellers on his team didn’t respond with “Please be more careful next time.” They jumped in, treated it as a shared problem, and worked through different options together. When there was a product bug, same thing. When the client’s business was struggling and they needed to cut spend, same thing.

This mental model of co-ownership rather than vendor management is where long-term trust actually gets forged. It’s not in the renewal meetings or high-five good times; it’s in the messy, stressful moments when something goes wrong and the easiest thing to do would be to point at whose fault it is.

The best salespeople stop acting like vendors long before the customer stops thinking of them as one.

3. Good salespeople nurture every lead. Excellent salespeople fish for “no” early.

This one is the most counterintuitive for a start-up founder like me.

Good salespeople are optimistic. They keep conversations alive, follow up diligently, and give prospects every chance to say yes.

But excellent salespeople, especially those as startups, recognize that their scarcest resource isn’t pipeline. It’s time spent in false hope.

David described a pattern he’s seen repeatedly in the startup world. You meet a potential customer. They’re enthusiastic. They love the product. They say all the right things. You spend weeks setting up a pilot… and then they come back and say, “Actually, priorities have changed.”

His advice is to create a little scarcity. Push gently for a decision. Don’t be afraid to say, “Look, we’re working with a small number of partners right now. We need to know if you’re really enthusiastic, or if we should revisit later.”

You can do this in a warm and respectful way. But you have to be willing to give them the chance to say no.

The best prospects, David noted, will actually appreciate this. They’ll say, “This sounds great, but it’s not for us right now. Come back when you have more traction.”

It’s the politely enthusiastic ones who won’t tell you that who will waste months of your life.

4. Good salespeople listen to the buyer. Excellent salespeople read the organization behind the buyer.

Good salespeople are attentive in meetings. They pick up on signals, ask good follow-up questions, and build rapport with their point of contact.

Excellent salespeople do all of that, but they’re also doing something bigger: they’re trying to figure out what the organization actually needs, which may be quite different from what the buyer is telling them.

David described this as a kind of detective work. Your contact at a company might be genuinely enthusiastic about your product. But their boss might be risk-averse. Their procurement team might have a freeze on new vendors. The person you’re talking to might want to be the kind of buyer who takes risks on new technology, but their organization simply isn’t structured to let them.

This matters enormously for startups, because the false positive is so expensive. You can spend all your time setting up a pilot with someone who looks like a perfect customer with the right DNA and words, only to discover weeks later that the organization was never going to move.

The sophisticated seller is not just hearing the words in the meeting. They’re interpreting incentives, internal politics, org structure, and actual decision-making power. They’re looking through what people say to understand what the organization can actually do.

5. Good hiring processes reference-check last. Excellent ones reference-check first.

Here’s a hiring insight that extends well beyond sales, though David felt it acutely in sales hiring.

Most companies treat references as the final stage. You’ve done five rounds of interviews. The team is excited. You’ve basically decided you want to hire the person. And then you call two or three references who the candidate hand-picked.

At that point, David argues, you’re not actually learning. You’re confirming. Sunk cost bias has taken over. Unless the reference says something catastrophic like “this person committed crimes on the side”, you’re going to rationalize whatever you hear and move forward.

David’s preference is to push reference checks much earlier in the process, particularly for sales hires, where the stakes of a mis-hire are immediate and quantifiable.

There’s also an art to the reference check. David lets the reference talk first just to see what naturally rises to the top. Then he pushes hard on weaknesses, and keeps pushing if they give him a non-answer. I love the sharpness of his favorite calibration question: “If you were starting a company tomorrow and making your first sales hire, would this person be it?”

That question forces the issue. If the reference says “they were top 30–50%,” that’s not a strong signal. If they say “this person would be one of my first two hires regardless of what I was building,” that’s as laudatory as it gets.

The gap between those two answers tells you far more than five hours of interviews.

6. Good sales orgs standardize the playbook. Excellent ones run an hourglass.

This concept really resonatedwith me as someone who’s also managed large teams.

The classic tension in any sales organization is: do you mandate a pitch from the top, or do you let every rep run their own playbook? Most leaders pick one side or the other, and both have obvious failure modes.

David described a different model that he called the hourglass.

You start by narrowing. The company creates a standardized pitch, collateral, and set of best practices based on what the top sellers are doing. This is the canonical way to tell the story. Everyone learns it.

Then you widen. Inevitably, individual reps discover that their clients are a little different. They come up with a better analogy, a more relevant example, a different way to frame the value. You want that experimentation to happen, but off the common playbook rather than from scratch.

Then you narrow again. The best innovations get pulled back in. The canonical playbook gets updated. The cycle repeats.

What I love about this method is that it’s a a rhythm, much like designing experiences or iterating on a product. Converge, diverge, reconverge, diverge… There’s recognition for innovation (“X came up with a better way to tell our story, and now it’s part of how we all operate”), which motivates further exploration.

The best sales organizations aren’t forcing a binary decision between discipline and creativity ,but rather enabling a structure between then that makes both stronger.

7. Good sales leaders celebrate the blowout quarter. Excellent ones ask what the volatility means.

David described a seller he’d managed who was extremely high-beta. Some quarters, he’d absolutely blow away his number and hit the ceiling of whatever the compensation plan allowed. Other quarters, he’d crater to the floor.

On the surface, this looks like a star performer with some rough patches. And the good quarters are exciting and made team meetings fun.

But then David said something surprising. Over time, he began to question what those swings actually signified. Was this seller creating consistent value for clients? Or was he convincing clients to spend big in ways that didn’t sustain, leading to the inevitable pullback the following quarter?

The volatility might not just be a performance signal about the rep, it may also point whether the sales motion was creating genuine value.

An okay sales leader looks at a blowout quarter and celebrates. An excellent sales leader looks at it and asks: did we truly earn this?

This kind of self-interrogation is uncomfortable because it means questioning your own revenue during the exact moments when things are going great. But David’s point is that sustainable excellence in sales is ultimately about creating real value, not about extracting the greatest short-term spend. The two might look similar in any given quarter, but time reveals the split.

What separates good sales from excellent sales?

Across all seven contrasts, one pattern kept emerging: excellent salespeople are relentlessly honest about the actual situation with themselves, with their team, and with their clients.

They don’t pretend to know the product better than they do; they actually learn it.

They don’t treat a client’s problem as someone else’s problem; they consider themselves part of the team.

They don’t nurture false hope; they confront truth early.

They don’t just listen to the buyer; they are gathering clues to read the organization.

They don’t hire based on how polished the interviewer seems; they look for real signal in references.

They don’t rigidly standardize or improvise; they build a rhythm that improves the whole team’s performance.

And they don’t celebrate revenue without asking whether it was earned.

What I’m learning is that if sales feels tricky, it’s rarely because of lack of effort or charisma. What’s far more common is that some flavor of harsh truth is being comfortably glossed over, whether about the product, the client, the pipeline, or the team.

Excellent sellers are the ones who are dead set on excavating that truth.