Co-Founder & CEO, https://xitter.cf/t.co/v6BlPWIyst | prev: GM @Amazon | Angel Investor in ambitious founders: @twilio @cred_club @turingcom and 40+ more at https://xitter.cf/t.co/2w8zBy2IG1
San Francisco, CA
Joined July 2008
- Tweets4.4K
- Following1.6K
- Followers4.1K
- Likes8.1K
Pinned Tweet
1/9 A 🧵with simple lessons on writing well from my dozen years of writing at Amazon. cc @shreyas @lennysan @sachinrekhi @hnshah @gokulr @gibsonbiddle @kunalb11
Looking forward to catching up with so many product/builder friends in SF @lennysan
There's a @studiogang tower across the street from my office. MIRA, 400 feet of bay windows that rotate a few degrees on every floor until the whole building looks like it's mid-turn. Just about every residence ends up a corner unit.
I stood in front of Aqua and the St. Regis in Chicago and had that odd mix of familiar and peculiar before I understood why. Same hand.
Apple has this. Pick something up and you can usually tell who made it before you find the logo. Very few software products carry a signature like that. Most of ours get assembled from the same component libraries and the same patterns, and you could swap the wordmark without anyone noticing.
What I keep thinking about is where hers comes from. At Aqua, the balconies started as a question about what a resident on the 43rd floor could see past the towers in the way, and they ended up shading the apartments below and breaking up the wind so well the building needed no tuned mass damper. At MIRA, the twist starts from the bay window, the most San Francisco element there is.
Solve the real problem the same honest way enough times and the signature shows up on its own.
What does it take to build a world class product?
Whoa! love living in the future with Matic
Introducing Cues: Voice & Gesture control for Matic
We raised $115M and spent 9 years to make the world’s first intuitive home robot
Say you spilled coffee: Point to the spill and say, "Hey Matic, clean this" and it will hear, see, locate in 3D, and go clean on its own.
Matic comes with a lot of features:
1. Say "Hey Matic" - it locates your voice, turns, and looks at you
2. Say "Hey Matic, follow me" and start walking. Matic will follow behind
3. Say "Hey Matic, go clean the living room". Since it knows your house map, it navigates and just does it
It's so easy, a 5 year old and an 80 year old can use it and it understands 75 different languages.
Matic has 8x the airflow, specialised cleaning algorithms for rugs, corners, toekicks, mopping, etc and cleans better than any other robot vacuum.
Also keeps improving with software updates.
13,000 families use and love Matic.
WIRED magazine gave it a 10/10 (the only hardware to receive this rating in a decade)
Buy yours at maticrobots.com and if you don't love it after 6 months, we'll give you a full refund.
To celebrate our launch, we're cleaning 300 homes with Matic in San Francisco and New York City.
Comment "Matic" below, we'll send you the link to sign up and come to your doorstep to clean your home with Matic.
Grateful to be in the @AforeVC family! Congrats on 10 years! lmk if you want a referral
For 10 years, @AforeVC has invested in People Before Ideas.
Today our new website puts that front & center 🚀 afore.vc/residence
Here's what #PeopleBeforeIdeas means to us 👇
Continuous improvement is comfortable and quietly dangerous. You drive a number down five percent a year, everyone claps, and you never notice you're optimizing a process that shouldn't exist. At Amazon I learned to ask a sharper question. What would this number be in a perfect world? If procuring servers takes a month and we're proud of shaving off days, the real question is: what if it were zero? Now you're not improving the process, you're asking whether it should exist. That gap, between where you are and the theoretical floor, is where the big ideas hide.
AI is what makes this exercise dangerous to skip. For a lot of metrics that used to have a practical floor, AI just moved the floor to near zero. Time to first draft. Cost to answer a support ticket. Hours to analyze a dataset. If you benchmark against last quarter you'll improve five percent. If you benchmark against "what if this were instant and free," you'll often find AI can actually get you most of the way there now, and the five-percent plan was aiming at the wrong target entirely.
At Olto we put the impossible-seeming number on the metrics that matter and ask what would have to be true to hit it. With AI, the honest answer is increasingly "less than you'd think."
Took my son to a potter’s studio.
The wheel doesn’t care how badly you want it. Three collapsed before one held. Clay on his shirt, his face, somehow in his hair.
There’s no undo on a wheel. You keep your hands in cold wet clay until it centers, or it doesn’t.
Most of what he’ll grow up doing happens on a screen, and screens forgive you in ways that teach very little. Clay corrects you honestly. That’s the whole reason we started Makerhead.
We drove home grateful. That he has hands and a free Saturday, and that a working potter would put a child on a real wheel instead of handing him a kit.
The pot is not good. He’s proud of it, and so am I.
A pattern I use almost daily: don't decide until you actually have to. There's a term from lean construction, the last responsible moment, the point where waiting longer starts killing good options. Before it, every day you wait buys information. After it, the decision gets made for you by default, and defaults are usually bad. Most teams do the opposite, rushing to close an open question months before they need to, while uncertainty is highest, then defending a call they made blind.
This is a superpower for AI-native teams, if you're built for it. The pace of model improvement is so fast that a decision you're agonizing over today can be trivial in three months when the next model ships. The question isn't just "what's the answer," it's "do I have to answer this now, and can I build the thing so this decision stays changeable in real time?" Hard-coding a guess about model behavior today is how you get locked into yesterday's capabilities. Design so you can swap the guess later.
At Olto the instinct to look decisive by deciding early is a trap. It's almost always easier to change a decision you haven't made, and in AI the ground shifts fast enough that the wait usually pays.
A product and a business are not the same thing, and they don’t run on the same fuel.
Building the product is a taste problem. What to include, what to cut, how it should feel in the hand. No spreadsheet tells you that. It’s judgment, and judgment is taste.
I didn’t fully understand the rest of this until I moved from a product leader to a GM at Amazon. As a product person, taste was the whole game. As a GM, the questions changed. Will customers hand you their money, their data, their reputation, and do it again next year. None of that turned on how elegant the product was. It turned on whether we kept our word when keeping it cost us something.
That’s the trap I’d fallen into: thinking great taste earns you the business. It doesn’t. It earns you the product. Plenty of beautifully built things die because the company behind them couldn’t be relied on.
So the question I ask myself now isn’t “is this good.” Any product person can answer that. It’s “would a customer bet their own job on us,” and that one took leaving the product seat to even hear.
The most useful trait I saw at Amazon wasn't intelligence or speed. It was tolerance for looking wrong in public for a long time. When the team pushed to build a self-service publishing platform, plenty of smart people thought it was pointless. Millions of books already existed. More titles wasn't the bottleneck for a device that didn't exist yet. It became one of the biggest things the Kindle ecosystem ever produced, because the catalog was full of books priced too high or never digitized, and independent authors poured through that gap.
This matters more in AI right now, not less, because the noise is deafening. Every week there's a consensus about what AI is for, what it can't do, what's overhyped. Consensus moves so fast that betting against it feels reckless. But conventional wisdom is usually right, which is exactly why the rare correct contrarian bet pays so well, and why so few people can sit through the "this is dumb" phase long enough to find out.
The skeptics usually aren't stupid. That's the point. If everyone immediately agrees your AI idea is obvious, it's probably too small to matter.
At Olto I've started treating early discomfort as a signal rather than a warning. Easy consensus usually means we're playing it safe.
People think the hard part of innovation is the idea. It isn't. Ideas are cheap, and AI made them cheaper, you can brainstorm a hundred plausible products before lunch. The place things actually die is far less glamorous. It's the moment you commit real resources, name one accountable owner, and pull people off other work to chase something unproven. That's the leap, and it's where I've watched more good ideas quietly expire than anywhere else.
AI raises the stakes on this. When generating options is nearly free, the constraint shifts entirely to commitment. You'll have ten exciting prototypes and the temptation to keep all of them alive at 10 percent effort, which guarantees none of them get the focus to become real. Optionality feels smart and is usually just fear of choosing.
At Amazon the unlock was a single accountable owner and a dedicated team. Not a committee, not a side project, not someone's spare time. One person whose job is to make this real or prove it can't be.
At Olto our job often isn't having the idea. AI gives us plenty. It's deciding which one we actually leap on, then living with the choice while the other nine prototypes sit untouched.
Everyone runs the 5 Whys on a failure. Something breaks, you dig to the real cause, you fix it for good. Almost no one runs it on a success, which is a waste, because a win you can't explain is a win you can't repeat. A campaign drives a big jump in conversion, the easy move is to celebrate and move on. The valuable move is to interrogate it: what actually drove the lift, which customers, which message, the idea or something specific about the execution? Until you know, you got lucky.
This is acute with AI, because AI results are genuinely hard to attribute. A prompt change, a model upgrade, a tweak to retrieved context, and suddenly quality jumps. If you can't say which one caused it, you can't reproduce it, and worse, you'll credit the wrong thing and double down on it. The teams that get durable gains from AI are the ones disciplined about isolating what actually moved the metric, usually through real evals rather than vibes. An unexamined AI win quietly convinces you you're better than you are, right up until it regresses and you have no idea why.
At Olto we autopsy the wins as carefully as the losses. The goal is turning "that worked" into "that worked because of X," the only version you can run again.
I drew this on a notepad last week trying to explain what we’re building.
Most companies treat product and revenue as two separate worlds. Product people ship features. Revenue people sell them, support them, renew them. The two teams meet in a quarterly review and mostly talk past each other.
But the place where customers actually feel the company is the overlap. It’s where what you built meets what they bought. Where the demo a presales engineer ran has to match the thing the customer logs into on day 30. Where customer success can only save an account if the product earns its keep.
That overlap is where Olto ✦ lives. Not in product for its own sake, and not in revenue for its own sake, but in the messy middle where the two have to agree.
We started here because it’s the part everyone owns and no one runs.
Revenue is an output. Margin is an output. Cash flow is the one that matters most, and none of them are things you can move directly. At Amazon this took me a while to respect. You don't improve sales by staring at the sales number. You improve the inputs that feed it: selection, in-stock, discoverability, price, delivery speed, the on-site experience, service when something breaks. Fix those and the output moves on its own.
AI gives founders a dangerous new way to confuse the two. It's now easy to instrument everything and build a dashboard that tracks a hundred output metrics in real time, which feels like rigor and is mostly just a fancier scoreboard. The discipline is the same as it ever was, harder now because the temptation to watch is greater. Find the few inputs that actually drive the result.
For an AI product that's often things like the quality of your eval set, the relevance of your retrieved context, the latency the user feels, the rate at which you catch and fix bad outputs. Obsess over those. The output metrics are lagging confirmation.
At Olto we spend our time on the controllable inputs and treat outputs as a check on whether we picked the right ones. You don't win by watching the scoreboard, no matter how good the dashboard looks.
What does thinking big look like?
New project: interceptfund.com.
More: interceptfund.substack.com/p….
