Keeping up
What we've been keeping up with.
A running collection of the essays, talks, and threads that shape how we build — and what we took from each.
The AI Maturity Curve — Squint Cognition
Most AI failures happen when a system is deployed past its maturity level. We start every build at the automation rung and only hand over autonomy once the tool has earned it.
Why You Can’t Copy Palantir — Ethan Ding
Palantir’s real product is the shape of the company: engineers taught operations, deployments that feed the platform, and a willingness to lose money in year one to win the account. We can’t copy the scale, but we can copy the shape.
Deming’s List — the FDE market map
A map of every forward-deployed AI shop, from the ex-Palantir independents to the labs’ own deployment arms. Read the pitches side by side and they blur together. They’re all pointed at the enterprise. We’re pointed at the businesses that underpin the economy.
81% of America is Underfunded — Vijen Patel & Grady Buchanan
Nineteen percent of US GDP takes more than half of all venture capital. The other 81% is the middle market we serve: real businesses, real margins, and almost nobody building for them.
The 10,000 secret rules that run businesses — Markie Wagner, Poetic
Every business runs on thousands of unwritten rules that live in employees’ heads. Most enterprise AI fails because nobody extracted them first. Discovery is the product.
There are no lossless transformations of natural-language text — Sophie Alpert
Writing is thinking. Let AI brainstorm and proofread, but stand behind every sentence you send a client, because a summary you didn’t write carries ideas you didn’t think.
The Inner Ring — C.S. Lewis
The pull to get inside the circle never pays off, and it’s how decent people end up making small compromises. Do sound work with people you like and the real circle forms on its own.
Why Agentic Systems Need Ontologies — Frank Coyle
Probabilistic reasoning inside the model, hard logic outside it. An order can only be refunded once, and that rule belongs in a validator, not a prompt.
Harness Engineering — Ryan Lopopolo
Hold the model constant and shape the environment: context, tools, and executable checks that carry a company’s judgment. Every correction should make the harness smarter, so the second build is better than the first.
The architect’s guide to harness engineering — Alex Ker
Buy, customize, or build. Pick one harness, learn its internals, and get fluent, because the model and the harness now get optimized together.
If you stop, you’re dead — Thorsten Ball
Frontier models are a commodity, and every lab has to reinvest everything just to stay in the race. Build for where models and compute will be next year, not where they are today.
The next Horizon in agents — Bret Taylor & Clay Bavor
Agents that pursue a business outcome over months, not a single conversation, and get paid on the outcome. Outcome pricing is becoming the standard at the top of the market. We’ve priced that way from day one.
How do I build something people want? — Justin Jackson
You can’t know what a market wants until you live in its communities. Their forums, their trade shows, their complaints. That’s where our next build comes from.
Q2 2026 Business Update — Palantir
US commercial revenue up 149%, with case studies measured in EBITDA margin and dollars saved. Proof that the market pays for outcomes, not seats.
Long Humans — Josh Kushner
The opposite of the AI-doom trade: as routine work gets cheap and abundant, earned trust and judgment become the scarce assets. It's the same bet our HVA framework makes on every engagement.
$1 of EBITDA — Jack Soslow
Most AI companies can't point to a single dollar of real value they've created. We agree — it's why we price on outcomes, and why the only proof we count shows up in the P&L.
Reflections on Palantir — Nabeel Qureshi
Solve one client's exact problem. We live by "to hell with generalizability."
Understanding the code your agents write — Geoffrey Litt
Directing AI doesn't excuse losing fluency in what it builds — that caps our next idea. We are active participants in our own builds.
"It's Hard to Eval" Is a Product Smell — Hamel Husain
If you can't tell whether the AI got it right, the product is designed wrong. We build verification into the tool itself.
Making AI Useful — Tyler Kim & Caelin Sutch
Bolting a chatbot onto a site changes nothing. AI should be the interaction layer itself — answering with cards, timelines, and actions instead of paragraphs. (We hate reading the same stuff too.)
98% isn't very much — Hugo
A standing reminder that the edge cases are where trustworthy software is actually earned.
AI will improve our economy. But will we let it? — Tyler Cowen
The economics of AI are ready; whether the gains land depends on institutions and habits. Adoption is the bottleneck, not capability.
Why I Stopped Arguing With People — Cong Wang
Stop spending energy winning debates and spend it getting better. An internal memo to help serve our clients best.
Reading the same things? Let's compare notes.
A 30-minute call. We'll map where your hours go and what's worth automating first.
Learn more