Reaching escape velocity
Shifting the unit of AI productivity from people to platforms.
One argument. Two sessions.
Session one makes the argument and rebuilds the roadmap. Session two builds the machine.
this week
This session is the map — the next one is the aircraft.
The AI Adoption
Continuum
Ground → Airborne → Breaking atmosphere
The discrepancy
The market can’t agree on whether AI works. Both sides are looking at the same machine.
Two realities coexist in the world
Your company bought the tools and ran the pilots. So where are the gains?
of businesses fail to see meaningful returns on AI.
Solo builders ship in a week what used to take a quarter.
How can both of these be true?
Both are true. The difference is how they move, not how hard they work — that's the map, next →
Same car. Different driver.
“AI hallucinates.” “It can't do X.” “We've seen it fail at other companies.”
Put an automatic-only driver in this car and they spin it into the wall...then tell you why it's undrivable. Put a pro in the same seat, and it's the fastest thing on Earth. Both are correct.
So the question isn’t whether the car is good. It’s how many laps you’ve driven.
I paid for the feature. Then I refused to press the button.
I had reasons. Good ones. I said them out loud, with confidence, for months.
Nothing about the car changed that week. I did. Trust wasn’t a decision — it was residue from reps.
The car kept updating.
I didn’t.
Every dismissal has a version number on it. Almost nobody checks the date. Most confident takes about what AI “can’t do” were formed on a release that has since been deprecated — twice.
What sounds like rigor, is often a skills gap.
The most responsible-sounding teams are often the least practiced.
before the reps are in
a hallucination, some slop
must be the tool
knew I couldn't rely on it
known beats unknown
Every stage of this loop masquerades as good judgment; however it is really a signal of low proficiency + low agency. Distrust is a status report on your scaffolding.
And the hallucination fear assumes the model is always guessing...it isn't. Hold that thought.
“Five years away” was the smart take — for fifteen years.
It was sophisticated. It was evidence-backed. It was right, over and over. People built careers on being right about it.
Your skepticism has been right many times. That’s exactly what makes it expensive now. Being right about hype nine times is what trains you to dismiss the tenth.
"Using AI to code" is three very different crafts
You chat, it generates, you eyeball the result and ship. No specs, no checks, nobody reads the code.
Agents produce bounded chunks of code. A person reads every line before it lands. Quality lives in the review.
You author specs and scaffolding. The system builds, verifies against acceptance criteria, and returns evidence. Quality lives in the spec.
High-leverage teams don't just use agents to code; they build systems that govern how agents code.
The map
A continuum of operating modes, and where you sit on it.
Think of AI adoption as modes of transport.
Six ways to move. Only some leave the ground.
The shift from player, to coach, to conductor.
We're widely still in the ground transport phase.
Driving gets you ~10–20% faster, but not the 10× solo builders see. The difference lies in the mode of transport.
The Time Traveler's Dilemma
WHY THE ENDS CAN'T TALK
Two people at opposite ends of the continuum aren't disagreeing; they're having entirely different conversations. One is describing a future the other can't see from the ground.
You won't know what you don't know.
The system
Closing the gap is organizational, not individual. A faster task is not a faster product; the team is the unit of change.
A faster task is not a faster product
Accelerate one box in isolation, and the work piles up at the next boundary.
The bottleneck moved. The system didn't.
Local speed is not system throughput.
Faster individuals create more review, more rework, more undecided work in flight. The team is the unit of change.
Make intent executable;
keep execution governed.
AI changes product development when a team turns shared intent into executable artifacts that agents can build from, and people can govern.
A prompt disappears. Artifacts persist.
These files are the contract between the roles, and the interface the agents build from.
Delegation returns evidence
The model generates; deterministic tools verify.
We never rely on the model's opinion of its own work. Done means a green, verifiable verdict.
The path
What this means for you — and where we go this week.
Getting airborne requires organizational transformation
Buys the tools. Rolls them out. Provides foundation & infra
The people who write the code. The org can't get airborne until we do.
System-wide change — but only after Helicopter is proven.
One argument. One week.
The argument: from faster tasks to a different product-development system.
Make an agentic-ready roadmap — the artifact rebuilt for an agent reader, and you build one in the lab.
Develop an idea with an LLM — name the destination, map the decisions, work the frontier.
Synthesize those decisions into an agent-ready repository of artifacts.
Getting Airborne. Zero application code to a working product — the harness builds PEDAL live from the artifacts.
Breaking Atmosphere. When software starts learning — repeated behaviors compile into deterministic rules.
We're not selling a framework; we're opening a live lab. Bring your product partners.
Don't do it for me.
Don't do it for Lyft.
Do it for yourself.
This is about your career, your growth, and your future. The sessions ahead are the on-ramp; the reps are yours.
INVEST IN YOURSELF · START THIS WEEK