Talk draft · 2026

One abstraction up

What AI actually changes about building software — and what it does not.

The pattern

Every layer absorbs the one below it

And after a lag, the lower skill stops being a differentiator. Hiring does not reward effort. It rewards whatever is scarce at the current layer.

The ladder

Punch cardswiring the machine

Assemblytalking to the hardware

Cmanaging memory

Java, OOPstructuring large systems

Python, JSwriting the program

React, frameworksassembling an application

Orchestrating AIshipping and operating a product

What the old tests looked like

2010s

A calculator, or a small game. Basic Python was enough to get hired.

2020s

A CRUD app. A Netflix clone with login, a database and a few screens.

Now

All of that, and more, is a prompt away. The artifact stopped being the proof.

So what is the proof?

Something live. Real users. A bug you had to fix at 2am. Consequences.

Correction people get wrong

The abstraction removes the need to demonstrate the layer below

It does not remove the need to understand it. React developers still need the DOM. AI-assisted builders still need HTTP, data models and failure modes.

Say “the test moved”, never “you are obsolete”.

The adoption gap — sourced

2.2%

of US households had a paid AI subscription, April 2026.

0.1%

was the same figure in January 2023. Fast growth, tiny base.

~30%

of S&P 500 companies report some “quantifiable impact” from AI.

~2%

of those same companies track any metric for it at all.

Adoption is broad and shallow. Even the companies claiming impact mostly cannot measure it. The gap is not model quality. It is execution.

Source: PNC Research internal data (13 July 2026), via a16z State of Markets II. Note: the 2.2% measures paying households, not usage — free usage is far larger. Say “paid penetration”, not “adoption”.

Because look at the ceiling

What people are already doing

Porting PS5 drivers to PC.
Porting and modding games, merging two unrelated titles into one.
Multi-day autonomous pipelines that research, build, test and deploy.
A psychiatrist generating music with AI as a tool for his patients.

The bottleneck, from MIT

95% of enterprise AI pilots produce no measurable impact

MIT NANDA, The GenAI Divide: State of AI in Business 2025. 150 leader interviews, 350 employees surveyed, 300 public deployments analysed.

The report’s own conclusion: the problem is not the quality of the models. It is the learning gap — for the tools and for the organizations using them.

Say the criticism before someone else does: the sample skews to large enterprises and “failure” is defined loosely. The finding that survives is integration, not capability. Lead author: startups run by 19- and 20-year-olds “pick one pain point, execute well” and go from zero to $20M in a year.

So the bottleneck is orchestration

Capability is abundant. Integration is scarce.

Models

Solved, commoditized, cheap, improving monthly.

Pilots

Bought, demoed, shelved. Stuck at the starting line.

Orchestration

Rare. Picking one pain point and running it end to end.

Which is why the reward goes to whoever can take the gains of the technology to people who are not getting them yet.

Before you pick a market

TAM is a ceiling, not an access route

Number	What it answers
Total market	How big the prize is if everything goes right
Reachable segment	Who you can actually get in front of, this year
Your channel	How the first thousand users arrive, by name

A 2B market you cannot reach is worth less than a 50M market with distribution. Name all three, or the size is just a borrowed benchmark.

My own case

I built a DAG of agents that ships a product

Read the MS certificate spec→ Construct test DNA→ Find published test books→ Scrape and OCR them→ Solve every question→ Verify the answers→ Build gradient descent→ Deploy backend→ Frontend and app UI

Not one task automated. A workflow, end to end, days long, on the free tier of a serverless platform.

And it has real numbers

150K+

learners on the flagship product.

1M+

practice results checked by AI.

4.8

store rating across Google Play and the App Store.

4

exam modules, Speaking, Writing, Reading and Listening, A1 to C2.

Shipped end to end: iOS, Android and web, in Uzbek and Russian, marketed to a market that already had an incumbent. Live at spiko.uz.

What the pipeline actually produced

Cost

Runs on free-tier serverless. Holds thousands of users before a bill exists.

Product

A live test-prep engine, competing where an incumbent already sits.

The work I did

Ideas, orchestration, verification, product decisions. Not typing.

The work that remains

Getting users. Activation, retention, conversion, and finding fit.

Who is talking

Azizbek Umidjonov

Android and full-stack engineer, based in Seoul.
Builds AI education products used by students in Uzbekistan — spiko.uz and arbee.uz, under Milliy Technology.
Shipped a live product to 150K+ learners largely by orchestrating AI, which is the reason this talk exists.
Works remotely with teams anywhere. typosbro.me

Where the moat moved

	Was	Is
Scarce	Writing the code	Operating the thing
Proof	The artifact	The users
Hard part	Building it	Getting anyone to care
Failure mode	It does not compile	It compiles and nobody comes

The durable skill

Not the current tool. The ability to be re-certified at the next layer.

Each cycle absorbs the last one. The framework you learn today is the CRUD app of 2029. What survives is your rate of re-registration.

So what do you do on Monday

Ship something small that a real person can use. Ten users beats zero users.
Keep it running. Fix one production bug. That is the story that gets you hired.
Contribute where other people depend on your work.
Get paid for something, once, at any size.
And validate before you analyse: the market is genuinely brutal right now. That part is real.

Close

The gap is not model quality. It is how many ideas get executed end to end.

Two percent pay for this. The other ninety-eight percent are still writing emails with it.

Notes to self — delete before presenting

Now sourced

2.2% of US households paid, April 2026 (PNC Research via a16z State of Markets II). The 30%/2% S&P 500 split is from the same piece. The deck quotes the number; do not round it back down to a guess.

My timeline

AI-pilled around January 2026. Before that: mostly copy-pasting from earlier models. The iOS move was recent, not an eighteen-month call.

Honest state

The engine is live and pre-revenue. Do not imply traction that does not exist. The pipeline is the story.

Order that works

Validate the market first, then the frame, then the smallest first step. Skipping validation is what makes people defensive.

Sources

Household paid penetration: PNC Research via a16z State of Markets II. Enterprise pilots: MIT NANDA, The GenAI Divide 2025, via Fortune. Market-first framework: Michia Rohrssen, “Give Me 24 Minutes…” — his framing, cited as such.