Skip to main content
Akella inMotion Lab

Lab.

Internal products and open-source tools we build between client work. For client case studies, get in touch.

Interactive model

Parity

When does a vibe coder match a software engineer?

A Monte Carlo model of the answer, with every assumption exposed as a control.

recomputing
starting worker
The brief

Two people, same raw intelligence. One spent ten years learning to build software, the other did not. Both use the best AI available. This works out when the second one catches the first, and what would have to be true for that to never happen.

The interesting result is not a date, it is a piece of algebra. Work out the largest project each person can build and safely ship, take the ratio between the two, and the AI capability term cancels out. Better models raise both ceilings by the same multiplier and close nothing.

Nothing here is a single date. Every answer is a range with a probability attached, because every number going in is uncertain too. Move anything you disagree with.

The whole model, in one line
X_max = delta * H * [ (1 - r) / (r - c) ]1/k
X_max
the biggest thing this person can ship
delta
how much of the AI they can actually direct
H
what the AI can do on its own
r
how right it has to be before you ship
c
what gets caught before it ships
k
how fast reliability falls as size grows

Divide one person's ceiling by the other's and H disappears. It sits in both, identically. That is the finding.

Ships a paid product

Where the gap settles

Turning point

When AI can out-build a good engineer working alone. Before it, the non-engineer is closing on a stationary target. After it, both ride the same escalator.

The two ceilings over time, and the distance between them. Move anything in the rail and watch which chart reacts.

Scenario
Nothing pinned, everything sampledrecomputing

The evidence as it stands, with all the uncertainty left in.

running the simulation

The same seed gives the same answer twice. Capability is anchored on METR's task horizon series, forecast uncertainty is calibrated against Farmer and Lafond's measured forecast error across 53 technologies, and verification behaviour is anchored on published defect and code review studies. This is a model. It is wrong in ways the Limits tab tries to be honest about.