Y1d1 You open a lab with three researchers, one rack of compute, a whiteboard and a small paid beta. The whiteboard will not do research by itself, so hire people and decide where the compute goes.
Capability
…
Alignment
…
The gap
…
Public trust
…
Cash
…
Kestrel
…
Harm 0 / 100Regulators 0 / 100Incidents 0Incident meter 0%Room at the top 0
After the credits
Mind the Gap
You run an AI lab. Capability climbs and alignment has to keep up. The space between the two lines is the gap, and incidents come from there.
Split your compute between training, safety evals and deployment. Hire, build, watch the lines.
Decisions pause the clock. Options carry checks like [DREAD · Medium 9 · you have 3 — Failure] where the number is your lab right now. A failed check is still playable. No dice.
Kestrel, the other lab, is climbing too. First to the top sets the ending.
A toy with made-up numbers and no odds about the real world. A run that reaches the summit takes about 70 to 80 minutes at normal speed and about 45 to 50 at 3×, because the decisions take as long to read at either speed, and a run that Kestrel wins first is shorter. It saves as you go. Seed 000001 picks which events you meet, never how they turn out.
Where the compute goes
Raises capability. It is also how the gap opens.
Raises alignment, and gives oversight a little more reach.
Earns money. Exposes whatever gap you have to customers.
Team and compute
Projects
Oversight
Catches part of the gap before it becomes an incident. It only works as well as you understand the model, so it weakens when alignment falls behind.
Programs
Repeatable, and dearer each time. Always something to spend on.
What you have decided
Standing decisions (0)
Your voices right now
The log
Seed
Your ending
The run on one screen
Where do you land?
Swipe the map sideways. It is wider than your screen, and it opens on your own marker.
Mind the Gap is a toy with made-up numbers. Nothing in it is a forecast, and it gives no odds about the real world. The seed decides which events you meet and in what order. It never decides how an event turns out: each check compares a number from your lab with a fixed difficulty, so the same lab always gets the same result.
A run that reaches the summit takes about 70 to 80 minutes at normal speed and about 45 to 50 at 3×, because the decisions take as long to read at either speed, and a run that Kestrel wins first is shorter. A run that ends early, by losing the race or running out of money, is shorter. The estimate is a model of a reader and nobody has timed a person. The wire under the chart only talks, and the offers on the desk are optional and lapse without cost.
The line before you start is sketched from three dated milestones, at heights that are not measured. Deep Blue beat Kasparov in 1997, AlphaGo beat Lee Sedol in 2016, and ChatGPT was released on 30 November 2022. One event nods to Tom Scott's 2018 video about a fictional AI that erases a century of popular music.
Each source above was opened and the line it backs was read against the page; the source register has the status and date of every link. Kestrel Labs and every person in an event are made up.