Adjust the heating.
Heating powerExample
Less heatMore heat
Move right to raise pressure and earn points faster. Move left to reduce pressure when risk rises.
AN EDUCATIONAL SIMULATION GAME
You control the heating and shape of a simulated fusion plasma. Keep it stable for six simulated seconds. Higher pressure earns a higher score.
About two minutes. No physics background required.
The controller is playing here. Press “Play the game” to take your turn. This uses a simplified physics model; the 3D geometry is illustrative.
HOW TO PLAY
Press Play the game, then Start shot. Move the heating and shape sliders while the clock runs. Your goal: keep the plasma stable for six simulated seconds and earn the highest score you can.
Move right to raise pressure and earn points faster. Move left to reduce pressure when risk rises.
Move right to raise the stability limit. Shape changes take time and reduce heating efficiency, so plan ahead.
Watch the pressure, stability margin and risk forecast. If the forecast rises or the margin shrinks, pull the heating back.
Finish the round: stay stable until the clock reaches six simulated seconds. A disruption ends it early. Your score adds up pressure over time.
Start slowly: practice takes about 60 real seconds. The real-time round takes six. Then watch the controller and compare scores under the same conditions.
THE AUDIENCE
For people exploring how AI should act, how humans stay responsible, and what good control looks like. No plasma-physics background required.
Make the conversation about automation concrete. Explore which decisions people should make, which can be delegated, and what needs to be designed before deployment.
Pressure-test the phrase “human in the loop.” Consider response time, operating limits and evidence when deciding whether oversight is practical.
Learn through your hands. Fly the plasma, compare the controller, and connect the experience to the science behind prediction and feedback.
THE QUESTION BEHIND THE BUILD
What does human oversight mean when a system moves faster than we can react?
I built Twenty Milliseconds to make that question tangible. In my work in AI delivery and governance, I focus on turning ambition into clear objectives, operating limits and evidence.
This experiment invites you to experience the timing problem yourself: take the controls, feel the pressure, then compare what happens when a controller acts every 20 milliseconds.
The idea I want people to explore is simple: meaningful human oversight includes deciding the goals, setting boundaries, and reviewing outcomes. Those responsibilities begin before the system starts making decisions.
Try it. Question the assumptions. Follow the evidence.
01 THE TIMESCALES
Put the controller, the plasma, and a human response on the same clock. Drag the timeline to see what fits inside 300 milliseconds.
Elapsed time
Ten control intervals fit inside the 200 ms forecast horizon.
5 intervalsin the model’s 100 ms pressure-response timescale.
10 intervalsinside a 200 ms forecast horizon.
12 intervalsbefore the assumed 250 ms human response.
20 ms and 200 ms: PPPL’s PACMAN report. The 100 ms response, 10–50 ms growth band, and 250 ms reaction are illustrative game assumptions; timing varies with the plasma and task.
02 THE THREE-SHOT CHALLENGE
Practice slowly. Try real time. Then let the controller fly. These benchmark runs show why the pace of the loop matters.
Practice gives your hands time to respond.
The same warning becomes a fraction of a second.
Repeated adjustments hold pressure below the limit.
Calculated from this game: 24 fixed seeds (11–34) per mode. The simulated human reacts after 250 ms. Traces show the same seed, 11. These are educational simulation results, not DIII-D measurements or predictions of your score.
Take the controls03 INSIDE THE SCIENCE
Heating raises pressure. Magnetic geometry changes the operating margin. The controller balances performance against instability.
THE PERFORMANCE LEVER
In this game, increasing beam power raises normalised plasma pressure, βN. Higher pressure improves the score, but crossing the changing stability limit can trigger a tearing mode.
Experimental work at DIII-D demonstrated machine-learning approaches to plasma control. Seo et al. studied tearing avoidance; PACMAN connects diagnostics, predictors, controllers and actuators.
Four equations, a scripted stability limit, a perfect forecast, and a rule-based controller. The score is βN·seconds, a pressure-time proxy. It does not measure fusion power.