Each dot is a neuron that fires most for one movement direction. A decoder adds up their spikes to guess where the user wants to move; use more neurons, longer bins or smoothing to see the guess sharpen, or silence electrodes to watch it degrade.
The science
Neurons in motor cortex are broadly cosine-tuned: neuron i with preferred direction θi fires with mean rate λi = b + m·cos(θ − θi) (rates are cut at zero). Spike counts in a bin Δt are Poisson with mean λiΔt. The population-vector decoder estimates the velocity direction as v̂ = (2 / N·m·Δt) · Σ (ni − b·Δt) · ui, with ui the unit vector of θi. For uniformly spread preferred directions it is unbiased, E[v̂] = u(θ), and its variance is Σ Var = 4·Σλi / (N²m²Δt), so the error falls as 1/√N and 1/√Δt. The page checks that formula against the actual Monte-Carlo error bin by bin.
The cursor moves with a smoothed decoded velocity; the simulated "user" aims at the target each bin (closed loop), which corrects errors the way a real user watching feedback does. Smoothing v ← αv + (1−α)v̂ trades noise for lag.
Simplifications (honest). Real recordings have correlated noise, non-cosine and non-stationary tuning, speed coding, and need recalibration (modern decoders use Kalman filters or neural networks and the brain itself adapts). All rates, depths and counts are illustrative, not data from any device or patient. This is an educational model of the decoding principle, not a medical tool.