Bliss Chapman
About
Neuralink team member working on decoding, UX, and signal processing
Cast within
No topic-region cast yet — this appears once Bliss Chapman's compiled claims are aligned into a topic region's argument tree.
Claims by Bliss Chapman (12)
Open loop calibration is easier to debug than closed loop
Open loop calibration (where the user has no cursor control and follows a moving target by intention) is preferable to closed loop calibration partly because it is far easier to debug—there is no feedback cycle with the user in the loop—and a key heuristic for choosing which problem to tackle is selecting the one easiest to debug since a tighter iteration loop predicts success.
Poetry compresses meaning via listener's generator function
Some thoughts are most precisely described in poetry because language has a steep information bottleneck; being literal is suboptimal compression, whereas poetry seeds the generator function in the listener's brain so they reconstruct the true underlying meaning themselves, and it is in that act of reconstruction that the listener understands the meaning—similar to how a painting's beauty is in the thought process it evokes rather than its pixels.
Magnetic targets and predictive sizing improve control
The output displayed on screen is a function not only of the decoder but of what is on the screen; Neuralink uses 'magnetic targets' that index the screen and dynamically enlarge small hard-to-hit targets like a tiny close-tab X based on cursor dynamics, analogous to how an iOS keyboard enlarges likely-next keys via a language model, improving usability without improving the underlying decoder.
Scaling channel count improves reliability not just function
Scaling electrode channel count yields a logarithmic improvement in decode quality, increases the number of distinct actions a user can perform (since each channel maps to imagined movements), and crucially should improve system reliability against neural non-stationarity without any decoder changes, because as channel count grows the relative importance of any single channel's drift or noise diminishes.
Decoder offline metrics don't equal online performance
Building a BCI decoder is fundamentally a control problem, not a standard supervised learning problem, because offline validation metrics like validation loss do not necessarily correspond to online control performance—multiple models can achieve the same validation loss yet differ greatly in controllability—so auxiliary loss terms and nuanced label construction matter more than raw loss minimization.
Neural baseline rate drift causes cursor bias
The baseline firing rate of neurons shifts day to day (like weighing rice in a different pot each day), and since the brain represents motor intention via rate coding—the delta between baseline and modulated firing—this baseline drift is the first-order cause of cursor bias and the downstream non-stationarity problem that requires recalibration.
BCI latency already beats biological movement
Neuralink's end-to-end latency from brain spike to cursor movement is about 22 milliseconds, which is already competitive with and slightly faster than the roughly 75 milliseconds it takes for a neuron's command to actually move a human hand, giving BCI users an inherent reaction-time advantage in tasks like reaction-based gaming because they bypass the muscle pathway.
Decoder error cost differs for velocity versus click
In BCI decoding, how machine learning errors translate to user experience depends on the action: velocity output can be merely on-average correct because it is integrated over time with the user in the control loop steering toward the target, but a click is performed nearly instantaneously so its error has high cost, requiring much higher precision—meaning decoder design must assign cost to errors per action type.
UX is fundamentally how a thing works
User experience is not merely the visual interface or aesthetics but fundamentally how the thing works—spanning from feature detection in the brain, to decoder design, to what control surfaces are provided—and the ideal BCI UX fades into the background so the user subconsciously feels direct control, like the difference in response curves between a trackpad and a mouse that someone deliberately designed.
BCI independence is the core value for paralysis
The transformative value of a BCI for people with paralysis is not raw capability but ergonomic independence—unlike eye trackers or mouth sticks that require specific positioning, a caretaker to set up, or that fail with muscle spasms, a brain implant lets the user interact with the world at any hour without going through another person, which is profoundly game-changing such as being able to privately text a friend at 2am.
Spike band power restored degraded performance
After many of Noland's threads retracted from the brain causing performance to degrade, switching the implant's firmware from individual spike detection to measuring spike band power—averages of populations of neurons nearby each electrode—immediately restored and ultimately exceeded prior performance, demonstrating reliance on broader population activity is more robust than tracking individual neurons.
Bits per second metric for BCI performance
BCI cursor performance is benchmarked in bits per second (BPS), computed as the log of the number of targets on screen (optionally minus one for a delete key) times correct selections minus incorrect, divided by a time window; larger grids with more targets convey more information per click, and Noland reached 8.5 BPS versus a prior human world record of 4.6.
My Notes
Loading notes...