Thomas Reardon

The through-line

Reardon's consistent starting point is an asymmetry: machines pour information into us at an enormous rate, and we can barely answer back. "Machines sort of shouted us all day long in a very high bandwidth fashion but we aren't able to shout back at them in a high bandwidth fashion" [1]. Everything else follows from trying to fix that, and the fix he proposes is not a better keyboard or a better microphone but a different place to tap the nervous system, one close enough to the muscles that individual motor neurons can be heard, and far enough from the brain that no surgery is required. His work at CTRL-labs was framed around exactly this combination of neural interface design, computational neuroscience and machine learning [2][4].

The second through-line is a methodological argument he shares with his co-founders: neuroscience should aim at augmentation and at scale before it aims at pathology. The company was founded by "a set of neuroscientists who are really obsessed with trying to do something that wasn't based in where most academic Neuroscience is based today which is in neuropathologies," pointing instead at "the realm of neuro augmentation and maximizing what neural gifts most people already have" [1].

On the input/output bandwidth imbalance

The imbalance is structural, not incidental. Input is massively parallel: "millions of neurons that are directly taking in input from the world," including proprioceptive neurons nobody thinks about, with the brain effectively extending sensory neurons "throughout your entire body" [1]. Output runs through roughly 200 muscles, and muscles are slow. "In the end all you're doing is turning muscles on and off and that process is necessarily slow" [1]. Speech and typing are equally hobbled, because both are movement: "moving my lips moving you know my lungs and my pharynx Etc to communicate verbally to use my hands to communicate via keyboard or a mouse all of this is very very low bit rate" [1].

He is careful about the numbers while still using them. Estimates put the gap at "between six and seven orders of magnitude," which he glosses as "a million to 10 million times difference," while noting that "people have tried to write down a number for this I always think these things are a bit you know hard to verify" [1]. The point survives the imprecision: the bottleneck is on the human output side, and it is a movement bottleneck.

On neuro augmentation before neuropathology

Reardon frames the choice to build for the general population rather than for clinical populations as a philosophy about how science moves. He believes working on "scale problems first and then partnering with clinical practitioners who can then adapt sort of the scale discoveries to clinical populations" gets results faster, and that "a lot of Neuroscience does get slowed down by this kind of clinical first pathology first approach to understanding and decoding the brain" [1]. The mechanism he trusts is machine learning applied at volume: by "using effectively at heart machine learning at a very very high scale we would learn things about the brain that are hard to learn in this kind of bespoke single experiment at a time approach that academic Neuroscience is mostly centered around" [1]. The target is not a deficit to be repaired but a capacity already present and underused, "maximizing what neural gifts most people already have" [1].

On why the interface should be non-invasive, and why EMG

The commitment is to "non-invasive neural interfaces," meaning detecting nervous system activity "without effectively Perforating the human body," so no drilling into the skull and nothing inserted [1]. That constraint leaves a short menu, and he ranks it by resolution. EEG listens "at a distance" to large electrical waves generated by hundreds of thousands of cortical neurons; functional MRI is "even bigger," looking at millions of neurons and their metabolic signatures; EMG, by contrast, lets you "actually listen to individual motor neurons about the currency of that nervous system which is called an action potential or a spike" [1].

The deeper justification is about where control actually happens. "You don't actually use single neurons to control the world today you use muscles to control the world," and it takes hundreds of thousands or millions of forebrain neurons to coordinate the muscles that let you type [1]. So the design move is to skip the movement entirely: "let's not focus on the movement let's just actually listen to individual neurons down at that muscle interface and see if you can use that two controller machine to otherwise type without actually doing all the movement of typing" [1]. Surface electromyography is the technology used to pick up those signals [1].

On why the hand is the right place to listen

He does not dismiss voice. "I think voice is a great way to control computers especially since we use computers as a Communications tool all the time so using it to control communication streams is fantastic" [1]. But control of the physical world runs through the hands, whether that is driving a car or making a cup of tea, and the neurology reflects it: more forebrain neurons are dedicated to the hands than to any other part of the body, by estimates "around twice as many neurons" as are dedicated to the mouth for speech production and feeding [1]. The density exists because hands need adaptive motor skill rather than reflexive or repetitive motion. His way of putting it: "your hands make your legs look really really dumb," since legs "don't really do much adaptively" but hands "can do a whole lot" [1].

There is an anatomical wrinkle he likes. Most hand movement comes from the extrinsic muscles in the forearm, so "most of the quote-unquote intelligence in my hands is actually embodied in my forearms" [1]. There are intrinsic muscles in the hand itself, such as the one manipulating the thumb, but those are not the ones he listens to; the interface targets the extrinsic forearm muscles [1].

On dimensional collapse and why it is an opportunity

Reardon describes the motor system as a funnel. Perhaps 50 to 100 million cortical neurons at the top are dedicated to controlling the hand muscles; by the time signals pass from upper motor neurons in cortex, down the long synaptic wire, to the lower motor neurons in the spine, a single muscle may be driven by only four hundred to a thousand neurons, "about four orders of magnitude or more smaller" than the cortical population [1]. He takes the standard view of why: "control has consequences," and precise control of the end effector demands that collapse, with the cortical computation existing to deliver "that fine control that's only a couple hundred neurons and then ultimately down to just those 14 muscles" [1]. For an engineer, that funnel is a gift. Measuring at the muscle abstracts away the complexity of control and gives you a highly processed, high-information output modality without touching cortex [1].

On breaking motor recruitment

The scientific bet that most excites him concerns Eccles's 1950s theory of motor recruitment, taught in the first semester of any neuroscience class and emphasized in Principles of Neural Science, "the Bible of neurosciences," which he calls "a fantastic book" though "not one that I would just pick up and start reading" [1]. Recruitment says the neurons innervating a muscle always turn on in the same fixed sequence, Alice then Bob then Charlie, and de-recruit in reverse. Mathematically that means "there's only one dimension of activation," and from a measurement standpoint you can see "effectively only a single variable something that goes from you know zero to a thousand but it's in the same sequence every time" [1].

His goal was to falsify that as a hard constraint: "our goal at control Labs was what happens if that's not true," on the suspicion that "maybe people just haven't really tried hard enough to look at how you might activate the neurons differentially" [1]. He says they have "proven to ourselves again and again and again is that motor recruitment is not a guarantee" [1]. The payoff is combinatorial. If a fixed order holds, "a predicts b b predicts C" and information collapses; if Alice can fire alone, then Charlie alone, then Alice and Charlie, then Alice and Bob, "I now combinatorically massively expand just with three neurons I've gone to say eight states of activation," with a rate code layered on top [1]. Framed at the level of the whole hand, the question is whether you can "break that model" of fourteen one-dimensional muscles and instead "treat it as a multi-dimensional object," so as to "massively increase the dimensionality of an individual muscle" [1]. He concedes this "seems like a mathematically abstract concept," and defends it as exactly the trade of computational neuroscience: "we like to read these kind of like gross simplifications to try to understand the information that's being carried from the brain up to muscles" [1].

Takeaways

• The human-machine bottleneck is on the output side, with estimates of six to seven orders of magnitude between what we take in and what we can send back, because all output ultimately runs through roughly 200 muscles [1].
• Speech and typing are both movement, and movement is slow, so a higher-bandwidth interface means bypassing movement rather than optimizing it [1].
• Build for the whole population first: working on scale problems with machine learning and handing discoveries to clinicians afterwards beats the "clinical first pathology first" default of academic neuroscience [1].
• Among non-invasive options, EMG is the choice because it resolves individual motor neuron spikes, where EEG sees hundreds of thousands of neurons and fMRI sees millions [1].
• The hand deserves roughly twice the forebrain real estate of the mouth, and most of its control lives in the extrinsic muscles of the forearm, which is where the signals are read [1].
• The motor system's collapse from ~50 to 100 million cortical neurons to a few hundred per muscle is what makes a muscle-level tap a highly processed, information-rich place to listen [1].
• Eccles's fixed-order motor recruitment is not a guarantee; if neurons innervating a muscle can be activated in arbitrary combinations, the state space expands combinatorially rather than staying one-dimensional [1].