The Brain-Computer Bandwidth Curve

How much reliable intent crosses from cortex to computer, how fast that is rising, and why the constraint has moved from decoding the signal to keeping it.

Research through September 2026

Figure 1

The records bent upward. Then they went home.

Electrode counts barely moved. What changed was the thing being decoded, and then the number of hours it worked.

Independent home use · 2026

56.1 wpm · 256 intracortical · >99% prompted; 92.3% of sentences rated mostly correct. One participant used the speech and cursor system nearly daily for about nineteen months: more than 3,800 hours and more than 180,000 sentences, at an average 56.1 words a minute, without researcher support.

Figure 1: Reported communication rates, converted to words per minute at five characters a word where studies report characters. The points are not like-for-like: vocabularies, error rates, participants, cortical targets, and protocols all differ, and each point is a single study, usually a single participant. Laboratory records are peak or median rates in structured sessions; sustained points average months or years of use. Select a point for its electrode count and error rate. Sources are listed at the end of the report.

~4×Growth in electrodes in landmark human intracortical studies from 2006 to 2026: 96 to 384. About 7% a year, nothing like a semiconductor curve.

~10×Growth in the laboratory record from 2017 to 2023, from about 8 to 78 words a minute, on essentially the same classes of electrode.

3,800 hoursIndependent home use of one speech BCI over about 19 months, at an average 56 words a minute, without researcher support.

What the record shows

The useful output of the best interfaces is tens of words a minute: single-digit to a few tens of bits a second. Natural conversation is about 160 words a minute, so the gap is now a factor of three, not a factor of a thousand.
Electrode count is not bandwidth. The laboratory record rose tenfold in six years while channel count rose fourfold in twenty, because decoding moved from cursors to handwriting to phonemes, and language models filled the gaps.
The binding constraint has migrated from decoding in a session to stability across months: neural drift, calibration, heat, packaging, and surgery.
Writing to the brain is far behind reading. Current stimulation produces a few usable percepts, not a display, and its value shows up as better control rather than more bits.

This report separates raw data rate, usable neural information, decoded symbols, error-corrected output, and correct information delivered per day. They are different quantities and should not be placed on one curve.

Figure 2: The interface trade space

Every interface buys signal with surgery

Proximity to neurons, surgical burden, and chronic stability trade against one another. The families are points on one frontier, not contestants in one league table.

The highest demonstrated freedom to decode arbitrary movement and speech

Observes: Spikes, multi-unit activity, and local fields
Resolution: Tens of micrometres; sub-millisecond
Best evidence: 62 wpm in the lab; 56.1 wpm averaged over 3,800 hours at home
Stability: Arrays have recorded for up to 7.6 years; spiking on ~36% of electrodes on average

Binding constraint: Rigid silicon in soft, pulsing tissue. Scarring, micromotion, and unit turnover make the recorded signal drift even when the hardware survives.

Surface and scalp approaches lose spatial detail before software sees the data. Penetrating approaches keep the detail but inherit a biological interface that changes over months. Neither problem is solved by a better decoder alone.

Part I: First principles

A BCI is a communication channel with a biological end

Every stage between a neuron and a sentence loses or adds information. Bandwidth is what survives the whole chain, not what enters it.

Useful bandwidth≈correct information delivered to the application÷time in use + calibration + downtime

01

Sample the neurons

Intent is distributed across large populations. An electrode sees a projection of that activity: a few nearby cells, a local field, or, from the scalp, a blur of millions.

Measure: Neurons or fields sampled, cortical area covered
Failure boundary: Sampling cortex that does not carry the intended variable, or too little of the cortex that does

What the record shows at this step

Reported evidence: Extracellular spikes are 50-500 µV close to a cell and vanish into background beyond about 150 µm; skull and scalp low-pass filter everything EEG records.
Where it is moving: Wider coverage of speech and hand cortex, including the sulcal banks that current arrays cannot reach.
Principal risk: Sampling cortex that does not carry the intended variable, or too little of the cortex that does

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

02

Couple electrode to tissue

Geometry and chemistry set how much signal reaches the contact. The boundary is alive: it scars, shifts, and changes which cells a contact hears.

Measure: Impedance, usable channel yield, signal-to-noise
Failure boundary: Scar tissue, micromotion, and unit turnover that make the signal drift even when the hardware works

What the record shows at this step

Reported evidence: Across 14 BrainGate participants, arrays recorded spiking on an average 35.6% of electrodes, for up to 7.6 years.
Where it is moving: Softer materials and designs that trade some selectivity for decade-long stability.
Principal risk: Scar tissue, micromotion, and unit turnover that make the signal drift even when the hardware works

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

03

Amplify and digitise

Microvolt signals must be amplified against thermal noise, muscle artefact, and mains hum, then sampled fast enough to catch a spike.

Measure: Input-referred noise, common-mode rejection, power per channel
Failure boundary: Front-end power that scales with channel count inside a sealed, heat-limited package

What the record shows at this step

Reported evidence: Thermal noise at a typical microelectrode is several microvolts, which already sets the ceiling on extracellular signal-to-noise.
Where it is moving: Sub-microwatt analogue feature extraction on the implant itself.
Principal risk: Front-end power that scales with channel count inside a sealed, heat-limited package

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

04

Compress and transmit

Raw waveforms from a thousand channels are hundreds of megabits a second. Fully implanted systems extract features first and send a thousandth of that.

Measure: Transmitted rate, energy per bit, implant temperature rise
Failure boundary: Compression is irreversible: what the implant discards, no later decoder can recover

What the record shows at this step

Reported evidence: Threshold crossings and spike-band power carry most of the decodable information, which is why high-performance speech systems use them.
Where it is moving: Implanted wireless packages that replace the percutaneous connector without losing decoding performance.
Principal risk: Compression is irreversible: what the implant discards, no later decoder can recover

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

05

Decode the intent

A model estimates a latent variable: cursor velocity, a letter, a phoneme. Neural representations overlap and drift, so the decoder makes substitution and timing errors.

Measure: Raw phoneme or character error, latency, calibration data required
Failure boundary: Drift across days that silently degrades a fixed decoder

What the record shows at this step

Reported evidence: A fixed brain-to-voice decoder degraded noticeably within about two weeks; a fixed typing decoder reached 12-15% word error within days.
Where it is moving: Self-supervised recalibration during ordinary use, and models pretrained across many people.
Principal risk: Drift across days that silently degrades a fixed decoder

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

06

Apply a language prior

A language model turns noisy symbol estimates into probable sentences. It is the reason speech BCIs work, and the reason their accuracy cannot all be credited to the brain.

Measure: Error before and after the language model, accuracy on improbable sentences
Failure boundary: Fluent output that reflects what was likely rather than what was meant

What the record shows at this step

Reported evidence: Handwriting decoding at about 94% raw character accuracy rose to over 99% with a language model; the 2026 typing system used a 5-gram model.
Where it is moving: Evaluation on names, novel words, and adversarially unlikely sentences, not only familiar corpora.
Principal risk: Fluent output that reflects what was likely rather than what was meant

Evidence statements describe published studies and biophysics. “Where it is moving” is an editorial reading of direction, not an announced capability.

A system can need tens of megabits a second of neural waveforms to recover a few tens of bits a second of human intention. That is not waste. It is the price of inferring a quiet variable from loud biology.

Figure 3 · Where the bits go

Megabits in, tens of bits out

An implant digitises millions of samples a second to recover a message worth a few tens of bits. That ratio is not waste; it is what inferring a low-dimensional intention from redundant biology costs. What limits the implant is the heat of moving those samples.

Digitised samples

92.2 Mbit/s256 channels × 30 kHz × 12 bits

Sent across the skin

25.6 kbit/sThreshold crossings and spike-band power, ~100 bit/s per channel

Human intent delivered

22 bit/sUpper bound at 56 wpm: 5 characters a word, log₂27 bits a character

Recording channels256Amplifier and converter power20 µW per channel

Estimated implant power10mW

14% of a ~70 mW heat envelope

Front end: 5.1 mW
Radio: <0.1 mW
On-implant processing: 5 mW
Samples per intended bit: ~345,493

Compression cuts the transmitted stream by about 3,600×. It is irreversible: waveform shape is discarded, so the implant can no longer separate neighbouring neurons on one contact. At 256 channels the front end alone uses 5 mW; the heat ceiling becomes a per-channel power ceiling.

Figure 3: An illustrative model, not a measurement of any device. Radio energy is set at 1 nJ per bit and on-implant processing at a flat 5 mW. The ~70 mW envelope is the middle of a 40-80 mW range that bio-heat models give for a skull-mounted implant held to about a 1 °C tissue temperature rise; real limits depend on geometry, placement, and duty cycle. The intent bound is an upper limit: English is predictable, so the true conditional information rate is lower.

Why output outran electrodes

From 2006 to 2024, most human records used the same kind of 96-electrode silicon array. The hardware barely changed; the representation being decoded changed completely.

The decoded variable changed

Cursor control asks motor cortex to act as a joystick. Handwriting and speech ask it to do what it evolved for: produce rich sequences of distinct gestures. Letters and phonemes are easier to tell apart than cursor directions, so the same electrodes carry more usable information.

Nearby neurons are redundant

Cortex coordinates movement through low-dimensional population dynamics. Ten to thirty latent factors often capture most of the variance in a hundred-electrode recording, so each added contact in the same patch adds less new information.

Language supplies the rest

Words are not equiprobable. A language model resolves ambiguous phonemes from context, which is why 62 words a minute did not need hundreds of thousands of independent channels.

Hardware scaling is heavy

Every added channel costs amplifier power, telemetry, packaging, surgical footprint, and a new failure point, inside a skull that cannot shed much heat. Software improvements cost none of these.

A 2026 typing study made the point directly. With about 140 randomly chosen electrodes, attempted typing produced 9.5% word error while attempted handwriting produced 27%. Same participant, same electrodes, different amounts of usable information.

The software dividend has a limit. A decoder cannot recover information absent from the signal, and Shannon’s bound, C = B log₂(1 + SNR), applies to the neural link as to any other. As the signal thins, a larger model learns priors instead of intent.

Implant lifetime is not signal lifetime

7.6 yearsLongest recording from a BrainGate Utah array in a retrospective study of 20 arrays in 14 participants, with a mean enrolment of 2.8 years.

~15 daysTime before a fixed brain-to-voice decoder, trained on one day, showed noticeable degradation in a 2025 study.

10⁸×Approximate stiffness mismatch between silicon and brain tissue, which moves continuously with every heartbeat and breath.

The operational question is not whether the same neuron stays on the same wire, but whether the decoder stays matched to a slowly moving population. Adaptive decoding and biological stability are substitutes up to a point, and only up to that point: if contacts fail or signal-to-noise falls irreversibly, no recalibration can restore the missing dimensions.

The discovery chain

Progress came in discontinuous changes of representation: switch, cursor, arm, handwriting, phoneme, voice. Each one reset the rate while the sensor stayed roughly the same.

1973
Evoked potentials as a switch
Vidal showed that scalp-recorded visual responses could drive a computer selection, establishing the closed loop without muscles.
1988
The P300 speller
Farwell and Donchin spelled by attention to flashing letters. One to two words a minute set the non-invasive baseline for two decades.
2006
Intracortical cursor control
A 96-electrode Utah array let a person with tetraplegia move a cursor, open email, and operate a television, years after injury.
2012-2013
Robotic arms
Participants reached, grasped, and drank from a bottle with a robotic arm; a second group achieved seven-dimensional control.
2021
Handwriting and touch
Decoding attempted handwriting reached 90 characters a minute. Separately, stimulated touch halved object-handling time with a robotic arm.
2023
Speech
Intracortical speech reached 62 words a minute and surface ECoG 78, both with roughly a quarter of words wrong on open vocabularies.
2024-2026
From records to routine
97.5% accuracy over months, then more than 3,800 hours of independent use at home: the first long stretch of the curve measured outside a laboratory.

Writing to the brain is a different problem

Reading can average many noisy channels into one confident estimate. Writing cannot: injected current spreads through tissue, neighbouring contacts interact, and forcing hundreds of neurons to fire together is nothing like the sparse, precisely timed activity of natural sensation. Charge is also bounded by tissue safety at each contact.

20.9 → 10.2 sMedian robotic-arm task time with vision alone, then with stimulated touch from somatosensory cortex (2021). A low-rate signal transformed the control loop.

96 contactsA visual-cortex array in a blind participant evoked stable phosphenes and some letter identification, but nowhere near 96 independent pixels.

>1 millionPeople using cochlear implants, the strongest precedent: 16 to 22 contacts, a sparse code, and a brain that adapts to it.

Stimulation resolution should be counted in distinguishable percepts and task improvement, not electrodes or pulse rates. A deep-brain stimulator delivers over a hundred pulses a second and transmits no semantic bits at all; it is a controller, not a channel.

Who is building what

Academic consortia hold most of the peer-reviewed human evidence. Companies are building the implantable packages that deployment requires. Search the record, or filter by interface.

BrainGate consortiumBrainGate2 pilot trialUtah arrays in motor and speech cortex with percutaneous connectors

Reported evidence: Twenty years of human data, from 2006 cursor control to 2026 home speech; across its first 14 participants, arrays recorded spiking on an average of 35.6% of electrodes for up to 7.6 years.
Announced next step: Fully implanted wireless hardware and unsupervised recalibration for daily use.
Unresolved risk: A connector through the skull, researcher-built systems, and a small cohort whose best results come from a few participants.

UC Davis Neuroprosthetics LabSpeech neuroprosthesis256 electrodes in ventral precentral gyrus decoded to text and synthesised voice

Reported evidence: 97.5% word accuracy over 8.4 months; later more than 3,800 hours of independent home use at an average 56.1 words a minute; brain-to-voice synthesis with about 10 ms of neural-to-voice computation.
Announced next step: Instantaneous expressive voice, and replication in more participants.
Unresolved risk: A fixed voice decoder degraded noticeably within about two weeks; results rest on one long-term participant.

Stanford Neural Prosthetics Translational LabHandwriting, speech, and typing decodersRepresentation-first decoding: handwriting, phonemes, and attempted finger movements

Reported evidence: 90 characters a minute from attempted handwriting (2021) and 62 words a minute from attempted speech (2023) on essentially the same class of array.
Announced next step: Open-vocabulary accuracy and lower calibration burden.
Unresolved risk: Language-model priors can make fluent output that the neural signal did not uniquely specify.

UCSF Chang LabHigh-density ECoG speech253-electrode surface grid over speech sensorimotor cortex

Reported evidence: Median 78 words a minute with text, synthesised audio, and avatar output after under two weeks of training.
Announced next step: Higher accuracy and longer-term stability from the cortical surface.
Unresolved risk: Around 25% word error at record speed; one participant and a craniotomy.

SynchronStentrodeElectrode array delivered through the jugular vein into a cortical vein

Reported evidence: First-in-human studies in people with severe paralysis delivered a thought-controlled digital switch for ordinary computer tasks, with no open brain surgery.
Announced next step: Larger pivotal trials for assistive computer control.
Unresolved risk: Distance from neurons caps bandwidth at clicks and selections; vascular-device risks replace craniotomy risks.

Precision NeuroscienceLayer 7 cortical interfaceThin-film high-density surface electrodes inserted through a narrow skull slit

Reported evidence: Temporary high-density recordings during neurosurgery; a regulatory clearance for short-term use was reported in 2025.
Announced next step: Chronic implanted communication systems.
Unresolved risk: Short-term intraoperative data does not establish chronic communication performance.

NeuralinkN1 / PRIME studyRobot-inserted flexible polymer threads with 1,024 electrodes and a fully implanted wireless package

Reported evidence: Participants have used cursor control at home; in the first participant most threads retracted after implantation, and the company compensated in software.
Announced next step: Speech decoding, more participants, and higher channel counts.
Unresolved risk: Peer-reviewed, long-duration communication outcomes are not yet published on the same curve as academic systems.

ParadromicsConnexusDense microwire bundles with integrated electronics for 1,600+ channels

Reported evidence: High-channel-count microwire recording, with a first brief in-human recording reported in 2025.
Announced next step: A chronic speech-restoration trial.
Unresolved risk: Channel count must translate into stable independent dimensions, inside a power and heat budget.

Blackrock NeurotechUtah arrayRigid silicon microelectrode arrays

Reported evidence: The sensor behind most human intracortical BCI records from 2006 to 2026.
Announced next step: Higher channel counts and implanted wireless versions.
Unresolved risk: An eight-order-of-magnitude stiffness mismatch with brain tissue drives micromotion and scarring.

University of Pittsburgh Rehab Neural Engineering LabsBidirectional motor and touch BCIIntracortical microstimulation of somatosensory cortex for artificial touch

Reported evidence: Adding stimulated touch halved median object-transfer time with a robotic arm compared with vision alone.
Announced next step: Richer, more natural tactile percepts.
Unresolved risk: Stimulation evokes buzzing or tingling rather than texture; percepts interact as contacts are added.

Feinstein InstitutesDouble neural bypassIntracortical decoding coupled to patterned spinal and cortical stimulation

Reported evidence: A 2026 participant with severe tetraplegia regained functional object manipulation with persistent gains in movement and sensation.
Announced next step: Closed-loop restoration that improves the underlying function.
Unresolved risk: Better analysed as a controller than a channel; generalisation beyond single participants is unproven.

Cochlear implantsClinical standard of care16 to 22 electrodes along the auditory nerve's frequency map

Reported evidence: The most successful neuroprosthesis, restoring speech comprehension to more than a million people with a sparse engineered code.
Announced next step: Better music and noise performance.
Unresolved risk: A reminder, not a template: a few contacts are useful, but they are not equal to natural hearing.

Company capabilities are reproduced from publications and public announcements. Installed contact counts are not usable channels, and an announced trial is not a published outcome.

Part II: Useful bandwidth

The unit that matters is correct words per day

A peak rate describes a session. The rate a person experiences is set by how many hours the system works, how much of it is spent recalibrating, and how many words come out wrong.

Figure 4 · Interactive model

How many correct words survive a day?

A record is set in a session. A life is lived in days. Useful bandwidth is correct information divided by all the time it took, including calibration and downtime.

Rate while in use56 words/minWord error rate3.0%Hours in use per day6.6 hSupervised calibration per day0 minSystem availability95%

2026 home use20,435correct words/day

≈ 8.4 bit/s averaged over a sixteen-hour waking day

At full rate, all hours: 22,176
Lost to calibration: 0
Lost to downtime: 1,109
Lost to errors: 632

An average adult speaks on the order of ten to twenty thousand words a day. Try the laboratory record against home use: a faster system used briefly with a quarter of its words wrong delivers far less than a slower one available all day.

Calculation & assumptions

Correct words per day = rate × 60 × (hours in use − calibration) × availability × (1 − word error rate). The waking-day figure converts correct words to an upper-bound bit rate at five characters a word and log₂27 bits a character, then spreads it across sixteen hours.

Only the home-use hours are derived from the record: about 3,800 hours across roughly 19 months is about 6.6 hours a day. Session lengths, calibration minutes, and availability in the other scenarios are illustrative. Word error in the home-use scenario is set between the >99% prompted accuracy and the 92.3% of free sentences rated mostly correct.

Figure 4: Illustrative structure with plausible magnitudes, not a clinical benchmark. Words corrected by a language model count as correct here, which is right for usefulness but overstates what crossed the neural interface; a word the user could not have produced at all has no place on this ledger.

AvailabilityAn interface that works all day at a moderate rate beats a faster one that needs an engineer in the room.

ErrorAt a quarter of words wrong, correction consumes much of the speed advantage, and the user bears the cost of every error.

CalibrationThirty supervised minutes a day is a direct tax on throughput and on independence.

What pushes useful bandwidth up next

01

Decoders that track drift

Unsupervised recalibration during ordinary use turns a slowly changing signal into a stable interface without anyone present.

Condition: adaptation that does not also learn the user’s errors

02

Fully implanted packages

Wireless, sealed devices remove the percutaneous connector, the largest barrier to living with a BCI outside a study.

Condition: on-implant compression that keeps the features speech decoding needs

03

Wider, not denser, coverage

Sampling more of speech and hand cortex, including sulcal banks, adds new dimensions where adding contacts to one patch adds redundancy.

Condition: effective dimensionality that actually rises with added coverage

04

Models that transfer between people

Pretraining on pooled recordings could shrink calibration from hours to sentences for each new participant.

Condition: enough multi-site data, shared under consent terms that allow it

Part III: Where progress is stuck

The remaining constraints are biological and logistical

The field has moved from detecting a controllable signal, to decoding symbols quickly, to stabilising across days. The leading systems are working on the next two steps: years of independent use, and implantation routine enough to deploy.

Chronic stability

Implants survive for years while the neurons they hear change. Drift is gradual and measurable, and a decoder that is not continuously adapted loses accuracy in days to weeks.

Calibration burden

Every supervised minute of recalibration is subtracted from useful bandwidth and requires someone else to be present. Zero-calibration daily use is the target that matters most for independence.

The heat budget

A skull-mounted implant can dissipate only tens of milliwatts before warming tissue. Channel count therefore becomes a power-per-channel problem, and raw streaming is ruled out.

Surgery and hardware

Percutaneous connectors, craniotomies, and specialist centres limit who can receive a device. Deployment, not decoding, sets how many people the curve reaches.

The language-model question

Fluent output can hide weak neural evidence. Without evaluation on improbable sentences, accuracy figures cannot separate what the brain said from what the model guessed.

Cohort scale

The strongest results come from a handful of people. A one-participant record is evidence of possibility, not a platform.

An optimistic view, with conditions

Restoring conversation does not require reading thought

Speech cortex does not need to reveal every representation in the brain. It needs to supply enough stable information about attempted articulation for a sequence model to pick the intended sentence. The evidence through 2026 shows that a few hundred well-placed electrodes can already meet that narrower goal, in a few people.

Now to 2030

The threshold for consequence

For many people with locked-in paralysis, 20-30 words a minute at low single-digit error, sub-second latency, no daily calibration, and all-day availability would transform communication. The 2026 results meet parts of that in individual participants.

2030s

Reproducible, implanted, wireless

The same performance across dozens of participants, from sealed devices, with years of stability and a procedure that does not require a research hospital.

Longer horizon

Natural speed and a return channel

120-160 words a minute at conversational accuracy, about 40 bits a second, plus stimulation that supplies touch and proprioception good enough to close fine motor loops.

View the annual scorecard

Metric	Record through September 2026	Parity with natural function
Conversational rate	32 wpm sustained over months; 56 wpm averaged at home; 78 wpm laboratory peak	120-160 wpm
Open-vocabulary error	23.8% WER at 125,000 words (2023); >99% on prompted sentences at home (2026)	Under 3% in unprompted conversation
Daily calibration	Minutes of supervised retraining to none in the best home system	Zero, with background adaptation
Independent use	3,800+ hours over ~19 months, one participant	All-day, across many participants
Usable channel lifetime	Spiking on ~36% of electrodes, arrays recording up to 7.6 years	A decade, without revision surgery
Write resolution	Tactile percepts that aid control; a few dozen phosphenes from 96 contacts	Distinguishable percepts that scale with contacts

Baselines are reported results, most from single participants. Parity targets are editorial benchmarks derived from natural speech and typing, not forecasts.

There is more than one finish line

Signal detectedA controllable neural signal can drive a binary choice.
Symbols decodedLetters, phonemes, or words decoded at useful speed in a session.
Stable across daysAccuracy holds without supervised retraining.
Independent for yearsA person uses it all day, at home, without researchers.
ReproducibleThe same result across many people and centres.
DeployableImplanted wirelessly through a routine procedure and paid for.

Sources, method, and boundaries

Figure 1 plots rates as reported by each study, converted to words per minute at five characters a word where needed; it is not a benchmark series, and the points differ in vocabulary, error, protocol, and participant. Figures 3 and 4 are illustrative models rather than measurements. Bit rates for text are upper bounds that treat 26 letters and a space as equiprobable; real English carries less information per character. Nothing here is clinical advice.

Physical channel: An implanted contact, whether or not it records a usable signal.
Usable channel: A contact meeting a stated signal-quality or task-tuning criterion at a given time after implantation.
Word error rate: Substitutions, insertions, and deletions divided by words intended, usually measured after a language model.
Useful bandwidth: Correct information delivered per unit of total time, including calibration and downtime.

Nature: Neuronal ensemble control of prosthetic devices (2006)Nature: Reach and grasp using a neurally controlled robotic arm (2012)Lancet: High-performance neuroprosthetic control (2013)PNAS: High-speed spelling with a noninvasive BCI (2015)eLife: High performance communication by people with paralysis (2017)Nature: High-performance brain-to-text via handwriting (2021)Science: A brain-computer interface that evokes tactile sensations (2021)Nature: A high-performance speech neuroprosthesis (2023)Nature: Speech decoding and avatar control (2023)NEJM: An accurate and rapidly calibrating speech neuroprosthesis (2024)Nature: An instantaneous voice-synthesis neuroprosthesis (2025)Nature Neuroscience: Restoring rapid natural bimanual typing (2026)Nature Medicine: Long-term independent use of an intracortical BCI (2026)medRxiv: Long-term performance of arrays in 14 BrainGate participants JCI: Visual percepts evoked with an intracortical array (2021)JNIS: Motor neuroprosthesis implanted with neurointerventional surgery (2021)Nature Neuroscience: How advances in neural recording affect data analysis (2011)Science Advances: Languages communicate information at a similar rate (2019)

Read More

The 20 books most relevant to this report, drawn from the reading lists of people worth listening to, via TopBooks.