The branches · 7 of 7

The link to machines

A direct connection between a brain and a computer, and how much can pass through it.

The hard version of merging with machines: a wire, a chip or a sensor that carries signals between a brain and a computer directly, with no hands, eyes or voice in between. The project's second number, how much information can pass through such a link each second, is plotted here.

7 upgrades

furthest stage reached: in wide use

The second number: how much can pass between a brain and a machine

8 figures, 2006 to 2024

The merger argument has rested on one number since 2017: how much information can pass between a brain and a computer each second, and whether that can beat the mouth and the hands. This is the record of what has actually been demonstrated, each figure in the unit its source used, from the peer-reviewed papers and the makers' own releases, against the rates a reader already has. Two kinds of figure are kept apart, because a rate of text and a rate of pointing at a screen cannot share an axis without an assumption the sources did not make.

The best figure so far for text is 78 words a minute, from attempted speech decoded by a brain implant, reported on 23 August 2023: Sean Metzger, Kaylo Littlejohn, Alexander Silva, David Moses, Margaret Seaton and colleagues at the University of California, San Francisco (Edward Chang, senior author). Speech in conversation runs at about 160 words a minute, so the fastest brain link yet talks at about half the speed of talking, with one word in four wrong.

For pointing at a screen, the best figure is 8 bits a second, the maker's own figure and not yet in a peer-reviewed paper, reported on 8 May 2024; the best peer-reviewed figure is 4.16 bits a second, from 21 February 2017. An engineer with a mouse scores about ten.

Text out of the head, in words a minute How fast a person produced text through a brain link, in words or characters a minute as the paper gave it, with the vocabulary and the error rate beside it. A high rate with a small vocabulary or a high error rate is not the same thing as a high rate with neither; the table says which is which.
0 40 80 120 160 20182020202220242026 Speech in conversation, 160 A fast typist, 120 Typing on a phone, 23 Feb 2017: 39.2 correct characters a minute (shown as 7.8 words a minute). Chethan Pandarinath, Paul Nuyujukian and colleagues at Stanford (Krishna Shenoy and Jaimie Henderson, senior authors), in the BrainGate2 pilot trial. May 2021: 90 characters a minute (shown as 18 words a minute). Francis Willett and colleagues at Stanford (Jaimie Henderson and Krishna Shenoy, senior authors), in the BrainGate2 pilot trial. Aug 2023: 62 words a minute. Francis Willett, Erin Kunz, Chaofei Fan and colleagues at Stanford (Jaimie Henderson, senior author), in the BrainGate2 pilot trial. Aug 2023: 78 words a minute. Sean Metzger, Kaylo Littlejohn, Alexander Silva, David Moses, Margaret Seaton and colleagues at the University of California, San Francisco (Edward Chang, senior author). Aug 2024: 31.6 words a minute. Nicholas Card, Sergey Stavisky, David Brandman and colleagues at the University of California, Davis, in the BrainGate2 pilot trial.
  • a demonstrated figure, from a peer-reviewed paper
  • a demonstrated figure, from the maker's own release
  • the best figure so far
  • a rate a reader already has

Characters a minute are shown as words at five characters a word, the convention typing speed is measured by (so 90 characters a minute is shown as 18 words a minute, which is also how the 2023 Stanford paper restates the 2021 figure).

Pointing at a screen, in bits a second How much information a person's cursor control carried, in bits a second, measured by the field's grid task: targets appear on a grid, the person moves to and selects them, and the rate counts how many right selections were made per second against how many targets there were to choose from. It measures pointing, not language, so a figure here is not comparable to a figure for text.
0 20 40 20182020202220242026 Engineers with a mouse, 10 Speech, as information, 39 Feb 2017: 4.2 bits a second. Chethan Pandarinath, Paul Nuyujukian and colleagues at Stanford (Krishna Shenoy and Jaimie Henderson, senior authors), in the BrainGate2 pilot trial. May 2024: 8 bits a second. Neuralink, in its own progress post; measured by the company, not in a peer-reviewed paper.
  • a demonstrated figure, from a peer-reviewed paper
  • a demonstrated figure, from the maker's own release
  • the best figure so far
  • a rate a reader already has

Every figure, oldest first

  1. 13 July 2006 no rate stated peer-reviewed paper

    The first BrainGate participant moving a cursor, opening simulated e-mail and operating a television through a 96-electrode array in the part of the brain that plans hand movement. The paper states no rate: no bits a second, no characters a minute.

    Who Leigh Hochberg and colleagues at Massachusetts General Hospital and Brown University, with Cyberkinetics Neurotechnology Systems (John Donoghue, senior author). In whom A tetraplegic man, MN (Matt Nagle), three years after a spinal cord injury. Repeated The paper's readable text makes no claim of repetition; a supplementary video names a second trial participant, but the text does not report his results. Source doi.org. Nature 442, pages 164 to 171, published 2006-07-13. The abstract names the 96-microelectrode array and describes the neural cursor; it gives no rate. Belongs with A wired brain implant lets a paralysed person control a cursor Note Here as the series' starting mark: the first hundred-electrode implant in a person, from which every later point descends. The implant date (June 2004) is not in the paper's readable text and is not stated here.

  2. 21 February 2017 4.16 bits a second peer-reviewed paper

    Cursor control on a grid task: 4.16 bits a second for participant T5 on a 9 by 9 grid, which the paper calls, to its authors' knowledge, the highest documented brain-computer interface communication rate for a person with motor impairment. On the 6 by 6 grid the averages were 2.2 (T6), 3.7 (T5) and 1.4 (T7) bits a second.

    Who Chethan Pandarinath, Paul Nuyujukian and colleagues at Stanford (Krishna Shenoy and Jaimie Henderson, senior authors), in the BrainGate2 pilot trial. In whom Three people with paralysis: T6, T5 and T7, as above; the 4.16 figure is T5's. Repeated Three participants over two to five days each; the 4.16 figure is one participant's average over his sessions on the denser grid. Source doi.org. eLife 6:e18554, published 2017-02-21, Results and Table 1. Belongs with A wired brain implant lets a paralysed person control a cursor

  3. 21 February 2017 39.2 correct characters a minute peer-reviewed paper

    Copy typing on an on-screen keyboard by cursor control: 39.2 correct characters a minute (the paper's own 7.8 words a minute) for participant T5 with an optimised keyboard layout; 31.6 for T6 and 13.5 for T7. Free typing by T6 ran at 24.4 correct characters a minute over three days.

    Who Chethan Pandarinath, Paul Nuyujukian and colleagues at Stanford (Krishna Shenoy and Jaimie Henderson, senior authors), in the BrainGate2 pilot trial. In whom Three people with paralysis: T6, a woman of 51 with motor neurone disease; T5, a man of 63 with tetraplegia from a spinal cord injury; T7, a man of 54 with motor neurone disease. Repeated Three participants, over two to five days each; the paper reports every one. The 39.2 figure is T5's average, not a one-off peak. Source doi.org. eLife 6:e18554, published 2017-02-21. The Results give the per-participant figures; the Discussion restates them as 6.3, 7.8 and 2.7 words a minute. Belongs with A wired brain implant lets a paralysed person control a cursor

  4. 12 May 2021 90 characters a minute peer-reviewed paper

    Attempted handwriting decoded into text: 90 characters a minute with 94.1 per cent raw accuracy online, and above 99 per cent offline with a general-purpose autocorrect. The paper compares it with smartphone typing in the participant's age group, 115 characters a minute.

    Who Francis Willett and colleagues at Stanford (Jaimie Henderson and Krishna Shenoy, senior authors), in the BrainGate2 pilot trial. In whom One man, T5, paralysed from the neck down by a spinal cord injury. Repeated A single participant. The paper reports consistency across its own decoder training runs and across ten days of data, not a second person. Source doi.org. Nature 593, pages 249 to 254, published 2021-05-12; the figures are in the abstract. Belongs with A wired brain implant lets a paralysed person control a cursor

  5. 23 August 2023 62 words a minute peer-reviewed paper

    Attempted speech decoded into text: 62 words a minute, with a 23.8 per cent word error rate on a 125,000-word vocabulary and 9.1 per cent on a 50-word vocabulary. The paper calls it 3.4 times the previous record and says it begins to approach the speed of natural conversation, which it puts at 160 words a minute.

    Who Francis Willett, Erin Kunz, Chaofei Fan and colleagues at Stanford (Jaimie Henderson, senior author), in the BrainGate2 pilot trial. In whom One person, T12, with bulbar-onset motor neurone disease, unable to produce intelligible speech. Repeated A single participant, over five days of attempted speech aloud and three of silent attempts; the paper says the results must be confirmed in more people. Source doi.org. Nature 620, pages 1031 to 1036, published 2023-08-23; the figures are in the abstract and the first results paragraph. Belongs with A brain implant turns intended speech into text at conversational speed

  6. 23 August 2023 78 words a minute peer-reviewed paper

    Silently attempted speech decoded into text, speech sounds and a talking avatar: a median of 78 words a minute (78.3 in the main text) with a median word error rate of 25 per cent (25.5), on a vocabulary of 1,024 words, from high-density recordings on the surface of the speech cortex. The paper calls this large-vocabulary decoding; the 1,024 words are those of the sentences used to test it (249 sentences, not used in training), and a larger vocabulary of 39,378 words was tried only offline, at a word error rate of 27.6 per cent.

    Who Sean Metzger, Kaylo Littlejohn, Alexander Silva, David Moses, Margaret Seaton and colleagues at the University of California, San Francisco (Edward Chang, senior author). In whom One woman, the trial's participant Bravo-3, with severe limb and vocal paralysis. Repeated A single participant; the paper's readable text makes no claim of repetition. Source doi.org. Nature 620, pages 1037 to 1046, published 2023-08-23. The 78 words a minute and 25 per cent figures are in the abstract. The vocabulary is read from the full text (open copy, PubMed Central PMC10826467, and the journal's page, both read on 2026-10-17): the legend of Figure 2 says its panels, including the decoded words a minute, "were computed using the real-time evaluation trials with the 1024-word-General sentence set"; the text of the Results gives the median rate as 78.3 words a minute and the median word error rate as 25.5 per cent on those trials, and says the sentences were 249 drawn at random from that set, not used in training. The sentence that gives 78.3 follows one about a freeform test, and the text cites its figure panel as 2f where the legend labels the decoded-rate panel 2e; the legend's statement that panels b to g were computed on the 1024-word real-time trials is what ties the rate to that vocabulary. Belongs with A brain implant turns intended speech into text at conversational speed

  7. 14 August 2024 31.6 words a minute peer-reviewed paper

    Attempted speech decoded into text at the participant's own pace: 31.6 words a minute with a 2.5 per cent word error rate on a vocabulary of over 125,000 words, sustained at 97.5 per cent accuracy over 248 hours of use across eight months. Slower than the 2023 decoders and far more accurate; included because the series records every demonstrated figure, not only the records.

    Who Nicholas Card, Sergey Stavisky, David Brandman and colleagues at the University of California, Davis, in the BrainGate2 pilot trial. In whom One man of 45 with motor neurone disease and severe speech impairment, with four 64-electrode arrays (256 electrodes) in the left precentral gyrus. Repeated A single participant, over 84 research sessions and eight months; the accuracy held across them. Source doi.org. New England Journal of Medicine 391, pages 609 to 618; published online 2024-08-14 and dated 2024-08-15 in the print issue (the date given is the first). The 31.6 words a minute and 2.5 per cent figures are in the Results, the 97.5 per cent and 125,000 in the abstract. The online day rests on UC Davis Health's own release of 2024-08-14, which says the study "was published today" in the journal (https://health.ucdavis.edu/news/headlines/new-brain-computer-interface-allows-man-with-als-to-speak-again/2024/08), and agrees with the PubMed record (entered 2024-08-14, print date 2024-08-15) and the DOI registry (record created 2024-08-14). The journal's own page sits behind a bot check that was not bypassed. Belongs with A brain implant turns intended speech into text at conversational speed

The rates a reader already has

The series is data/bandwidth.toml in the public repository, and the rules for a point are in its layout document there. A figure missing or misread: say so.

← Inheritance