The Neuralink launch presentation — a brain-machine interface offered as the answer to AI risk

prediction is the right kind and the proposal filed it correctly. But the id fuses two claims of different shapes, with different due dates and different verdicts, and an entry that grades them as one produces a wrong answer in whichever direction it leans.

The id's noun is symbiosis — a thesis about how humanity survives artificial intelligence, which has no natural due date and was not attempted. The proposal's grading hook is a schedule — "a human patient hoped for by the end of 2020; the first human implant came January 2024" — which is an ordinary dated engineering claim that was delivered, three years and one month late. Grade the pair together and you can write either "Neuralink failed" or "Neuralink delivered," and both sentences would be true of half the document and false of the whole. hinton-radiologists-2016 had to split one forty-second remark into three speech acts for the same reason. This file does the same, and the split is most of its work.

Two corrections before anything else, because this canon's method is grading dated text at its date and both corrections are about dates.

First, the sentence is older than the event. The formulation that travelled furthest — "From a long-term existential standpoint, that's, like, the purpose of Neuralink, to create a high-bandwidth interface to the brain such that we can be symbiotic with AI. Because we have a bandwidth problem. You just can't communicate through your fingers. It's just too slow." — is from the Joe Rogan Experience, November 2018, roughly eight months before the launch event. It is the version Zheng and Meister quote in Neuron when they set out to refute it. The July 2019 presentation restated the thesis in front of hardware; it did not originate it. The id's year is right for the device and about eight months late for the argument, and a file that grades sentences should say which sentence it is grading and when it was said.

Second, the event had a stated purpose, and the thesis was not it. Musk opened with "The main reason for doing this event is recruiting." That is unusually useful: a forecaster's own success criterion for the occasion, stated by the forecaster, at the start, on the record. It has to be graded alongside the claims, and on that criterion the event is a clean hit — see §3.4. An entry that grades only the propositions and ignores the speaker's own stated objective is grading the wrong document.

The descent

descends_from: [culture-banks-1987, superintelligence-2014]

culture-banks-1987 is documented, not inferred, and this canon already recorded it. "Neural lace" is Iain M. Banks's term for the Culture's brain interface, and it is where the implant got its name; the Banks file says so, and then says the thing this file inherits: "A naming is not a delivery, and an entry that let a borrowed word count as a fulfilled forecast would be doing the thing this section exists to prevent." Seven years on, that warning can be graded rather than merely issued. The borrowed word is still borrowed. What was built is a 1,024-channel motor-cortex recorder for people with quadriplegia, which is a real and valuable thing and is not a neural lace by any reading of Banks — the Culture's lace is grown through the whole cortex in infancy, is read-write, and backs up a personality. The distance between the name and the device is the entry's subject.

superintelligence-2014 is the premise, and the endorsement is dated and public. On 3 August 2014 Musk posted: "Worth reading Superintelligence by Bostrom. We need to be super careful with AI. Potentially more dangerous than nukes." That is two years before Neuralink was founded and five before the event, and the 2019 argument's premise — that a sufficiently capable AI leaves unaugmented humans behind by default, even without malice — is that book's thesis in compressed form. Musk's own phrasing at the event, "Even under a benign AI, we will be left behind," is close to a one-line summary of Bostrom's point that catastrophe does not require hostility.

But I am listing an ancestor on an inference, and the ancestor's own file warns against exactly this move. superintelligence-2014's sources section records, under Not obtained: "Primary evidence that Superintelligence contributed causally to the founding of OpenAI. Altman's and Musk's endorsements are documented; the causal step is asserted everywhere and sourced nowhere I could reach." The same caution applies here and I am not exempting myself from it. What I can show is a dated public endorsement of the specific book by the specific person, and a premise that matches the book's. What I cannot show is that he was reading it when he founded the company. The header claim is that the argument came out of that argument, not that the company came out of that book, and a reader who thinks that distinction too fine should downgrade the edge accordingly.

The ancestor I wanted and could not establish

kurzweil-2005. The Singularity Is Near is where the merger thesis lives in this canon: not "build a safe AI" and not "regulate it" but become it — nanobots connecting the neocortex to the cloud, the biological and non-biological parts of intelligence fusing, humans transcending biology rather than competing with it. That is Musk's 2019 argument with different plumbing, published fourteen years earlier by a man whose forecasts this canon already grades. I found no evidence that Musk's version came out of Kurzweil's, and descent in this canon means came out of, not resembles. It goes here instead of the header. The two files should nonetheless be read together, because the merger thesis has now been offered twice by prominent forecasters and attempted by neither.

The siblings, which are the point of the set

This is the canon's fourth graded Musk claim, and the set is deliberate. musk-robotaxi-2019 asked for this file by name. It quotes the owner's proposal — that this entry be "the second reading on the same forecaster, which is what makes a track record a base rate instead of an anecdote" — calls that "exactly right," and then says of itself: "this file cannot do it alone — one predictor graded once is an anecdote no matter how carefully it is graded." It asks that its own section 4 "be read beside it and probably shortened." §3.6 below does that work. The other two exist because tesla-volume-production was commissioned to fix a hole, on the grounds canon/proposals.md gave for it — "a canon that grades that and holds nothing on the same person's successes produces a base rate that is not a base rate" — and spacex-booster-reuse was written to pair with it.

That file also left a warning this one has to answer: "An entry commissioned to supply a hit is under structural pressure to find one." The pressure here runs the other way, since this entry was commissioned to supply a second miss and the id it was filed under names the half that missed. §3.4 is where I put what I did about it.

So the set now reads: two claims that landed years late (tesla-volume-production, spacex-booster-reuse), one that has not landed at any multiple (musk-robotaxi-2019), and this one, which landed 3.1× late and is therefore the only member of the set that sits inside the independent rule the canon uses to grade this forecaster. That is this file's most mechanical contribution and it is worth more than another miss would have been.

Two entries are cited in prose and are not ancestors. mycin-1976 is the canon's record of a system that performed and was never deployed, for institutional reasons its builders wrote down; hinton-radiologists-2016 is the record of a capability that arrived and displaced nobody, for the same class of reason. Both are diagnostic here — the binding constraint on this forecast was institutional, not technical — and neither is upstream of it. Descent is not the same as diagnosis.

What it is

The event

On 16 July 2019, live-streamed from the Morrison Planetarium at the California Academy of Sciences in San Francisco, Neuralink made its first public presentation. The company had been founded in 2016 and had said almost nothing in public since. As of that month it had raised $158 million, of which $100 million was Musk's, and employed about 90 people.

Musk presented, with Max Hodak, then the company's president, and Matthew MacDougall, its head neurosurgeon. (Several other founders — DJ Seo, Philip Sabes, Vanessa Tolosa — are named in coverage of the company at that period; I could not confirm from any source I reached which of them spoke on the night, and the video was not retrievable. See Sources.)

The presentation had three parts, and they do not sit comfortably together:

A medical device, described conservatively. MacDougall's framing was that the system was "only intended for patients with serious unmet medical diseases," with the initial target being people with complete paralysis from upper spinal-cord injury. Nothing in that half of the presentation is transhumanist and nothing in it has since been retracted.

A thesis about AI, described expansively. Musk's, and the reason the event was covered the way it was.

A schedule. Musk's, and the reason this entry is a prediction.

The two claims, verbatim

The thesis. Reported consistently across outlets that had reporters in the room:

> Even under a benign AI, we will be left behind. With a high-bandwidth > brain-machine interface, we will have the option to go along for the ride.

and

> With a high-bandwidth brain interface, I think we can have the option of > merging with AI.

with the goal stated as achieving "some sort of symbiosis with artificial intelligence." Musk also said, of the prospect, that it "will take a long time" and that participation would be optional — a hedge that is real and should be recorded, since much of the hostile commentary omits it.

The schedule.

> We hope to have this in a human patient by the end of next year.

with regulatory approval targeted for the end of 2020 and, per MIT Technology Review's report the following morning, a trial start intended for Q2 2020, pending FDA approval. That article — Antonio Regalado, 17 July 2019 — called the timeline "highly ambitious and pretty unlikely" on the day. The contemporaneous independent grade was correct and is worth noting, because a great deal of the later commentary implies nobody could have known.

The device that was presented, and what happened to it

The hardware shown on the night was not the hardware that was eventually implanted, and this is rarely mentioned.

The N1 chip held 1,024 channels in a 4 mm × 5 mm package. The presented architecture was four such chips, implanted through small skull openings — three in motor areas, one in a somatosensory area — with wires tunnelled under the scalp to an inductive coil behind the ear, pairing wirelessly to a battery-powered pod called the Link. Threads were to be inserted by a purpose-built neurosurgical robot at micron precision, avoiding surface vasculature.

By the August 2020 update that design was gone. What replaced it, and what is in people today, is a single coin-sized unit, roughly 23 mm across and 8 mm thick, seated in a hole in the skull, with the electronics within a centimetre of the electrodes and nothing behind the ear at all. The first human implant, in January 2024, used one device with 64 threads carrying 1,024 electrodes — that is, one quarter of the channel count the 2019 architecture called for.

This matters for grading in a specific way: the schedule claim was made about a device that was abandoned within thirteen months. That is normal engineering and is not by itself a criticism. It is a criticism of the schedule, because a seventeen-month horizon to first-in-human was given for an architecture that did not survive to month thirteen.

The white paper

The presentation was accompanied by An integrated brain-machine interface platform with thousands of channels, credited on its title page to "Elon Musk, Neuralink" — a byline that drew comment at the time — posted to bioRxiv in July 2019 and published in the Journal of Medical Internet Research, 2019;21(10):e16194, doi:10.2196/16194.

Its content is real and modest relative to the presentation. Arrays of flexible polymer threads carrying up to 3,072 electrodes across 96 threads; a neurosurgical robot inserting six threads (192 electrodes) per minute; a spiking yield of up to 85.5% — in freely moving rats. Not primates, not humans. The paper describes an engineering platform, and the claims it makes are platform claims.

The journal published a commentary alongside it on 31 October 2019 — Pisarchik, Maksimenko and Hramov — which is the most useful contemporaneous technical assessment I found. It credits the integration genuinely, and then notes that the paper asserts the electronics can stimulate every channel but does not present those results; questions whether stimulation and recording can run simultaneously without artefact; and names biocompatibility as "one of the main obstacles for clinical applications of implanted devices," with an explicit trade-off between impedance and biocompatibility.

Four and a half years later, the first human implant's threads retracted from the tissue. I want to be careful about what that does and does not show. Neuralink attributed the retraction to the brain moving further than expected, which is a mechanical explanation and not the same thing as a biocompatibility failure. What the commentary establishes is narrower and still worth having: three neuroscientists reviewing the platform in the same issue as the paper wrote down that the long-term behaviour of these electrodes in tissue was the unresolved question standing between it and clinical use, and that is the family of problem the first implant ran into. It is a good example of the difference between a demo in a rat and a device in a person, which is the difference the FDA was asking about for the next four years.

Why a reading would cite it

Five occasions, and the disclosure that has to come with them.

One: when AI risk is answered with a hardware bet instead of a governance one

This is the entry's primary use and the reason the owner proposed it. The "Safety and alignment" and "Policy and regulation" lenses will keep meeting claims of the form this technology is how we survive AI. This is the canon's best-documented instance of that move, made by someone with the capital to actually try it, and it now has seven years of record attached.

The finding a reading gets is not "he was wrong." It is more specific and more useful: the bet was funded, the company succeeded on its own medical terms, and the AI-risk thesis it was founded on was never tested by it. In the six and a half years between the presentation and the most recent count I could source — 21 participants, reported by Reuters on 28 January 2026 and repeated by MIT Technology Review in June — not one of those implants was performed for a reason having anything to do with artificial intelligence, and not one was performed on a person without a disabling neurological condition (§3.3). A reading meeting a hardware answer to AI risk can ask the question this case makes concrete: what would count as the first instance of the thesis being attempted, and is anything in the roadmap that thing?

Two: when a timeline is offered for AI in a regulated domain

This is the transferable engineering lesson and §3.5 develops it. The forecast missed because the binding constraint was an evidentiary standard held by an institution the forecaster does not control and had not yet approached — not because the threads did not work. The canon has two other entries on the same fault line (mycin-1976, hinton-radiologists-2016) and this is the third and the cleanest, because here the regulator's objections were reported in detail and the slip can be attributed rather than guessed at.

Any reading grading an AI-in-medicine, AI-in-aviation, AI-in-finance or autonomous-vehicle timeline should apply the test in §3.5 before believing the date.

Three: it repairs a base rate this project already runs on

musk-robotaxi-2019 records WIRED's May 2025 analysis — nineteen years of pledges — as "the most useful independent grade found," reporting timelines that overshoot by 2× to 5× with the underlying ambition often directionally real. That file then honestly notes the rule does not fit the robotaxi claim, which is already past 5× on a number that has not arrived at any multiple.

This claim fits it exactly: 3.1× (§3.2). So the canon now holds a Musk forecast that confirms the rule, two that landed late in its range or beyond (tesla-volume-production, spacex-booster-reuse), and one that broke it. Four graded claims is not a base rate either, and this file should not pretend otherwise — but it is the difference between an anecdote and the beginning of a distribution, which is precisely what jobs/canon.md says the kind is for.

Four: when "bandwidth" is offered as the bottleneck between humans and AI

The argument is back, in 2026, in the discourse about agents and interfaces, and usually without the counter-argument. This entry carries the counter-argument in its strongest form, from a peer-reviewed paper that names Musk explicitly, and carries Neuralink's own headline number as the evidence for it (§3.3). That combination is rare and a reading should use it: the company's best result is also its critics' best exhibit.

Five: as the case where a slip was absorbed rather than conceded

§3.6. The record of the January 2024 delivery contains almost no trace of the December 2020 promise, and at least one outlet performed the correction on the forecaster's behalf on the day. That is a distinct way for a base rate to be destroyed and the canon did not have an example of it.

The disclosure this entry owes, under README rule 7

Rule 7 names musk-robotaxi-2019 explicitly: "no vendor is ever flattered by a project that grades that vendor's predictions ... xAI is a candidate to compose the tweet — those two facts must never meet." This is the same forecaster and the rule applies unchanged.

And, as the robotaxi file worked out, the corruption risk here runs backwards. A model built by Anthropic is not tempted to flatter this subject. The convenient grade is the hostile one, and it is available in industrial quantity: the secondary writing on Neuralink is dominated by people who reached their conclusion before they reached the evidence, in both directions, and a file assembled from it would look fine at a glance.

Three specific guards, so the reader can check whether they held:

• §3.4 is the longest subsection in the grading, not a courtesy paragraph. What was actually achieved is stated at full weight, including every number that makes Neuralink look good, and including the one where the sceptical independent forecast undershot the company's real progress.
• The company's own claims are labelled as the company's own every time they appear — in particular the adverse-event record, which is a self-reported figure from a private company with no published trial results.
• The animal-welfare and regulatory-capture material appears only in "Commonly misused as" and in Sources, is reported as allegation and investigation rather than finding, and nothing in the grade rests on it. It is the easiest available way to make this file feel damning without making it more accurate.

What it got right, and what it got wrong

3.1 The two claims, with the four things each

Claim A — the schedule.

• The words: "We hope to have this in a human patient by the end of next year," with regulatory approval targeted by end of 2020 and a trial start intended for Q2 2020, pending FDA approval.
• Date made: 16 July 2019.
• Date due: 31 December 2020.
• What happened: First human implant 28 January 2024, announced by Musk on 29 January. Delivered, 37 months past the due date.

Claim B — the thesis.

• The words: "Even under a benign AI, we will be left behind. With a high-bandwidth brain-machine interface, we will have the option to go along for the ride," toward "some sort of symbiosis with artificial intelligence."
• Date made: 16 July 2019 (the formulation dates to November 2018).
• Date due: None stated at the event, and Musk explicitly said it "will take a long time." Restated in August 2024 with a date attached: non-disabled implantation "probably less than 10 years," and "if five years from now, we might be at a megabit." Those come due 2029 and 2034.
• What happened: Not attempted. Zero non-disabled implants by anyone, anywhere. The premise independently attacked at the root in a peer-reviewed journal, using Neuralink's own best number.

3.2 Claim A, graded

Delivered, late by a factor of 3.1.

The horizon promised on 16 July 2019 was 17.5 months. The elapsed time to delivery was 54.4 months. That is 3.1×, inside WIRED's 2×–5× band, which makes this the canon's cleanest confirmation of the only independent rule it has for this forecaster.

The intermediate record, because the shape of the slip is the interesting part:

• 16 July 2019 — human patient promised "by the end of next year"; regulatory approval targeted for end-2020.
• July 2020 — FDA breakthrough device designation, which is a review-priority status and not an approval to implant. It is routinely reported as though it were one.
• August 2020 — the redesigned device is demonstrated in a pig, Gertrude. The architecture presented in 2019 is already gone.
• April 2021 — a macaque plays Pong.
• early 2022 — Neuralink applies to the FDA for the first time. Thirty months after the date by which approval was supposed to have been granted.
• (2022) — the application is rejected. Reported by Reuters on 2 March 2023: concerns over the lithium battery, migration of the threads within the brain, and whether the device could be removed without damaging tissue, with dozens of issues to be addressed.
• 25 May 2023 — the FDA approves the trial.
• September 2023 — the PRIME study opens recruiting at the Barrow Neurological Institute, Phoenix.
• 28 January 2024 — first human implant. Noland Arbaugh, 29, quadriplegic; identified publicly on 20 March.

The single most damaging line in that list is the fifth. The 2019 claim was conditional — pending FDA approval — and the hedge was stated honestly. But the company did not test the condition for two and a half years after the date the condition was supposed to have been satisfied. A conditional whose condition the speaker never approaches is not a hedge. It is a forecast with a clause that makes it feel hedged, and this canon should say so, because the clause is the part most often cited in the forecast's defence.

The four Musk manoeuvres catalogued in musk-robotaxi-2019 §5 do not apply here, and the reason is that Claim A did not need any of them: it arrived. What was done to the record instead is §3.6.

3.3 Claim B, graded

Not attempted, and the premise is now contested on evidence that did not exist in 2019 — most of it Neuralink's own.

No company has implanted a healthy person, and none claims to have. Not Neuralink, not Synchron, not Paradromics, not Precision, not Blackrock, not Neuracle. Roughly 150 people have received a brain implant as of June 2026, against 67 volunteers in trials through the end of 2023, and every programme those numbers are drawn from is a therapeutic trial with a neurological inclusion criterion. I could not find a reported exception and I would expect one to be extremely well publicised. Neuralink's own programmes are Telepathy (cursor control for paralysis), CONVOY (an assistive robotic arm), a speech-decoding trial, and Blindsight (vision, breakthrough designation September 2024, not yet implanted in a human as of this writing). There is no symbiosis programme. There is no protocol under which a person without a medical indication could receive one, and no regulator that would approve it.

The premise took a direct hit, in a journal, from Caltech. Jieyu Zheng and Markus Meister, "The unbearable slowness of being: Why do we live at 10 bits/s?", Neuron, published December 2024 and carried in the journal's early-2025 issue. Its finding: human sensory systems gather on the order of 10⁹ bits/s, and human behavioural throughput is about 10 bits/s — and this ceiling appears to be set by central cognition rather than by the output channel. Section 5.3 is titled "The Musk illusion." It quotes the symbiosis statement, and concludes:

> Based on the research reviewed here regarding the rate of human cognition, we > predict that Musk's brain will communicate with the computer at about 10 > bits/s. Instead of the bundle of Neuralink electrodes, Musk could just use a > telephone, whose data rate has been designed to match human language, which in > turn is matched to the speed of perception and cognition.

And here is the part that makes this entry worth its length. Neuralink's flagship result is that Noland Arbaugh reached over 9 bits per second on Webgrid, roughly doubling the previous brain-interface record. Able-bodied people using a hand and a mouse score about 10. The company's best number is therefore a restoration of the human ceiling, not a breach of it — which is a magnificent clinical result and is precisely what Zheng and Meister predict and what the 2019 thesis requires to be false. Neuralink has not published this figure in a peer-reviewed venue; it is a company claim, and it is graded here as a company claim — but note which way the uncertainty runs. If the 9-plus is generous, the true figure is further below the ceiling, not above it. There is no version of this number that helps the thesis, which is why it is the strongest evidence in the file and why it did not have to come from a critic.

Independent voices on the premise, both sides recorded:

• Lee Miller (Northwestern), to MIT Technology Review, 29 September 2023: "The idea that we're going to hook up two people with bits of wire and do better than what you and I can do right now, speaking, is folly."
• Erik Hoel, neuroscientist, 10 September 2024: humans are "innately bottlenecked by biology, rather than merely inconvenienced"; the telepathy vision is "a rather useless product" at those rates.
• Against, and this belongs here: Matt Angle of Paradromics, in the same MIT Technology Review piece, argues bandwidth genuinely matters for speech restoration — "with a thousand electrodes, it will be as good as a cell phone transcribing your speech." That is a real use for the extra channels, it is a medical one, and by June 2026 speech decoding with voice cloning is among the demonstrated capabilities in the field. The bandwidth argument is correct for restoring a channel that is broken. It is the extension to augmenting an intact one that the 10-bit finding bites.

And the thesis had a cost, which is the entry's least-known 2026 fact. STAT, 5 January 2026, O. Rose Broderick: competitors, investors and former federal regulators say Neuralink's public rhetoric about machine-human symbiosis and implanting healthy people — rhetoric that "diverges sharply from the company's clinical work" — is hampering the whole brain-computer-interface industry's ability to win FDA approval as medical devices and to secure insurance coverage. The same piece reports that Neuralink recruited a senior official out of the FDA office that regulates it.

That is a claim by interested parties and is recorded as one. But it is a falsifiable, dated, on-the-record assertion by named-category sources that the symbiosis framing is negative-value for the medical programme it is attached to — which is a stronger and stranger finding than "the thesis has not happened yet," and it is the sentence a 2026 reading should take from this file.

One thing Claim B cannot be graded on, and this entry will not pretend otherwise. Its premise is conditional: if AI outstrips us, then we need this. The antecedent is the whole subject of this project's needle and is not this entry's to settle. What is gradeable is the consequent — whether the interface was built and whether it does what was claimed for it — and that is what §3.3 grades. A reading that cites this file as evidence about AI risk itself has misused it.

3.4 What it got right, and this canon should say so plainly

This section is not a courtesy. tesla-volume-production exists because a base rate carrying only failures is not a base rate, and the same discipline applies inside a single file.

The stated purpose of the event was achieved, comprehensively. "The main reason for doing this event is recruiting." In July 2019: ~90 employees, $158 million raised. By the $650 million Series E of June 2025: 400+ employees, total funding over $1.29 billion, a $9 billion valuation, and an investor list including ARK, Sequoia, Founders Fund, Thrive, Lightspeed, QIA and G42. Graded against the criterion the speaker himself set at the podium, the presentation is one of the more successful recruiting events in the history of the industry.

The medical device works, in humans, and that was not obvious in 2019. A fully implanted, wireless, battery-powered, 1,024-electrode recorder, inserted by a robot, worn under the scalp with no percutaneous connector — the pedestal sticking out of the skull was the standard of care in academic BCI for two decades and is a chronic infection risk. Neuralink removed it. As of 28 January 2026 the company reported 21 participants implanted across sites in the United States, Canada, the United Kingdom and the United Arab Emirates, up from 12 in September 2025, with no serious device-related adverse events — the company's own figure, unaudited and not peer-reviewed, and carried here as the company's own.

The performance is real. Over 9 bits/s on Webgrid, roughly double the prior record. Cursor control, gaming, chess, web browsing, CAD, and — via the field more broadly — decoded speech with voice cloning, and at least one participant who kept earning an income. For people with cervical spinal-cord injury or ALS this is not a demo, it is function.

The engineering recovered from the first serious failure without another operation. Within weeks of the January 2024 implant, roughly 85% of Arbaugh's threads retracted — the brain moved more than expected — and the usable channel count collapsed. Neuralink recovered performance in software, by re-deriving the decode from the surviving electrodes, and Arbaugh's numbers later exceeded his pre-retraction baseline. Surgical technique was changed for subsequent participants and no later participant has had retraction at that scale. That is a good engineering organisation responding to a bad week in public, and it should be said as plainly as the retraction is.

And the sceptical independent forecast undershot. MIT Technology Review, 16 January 2025, assessing Musk's stated goal of 20–30 implants in 2025, judged it optimistic and projected from trial registrations at least two more by the end of 2025, with eight more by the end of 2026 — call it five total by end-2025 and thirteen by end-2026. The actual figure was 21 by January 2026. Musk's own number for 2025 was also missed, and by more in absolute terms. But the careful independent forecast was wrong in the same direction as the sceptics' priors, and this file records that because a canon that only catches over-optimism is measuring its own thumb.

3.5 The failure mode this entry adds

spacex-booster-reuse §4.8 sets out the canon's current taxonomy of forecast failure for this forecaster, and disagrees with its sibling in doing so: three modes, not one — schedule slip, which resolves; magnitude inflation, which does not; and pass-through claims, which depend on a competitor existing. tesla-volume-production §3.2 offers the sorting test:

> Ask whether the forecast is a scale-up of something already running, or the > first instance of something that has never run. The first is usually late. The > second is usually not a schedule claim at all.

Claim A is the second type and yet it resolved, which the test does not predict. The reason is a mode neither sibling file has, and it is this entry's contribution:

> The date was set by an institution the forecaster does not control, had not > approached, and whose standard is evidentiary rather than technical.

Not all of the 37 months are attributable to that, and I will not claim they are. Some of the slip is ordinary engineering: the architecture presented in 2019 was abandoned within thirteen months and the replacement had to be built, and the primate and porcine work had to be done. What can be said precisely is that the slip's longest single stretch is regulatory and is dateable. The threads worked in rats in 2019, in a pig by August 2020 and in a macaque by April 2021. From that point the company had a device and a demonstration and no application — the first filing came in early 2022, was rejected, and approval did not follow until 25 May 2023. Between the macaque and the approval lie twenty-five months in which the binding question was not whether the hardware worked but what a regulator must be shown before a healthy-brained adult can be opened: the lithium battery, migration of the threads, explantation without tissue damage. No amount of engineering velocity compresses those, because the rate limit is a review process and a safety record, not a fab.

This mode has three properties that make it worth carrying separately:

• It resolves — unlike magnitude inflation, which does not. The thing arrives; the date does not.
• Its multiple is bounded and estimable in advance. A first-in-human IDE for a novel implanted neural device is a multi-year process from a standing start, and it was a multi-year process in 2019 for anyone who asked. That is why the 3.1× lands inside the 2×–5× band rather than running away like the robotaxi claim: the constraint had a natural length.
• It is invisible to the forecaster's usual competence. Musk's hits (tesla-volume-production, spacex-booster-reuse) are cases where the binding constraint was manufacturing or physics — domains where money, iteration speed and vertical integration genuinely compress schedules. Neither compresses an FDA review.

The test a reading should apply, and it is one question: is the thing that has to happen next an engineering problem, or a permission? If it is a permission, discount the date by the institution's historical clock and not by the forecaster's track record — and check whether the forecaster has yet filed anything at all. mycin-1976 is the same finding from the other end: a system that performed and was never deployed, for institutional reasons its builders documented in 1984. hinton-radiologists-2016 is the same finding at labour- market scale. This is the third instance and the first where the institution's objections are on the public record in enough detail to attribute the slip rather than infer it.

3.6 The self-grade, the independent grade, and the manoeuvre the table did not have

jobs/canon.md requires both grades and requires the entry to say they differ. Here the difficulty is unusual: there is no self-grade to record.

I could not find any instance of Musk grading the December 2020 miss — not conceding it, not defending it, not reclassifying it. I looked and I am recording the absence rather than manufacturing a quotation. What the record contains instead is the date being moved, serially: end of 2020, then 2021, then 2022, then — in late 2022 — "in about six months," then 2023, then delivery in January 2024. Reuters, reviewing the sequence in March 2023, counted the forecast being made at least four times since 2019 before the company had applied to the regulator even once. That is amodei-2024-loving-grace's manoeuvre — the date advances by roughly the elapsed time, so it never comes due — applied five or six times in a row.

And then something the canon's table does not contain happened: it arrived, and the arrival absorbed the schedule.

The coverage of 29 January 2024 is a near-uniform success story. The December 2020 promise is largely absent from it. Where it appears, it is frequently discounted on the forecaster's behalf. New Atlas, on the day, Loz Blain:

> Adjusted for Musk's famously optimistic approach to timelines, the company has > delivered right about on time.

That sentence is written approvingly, by a competent outlet, about a four-and-a-half-year delivery against a seventeen-month promise. It is the clearest single artefact this canon has of how a base rate is actually destroyed: not by the forecaster regrading, disowning, advancing or relabelling — the four moves musk-robotaxi-2019 catalogued — but by the audience applying the correction factor itself and then treating the corrected forecast as the forecast that was made.

Call it the fifth entry in that four-item table — musk-robotaxi-2019 also identifies a move Musk uses in person and keeps outside the table, concede the date and claim the direction, so anyone counting every documented manoeuvre gets six. What matters is not the number but what this one does that none of the others do:

• Regrade generously (kurzweil-2005) — the claim survives, loosely scored.
• Disown (minsky-1970) — the data point is deleted.
• Advance the date (amodei-2024-loving-grace) — never due.
• Relabel (musk-robotaxi-2019) — moved out of the gradeable class, with a court's help.
• Absorb (this entry) — the claim is superseded by delivery, and the audience retro-fits the multiplier. Net effect: the forecast enters the record as a hit, the miss leaves no trace, and — uniquely among the five — the forecaster does not have to do anything at all. The other four require a move by the principal or their counsel. This one is performed by third parties, for free, and is therefore the one a canon most needs to be watching for, because there is no statement to catch.

The independent grade, then, has to be assembled rather than quoted, and it differs from the retelling on three points.

The horizon was 3.1×, not "about right." Seventeen and a half months promised, fifty-four and a half delivered. The "Musk time" correction factor is not a property of the forecast; it is a property of the audience, and applying it silently converts every miss in the 2×–5× band into a hit. If the correction is legitimate it should be applied at the time the forecast is made, publicly, by the person applying it — which is exactly what MIT Technology Review did on 17 July 2019 when it called the timeline "highly ambitious and pretty unlikely," and it was right.

The thesis and the schedule are being scored as one claim, in the company's favour. The delivery graded above is Claim A. Claim B is the one the id is named for and the one the company is valued on, and it is at zero. Reporting "Neuralink delivered" without saying which of the two delivered is the same substitution musk-robotaxi-2019 §4.4 identified — "the self-grade survives only by substituting a vaguer claim for the specific one, and the specific one is the only kind a base rate can use."

The strongest evidence against Claim B is Neuralink's own best result. Not a critic's number, not a rival's: 9-plus bits per second, published by the company as its flagship achievement, sitting just under the ceiling that the peer-reviewed refutation predicts and the thesis requires to be breached. That is the single most useful line this file has, and it is available to anyone who puts the two documents side by side.

The bottom line for base-rate purposes. The same forecaster, in the same year, made two dated claims about hardware nobody had built. One was bounded by a regulator and came in at 3.1× — late, delivered, inside the rule. One was bounded by physics and vehicle economics and is past 5× with its mechanism withdrawn. The difference between them was not the forecaster's confidence, which was identical, and not the ambition, which was comparable. It was the shape of the constraint. Sorting a forecast by its constraint before discounting it is worth more than any hit rate, and it is what four graded claims now let this canon do.

Commonly misused as

Not required for prediction, but this subject attracts misuse from both directions in roughly equal volume, and the entry is worth less without the section.

"Neuralink was the first to let a paralysed person control a computer by thought."

No, by twenty years. Matthew Nagle, paralysed from the neck down, received a 96-electrode Utah array in June 2004 and controlled a cursor, opened email, drew, played a simple game, operated a television and used a robotic hand; published in Nature on 13 July 2006 by Hochberg and colleagues on the BrainGate trial, which is still running: MIT Technology Review describes the BrainGate/UC Davis work in June 2026 as part of a two-decade research effort. Neuralink's advance over that is real and is an engineering advance — channel count, robot insertion, full implantation with no percutaneous connector, wireless power and telemetry, industrialisation. It is not the first demonstration of the capability, and a reading that cites the January 2024 implant as a first-of-kind will be wrong in a way that a neuroscientist will notice immediately.

"The symbiosis thesis is what the AI-safety field thinks."

It is not, and nothing in the last seven years of alignment research is built on it. The canon's own record of what the field actually did is concrete-problems-2016, rlhf-christiano-2017, superintelligence-2014 and bletchley-declaration-2023: specification, oversight, interpretability, evaluation, governance. Merging is not a research programme; it has no literature, no benchmark and no funded lab. It is one man's answer, offered alongside three other incompatible answers from the same man — co-founding OpenAI in December 2015 (democratise it), founding Neuralink in 2016 (merge with it), signing the Future of Life pause letter in March 2023 (stop it), and announcing xAI on 12 July 2023 (build it correctly myself). Those are four different theories of the same stated risk, each executed as a company or a signature, within eight years. Recording that is a fact about the forecaster's consistency and is not an argument about which answer is right — and a reading that uses it as one has overreached.

"Neuralink failed."

False, and lazily so. Twenty-one people have received an implant; the first of them has had his for over two years and is still using it; the reported performance is at able-bodied cursor rates; the surgical failure of 2024 was diagnosed and engineered around without reopening the skull; and the field the company competes in has gone from 67 trial volunteers to about 150 over the same period. What missed was a date, by 37 months, and what has not been attempted is a thesis. Neither is "the company failed," and this canon should not let a grading exercise slide into a verdict on an organisation.

"Musk is always wrong about timelines, so discount everything."

A base rate of four is not a law, and two of the four landed. The canon holds tesla-volume-production and spacex-booster-reuse precisely so this move cannot be made from inside it. §3.5 gives the actual discipline: sort by constraint — engineering, physics, market, or permission — and discount by the constraint's clock, not by the forecaster's reputation. Citing a track record as a blanket multiplier is the same overconfidence as the forecast, wearing the opposite jersey.

"More electrodes will get us to symbiosis; it's just a scaling problem."

This is the misuse the entry most needs to block, because it sounds like the scaling-laws-2020 argument and is not. Zheng and Meister's finding is that the ~10 bits/s ceiling appears to be set by central cognition, not by the output channel — so widening the channel does not move it. Neuralink's own 9-plus bits/s is consistent with that and inconsistent with the scaling story. The honest form of the pro-bandwidth case is Matt Angle's and it is much narrower: more channels demonstrably help restore a channel that is broken — speech decoding is the live example — and say nothing about augmenting an intact one. Anyone invoking bandwidth should be made to say which of the two they mean.

"Neuralink's animal-welfare record proves the whole thing is a fraud."

The record is real and belongs in Sources: a USDA investigation reported in December 2022, Physicians Committee for Responsible Medicine complaints about the UC Davis primate work, an SEC matter reported in December 2024. None of it is a finding about whether the 2019 claims were true, and this file's grade does not rest on any of it. Reaching for it is the cheapest available way to make a grading exercise feel conclusive, and minsky-1970 set this canon's standard when it declined to read a procedural outcome as a finding about a person.

Sources

Primary where reachable; every fetch failure named at the end, per the rule that a failed search is noted and the work continues.

• Musk, E. and Neuralink, "An integrated brain-machine interface platform with thousands of channels." bioRxiv 10.1101/703801 (July 2019); Journal of Medical Internet Research 2019;21(10):e16194, doi:10.2196/16194, PubMed 31642810. Not read in full — bioRxiv returned HTTP 429 on repeated attempts and jmir.org returned an empty body to this machine. The technical figures used here (up to 3,072 electrodes per array across 96 threads; robot insertion of 6 threads / 192 electrodes per minute; spiking yield up to 85.5%; freely moving rats) are taken from the abstract as quoted in the accompanying JMIR commentary, below, and cross-checked against the bioRxiv listing. This is the entry's most important unread primary and deserves a direct read on any later pass.
• Pisarchik, A. N., Maksimenko, V. A., and Hramov, A. E., "From Novel Technology to Novel Applications: Comment on 'An Integrated Brain-Machine Interface Platform With Thousands of Channels' by Elon Musk and Neuralink." JMIR, 31 October 2019; read directly via PubMed Central (PMC6914250). Source for the quoted technical figures, for the observation that stimulation results are asserted but not presented, and for biocompatibility as "one of the main obstacles for clinical applications of implanted devices."
• The 16 July 2019 presentation itself. Not seen in the primary. The recording exists; neuralink.com is JavaScript-rendered and returned only a page title to this machine, and no transcript was reachable. Every quotation from the event in this file is assembled from contemporaneous reports by outlets that had reporters present, and only where two or more independent outlets agree:
• "Even under a benign AI, we will be left behind. With a high-bandwidth brain-machine interface, we will have the option to go along for the ride" and "A monkey has been able to control the computer with his brain. Just FYI" — Antonio Regalado, MIT Technology Review, 17 July 2019, read directly. Also the source for MacDougall's "only intended for patients with serious unmet medical diseases," for the Q2-2020 trial-start intention pending FDA approval, and for that article's same-day judgement that the timeline was "highly ambitious and pretty unlikely."
• "some sort of symbiosis with artificial intelligence", "With a high-bandwidth brain interface, I think we can have the option of merging with AI", "will take a long time", and the end-of-2020 prototype target — NBC News, 17 July 2019, read via the Euronews syndication of the same wire text.
• "We hope to have this in a human patient by the end of next year" — CNBC, 17 July 2019. The CNBC page returned HTTP 403 to this machine; the wording is taken from search-result summaries of it and is corroborated by Fox News and Axios reports of the same statement from the same day. Recorded as the weakest-sourced direct quotation in this file.
• "The main reason for doing this event is recruiting" — reported consistently across contemporaneous coverage, none of which I reached as a fetched page; the phrasing varies slightly between outlets ("for doing this event" / "for this presentation") and the version given here is the more commonly attested. Since this quotation carries a whole grading criterion in §3.4, its second-hand status matters more than most.
• Zheng, J. and Meister, M., "The unbearable slowness of being: Why do we live at 10 bits/s?" Neuron, December 2024 (arXiv:2408.10234). Division of Biology and Biological Engineering, Caltech. Section 5.3, "The Musk illusion," read directly in the arXiv HTML, including the full quotation given above and the paper's reference [47], which identifies the symbiosis quotation's source as a Joe Rogan interview clip, November 2018 — not the 2019 event, which is the basis for the first correction at the head of this file. The Caltech PDF was fetched but returned as unparsed binary.
• Antonio Regalado, "Elon Musk wants more bandwidth between people and machines. Do we need it?" MIT Technology Review, 29 September 2023; read directly. Source for Lee Miller's "folly" quotation, for Matt Angle's counter-case on speech restoration, for Vikash Gilja, and for the ~40 bits/s figure for human information sharing across languages.
• Erik Hoel, "Neuralink's grand vision isn't scientifically supported," The Intrinsic Perspective, 10 September 2024; read directly. Source for "innately bottlenecked by biology, rather than merely inconvenienced."
• O. Rose Broderick, "Neuralink's big vision collides with reality of brain implants," STAT, 5 January 2026. Paywalled at the point where the named sourcing begins; the thesis, the characterisation of the rhetoric as diverging sharply from the clinical work, the claim by competitors/investors/ former regulators that it hampers FDA approval and insurance coverage, and the hire out of the FDA office that regulates the company are all from the accessible opening and from the article's own summary. The named quotations behind those claims were not read, and the finding is carried here at that strength and no higher. This is the file's most consequential 2026 source and the one most in need of a direct read.
• Jessica Hamzelou, "Brain-computer interface trials are taking off," MIT Technology Review, 19 June 2026; read directly. Source for ~150 people implanted field-wide against 67 volunteers in trials through end-2023, for the 21 Neuralink participants over two years, for the roster of companies in trials, and for the demonstrated capabilities including speech decoding with voice cloning.
• Antonio Regalado, "What to expect from Neuralink in 2025," MIT Technology Review, 16 January 2025; read directly. Source for the three-patient count in January 2025, for Musk's 20–30 goal, for the independent projection that undershot (§3.4), for Arbaugh's 9-plus bits/s against ~10 for able-bodied users, and for the 45-minute recalibration sessions.
• Reuters, 28 January 2026 — 21 participants enrolled worldwide, up from 12 in September 2025; Neuralink's own claim of zero serious device-related adverse events; the company statement that "a primary aim of our expanding clinical trials is to better understand these variations." Read via Yahoo's syndication; reuters.com is not accessible to this machine's user agent.
• Reuters, 2 March 2023, on the FDA's rejection of the early-2022 application — the lithium battery, thread migration and explantation concerns, and "dozens of issues." Not read directly. Taken from search-result summaries of the CNBC, Forbes, TechCrunch, MedTech Dive and Fierce Biotech reports of it, which agree on all four points.
• Lex Fridman Podcast #438, "Elon Musk: Neuralink and the Future of Humanity," August 2024; transcript read directly at lexfridman.com. Source for the 2024 restatements: "We could better align collective human will with AI if the output rate especially was dramatically increased"; "There's potential to increase the output rate by, I don't know, three, maybe six, maybe more orders of magnitude"; "We're hoping to do 10 by the end of this year, total of 10"; "Maybe if five years from now, we might be at a megabit"; and, on non-disabled implantation, "Probably less than 10 years." The episode number is on the page; the exact date is not, and August 2024 is from the publication record.
• Loz Blain, "Neuralink fits its first human patient with a brain interface," New Atlas, 29 January 2024; read directly. Source for the implant details (R1 robot, coin-sized N1, 64 threads, 1,024 channels) and for the "delivered right about on time" sentence quoted in §3.6.
• Wikipedia, "Neuralink," accessed 20 August 2026. Used for the dated skeleton only — founding date and founders, the July 2020 breakthrough designation, the Gertrude and Pager demonstrations, the 25 May 2023 approval, the PRIME study opening at Barrow, the identification of Noland Arbaugh on 20 March 2024, the 85% thread retraction reported in May 2024, the second and third participants, the UK and Canadian sites, the September 2024 Blindsight designation, and the animal-welfare and SEC matters. Every load-bearing figure in this file is sourced elsewhere as well; nothing in §3.2 or §3.3 rests on Wikipedia alone.
• Hochberg, L. R. et al., "Neuronal ensemble control of prosthetic devices by a human with tetraplegia," Nature 442, 13 July 2006, and the companion paper "A high-performance brain–computer interface" in the same issue. Neither read. The Nature listings were located but not fetched, and the details used here come from search-result summaries of institutional write-ups (Brown, UChicago Medicine, ScienceDaily, EurekAlert). Source for Matthew Nagle's June 2004 implant, the 96-electrode array, and the 2004–2005 session record used in "Commonly misused as".
• Neuralink's own Series E announcement, June 2025. The company's page was fetched and returned only a page title; the figures — $650 million, ~$9 billion valuation, the investor list (ARK, Sequoia, Founders Fund, DFJ Growth, G42, Human Capital, Lightspeed, QIA, Thrive, Valor, Vy), 400+ employees, cumulative funding above $1.29 billion — are from search-result summaries of the Fierce Biotech and SiliconANGLE reports, which agree.
• Musk, post on X, 3 August 2014 — "Worth reading Superintelligence by Bostrom..." Widely reported the following day (NBC News, CNBC, Christian Science Monitor) and already recorded in superintelligence-2014, from which it is taken here rather than re-derived.
• Within this repository, cited as record and never as evidence: canon/musk-robotaxi-2019.md (WIRED's 2×–5× rule, the four-manoeuvres table, the §4 pairing request, the rule-7 inversion), canon/tesla-volume-production.md (the scale-up-versus-first-instance test), canon/spacex-booster-reuse.md (the three-failure-modes taxonomy), canon/culture-banks-1987.md (the neural lace attribution and its warning), canon/superintelligence-2014.md (the Musk endorsement and the causal-inference caveat), canon/kurzweil-2005.md, canon/mycin-1976.md, canon/hinton-radiologists-2016.md, canon/minsky-1970.md, canon/amodei-2024-loving-grace.md, canon/concrete-problems-2016.md, canon/rlhf-christiano-2017.md, canon/bletchley-declaration-2023.md, canon/scaling-laws-2020.md, and LENSES.md, from which only two lens names are borrowed, in "Why a reading would cite it", and which this file reads and does not write. Nothing is taken from this project's own readings, digests or ledger.

Fetch failures, recorded

neuralink.com and all its subpages (JavaScript-rendered; page titles only); reuters.com, nytimes.com, apnews.com, wsj.com, ft.com, theverge.com and arstechnica.com (refused to this machine's user agent, so Reuters had to be read through syndication throughout); cnbc.com (HTTP 403); axios.com (HTTP 403); fiercebiotech.com (HTTP 403); technology.org (HTTP 403); the NIHR/UCLH BRC page on the UK trial (HTTP 403); feeds.bbci.co.uk (HTTP 404); bioRxiv full text (HTTP 429 on both attempts); jmir.org (empty body on both attempts); beckersspine.com (DNS failure); statnews.com (paywalled below the opening); the Caltech PDF of the Zheng and Meister paper (returned as unparsed binary — the arXiv HTML was used instead, so nothing was lost there).

A note on what that list does to the file. The two documents this entry most needed in the primary — the recording of the presentation and the STAT investigation — are the two it did not get, and they sit at opposite ends of the argument: the first is what Claim A's date rests on, the second is what the 2026 finding in §3.3 rests on. Neither was replaced by a weaker source without the substitution being named at the point of use.

Where this file could be wrong

The event quotations are second-hand. Every one is corroborated across independent outlets and none is used to carry an argument alone, but the recording was not heard and a later pass should hear it. The most consequential of them, "We hope to have this in a human patient by the end of next year," is the one whose primary source I could not reach directly, and Claim A's due date depends on it. If that wording is wrong, §3.2's multiple is wrong.

Which of the founders spoke on the night is not established. Hodak and MacDougall are confirmed by coverage; the header does not name anyone else.

The 3.1× is arithmetic on two dates and is therefore only as good as they are. 16 July 2019 to 31 December 2020 is 17.5 months; 16 July 2019 to 28 January 2024 is 54.4 months. Anyone who thinks the horizon should be measured from the Q2-2020 trial-start intention instead of the end-2020 patient promise gets 4.7× rather than 3.1×, which is still inside the band. Anyone who dates the claim to the November 2018 Rogan formulation gets a longer elapsed time and a larger multiple. The band conclusion survives all three choices; the specific number does not, and it is the number, not the conclusion, that is quoted here.

The strongest objection to §3.3, stated so it can be argued with. Claim B was never given a due date at the event, and Musk said explicitly that it would take a long time. Grading it "not attempted" after seven years could be called premature — a fifty-year thesis is not refuted at year seven. The response, and it is not a full answer: the entry does not grade the thesis false, it grades it unattempted, with its premise contested, and it records the 2029 and 2034 dates Musk himself attached to it in 2024 as not yet due. If a healthy person is implanted before 2029, this section will need rewriting and the canon's append-only discipline does not apply to canon/ — a later pass should correct it in place and say that it did.

The most likely direction of error in the whole file is over-severity toward the thesis and over-generosity toward the schedule, because the thesis has abundant hostile literature and the schedule has abundant friendly literature, and both were easier to find than their opposites. §3.4 exists as the counterweight and the reader should judge whether it is enough.

On the boundary

This file grades dated claims and nothing else. It deposits nothing in the evidence ledger, writes no digest, touches no other entry's file, places no needle, no score and no landmark position, and does not run build.py. It writes exactly one file: canon/neuralink-symbiosis-2019.md.

Nothing in it is evidence. The 2026 facts in §3.3 and §3.4 — the participant counts, the STAT reporting, the field-wide implant total — appear as record, so that a reading citing this entry knows where the claims stood when the file was written, and they are not deposits and do not move the needle. Where a number is the company's own, it is labelled as the company's own every time it appears. Where a source was not read, the sources section says so.