Tesla Autonomy Investor Day

This entry is named in README rule 7, and it is named as the reason a rule exists. The clause reads: "no vendor is ever flattered by a project that grades that vendor's predictions (musk-robotaxi-2019 is in the canon; xAI is a candidate to compose the tweet — those two facts must never meet)." Five other files in canon/ point here before it existed. turing-1950 calls it "the canon's standing case" for the prediction kind and, separately, "the ceiling" against which a vendor's timeline is graded. darpa-grand-challenge-2005 declines to grade Musk's robotaxi claims at all on the ground that "that is that entry's whole job." terminator-1984 names it as the canon's record of a failed date being quietly rolled forward. llama-weights-2023 and superintelligence-2014 both cite it in their own rule-7 disclosures.

So the file arrives pre-loaded, and it arrives with a hazard the other prediction entries do not have.

The corruption risk here runs backwards. Rule 7 is written against flattery — a grade going quietly soft because the vendor is adjacent to the project's plumbing. But this subject is not one a model built by Anthropic is tempted to flatter. The convenient grade here is the hostile one, and it is available in enormous quantity: the secondary literature on this prediction is mostly written by people who reached their conclusion before they reached the evidence, and a file assembled from it would be indistinguishable from a good one at a glance. superintelligence-2014 noticed the same inversion about its own subject and wrote a longer disclosure because of it. This file does the same, in section 9, and the discipline it imposes on sections 3 and 4 is the main thing that separates this entry from the twenty trackers it cites.

The finding, stated up front so the rest can be checked against it: the prediction is one of the largest misses this canon holds, the direction it pointed in was right, somebody else built the thing, and they built it with the component he said was fatal. All four of those are true at once and an entry carrying fewer than four is worth less.

1. What it is

The event

On 22 April 2019, at Tesla's headquarters in Palo Alto, Tesla held an invitation-only presentation for institutional investors and analysts, billed as Autonomy Investor Day and universally called Autonomy Day. It ran about three hours and was livestreamed. The running order was, roughly: Pete Bannon on the custom FSD Computer silicon Tesla had designed to replace NVIDIA's part; Andrej Karpathy, then Tesla's director of AI, on the vision neural network and how the fleet trained it; Stuart Bowers on Autopilot deployment; and Musk throughout, in question-and-answer and in interjection.

It was a financing event as much as a technical one. Tesla raised roughly $2.7bn in equity and convertible notes in the weeks that followed. That context is not a slur — investor days exist to be persuasive — but it is load-bearing for section 5, because it is why a court eventually had to decide what kind of speech act the claims were.

The claim, verbatim

Four sentences carry the prediction. They should be read together, because they are usually quoted apart and the mechanism lives in the ones that get dropped.

> "I feel very confident predicting autonomous robotaxis for Tesla next year."

> "By the middle of next year, we'll have over a million Tesla cars on the > road with full self-driving hardware, feature complete, at a reliability level > that we would consider that no one needs to pay attention."

> "From our standpoint, if you fast forward a year, maybe a year and three > months, but next year for sure, we'll have over a million robotaxis on the > road. The fleet wakes up with an over the air update; that's all it takes."

> "Not in all jurisdictions, because we won't have regulatory approval > everywhere, but I am confident we will have at least regulatory approval > somewhere, literally next year."

And, in the same session, the hedge he gave himself before anyone else could:

> "Sometimes I am not on time, but I get it done."

That last line matters and is almost never quoted by the trackers. It is a pre-emptive concession on schedule and an assertion of eventual delivery, made on the day, unprompted, in front of the investors. Section 4 is about what happens to a base rate when a forecaster builds the excuse into the forecast.

The mechanism, which is the part that matters

The claim that got repeated for seven years is "a million robotaxis by 2020." The claim that is actually gradeable, and that the canon should carry, is narrower and much stronger:

A million cars that Tesla had already sold, already carrying the necessary hardware, would become driverless commercial vehicles by an over-the-air software update.

That is not a manufacturing forecast. Tesla did not promise to build a million robotaxis; it promised that the million it had built were robotaxis in waiting. Everything else in the presentation follows from it — the economics, the appreciating-asset argument, the fleet-training story, the case for buying the Full Self-Driving option at the price Tesla was then charging for it. A prediction whose mechanism is "the hardware is already in your driveway" is far more falsifiable than one whose mechanism is "we will build a factory," and it is falsifiable in a specific way that section 3.4 records.

The economics

Musk put numbers on it. Each robotaxi-enabled Tesla was projected to generate on the order of $30,000 a year in gross income for its owner. Tesla would take 25–30% of ride revenue, running an Uber/Airbnb-shaped network into which owners would add and subtract their own cars; where too few owners participated, Tesla would run dedicated fleets. Operating cost was put at $0.18 per mile. And the conclusion:

> "If you buy a Tesla today, I believe you are buying an appreciating > asset — not a depreciating asset."

with the corollary that buying any other car would soon look like buying a horse.

On the day, one of the presenters said the numbers were not real. Asked about the financial projections on the slide, the answer given was "we just randomly threw some numbers on there." This is in the contemporaneous reporting and it is the single most useful sentence to have available when a 2026 principal shows a slide with a per-unit economics table on it.

The lidar sentence

> "LiDAR is a fool's errand. And anyone relying on LiDAR is doomed. Doomed. > Expensive sensors that are unnecessary."

Sometimes rendered with the coda "It's like having expensive appendices. You'll see." Tesla's stack was, and remains, camera-first: eight cameras, no lidar, no HD prior maps of the kind Waymo builds. This is a separate claim from the timeline claim and it grades separately. Section 3.5 does that, and the answer is more interesting than either camp's version.

The credible half

An entry that only recorded the boasts would be misreporting the event, and the project would be worse at its job for it.

Karpathy's presentation was substantive and much of it aged well. The core argument — that a fleet of hundreds of thousands of customer cars can be asked to return targeted examples of specific rare situations, and that "shadow mode" lets a candidate stack be scored against human decisions at a scale no test fleet can reach — was a real structural advantage, described in technical terms, and it is recognisably the data-engine pattern the whole industry later adopted.

The best contemporaneous witness is a hostile-by-default one. Brad Templeton, who had worked on Google's self-driving project and had spent years writing sceptically about Tesla's approach, published the same day that the presentation "significantly improved my impression of Tesla's methods and chances," and followed up a week later with a piece arguing that shadow testing was a genuine advantage on the hardest problem in the field. He nonetheless named the unresolved question exactly: whether cameras alone could reach "the final 9s of reliability needed to remove the steering wheel from the car." He also checked the money — putting the industry estimate for per-mile operating cost at about 37 cents against Tesla's 18 — and doubted that enough owners would keep a car dispatchable.

Hold that. A named expert, on the day, in public, said the technical argument was better than he expected and the timeline and the economics were not supported. Section 7 uses it against the claim that nobody could have known.

2. On the kind, and on the descent

The kind

prediction is correct and proposals.md filed it correctly. The kind requires four things and all four are cleanly available: the claim in the predictor's own words (§1), the date made (22 April 2019), the date due (mid-2020, with a stated outer bound of "a year and three months," i.e. July 2020), and what happened (§3). This is the least contested kind assignment in the canon's prediction set — minsky-1970 had to argue past a disputed attribution and turing-1950 declined the kind entirely. Here the speaker is on video, the date is fixed, and the claim is arithmetic.

moment was considered, on the grounds that Autonomy Day was an event with consequences — a $2.7bn raise, a strategic commitment to camera-only that the company still runs on. It is rejected because those consequences are all downstream of the claims, and a moment entry would have to grade them anyway while pretending the file was about a room in Palo Alto.

descends_from: [darpa-grand-challenge-2005, alexnet-2012]

darpa-grand-challenge-2005 is the strong edge and it is confirmed from the other end. That entry says, in its own words: "musk-robotaxi-2019 is proposed and unwritten. It is the natural descendant of this entry on the forecasting side, and section 3 deliberately does not grade Musk's robotaxi claims, because that is that entry's whole job." The descent is more than rhetorical. The 2004–2007 Challenges produced the personnel, the sensor industry and the institutional belief that autonomous driving was a solved-in- principle problem awaiting engineering; every serious 2019 autonomy programme, Tesla's included, was staffed and financed inside the expectation those races created. The DARPA entry's own most durable finding is the interval it measures — fifteen years and a day from Stanley finishing the desert course (8 October 2005) to Waymo opening a fully driverless public service in Phoenix (8 October 2020) — and this entry is what a prediction looks like when it is made in year fourteen of that interval by someone who thinks it is year fourteen of a two-year problem.

alexnet-2012 is the technical edge and the presentation itself supplies the grounds. Autonomy Day's load-bearing claim was not about cars. It was that a sufficiently large convolutional vision network, trained on data returned from a fleet, could substitute for the geometry that lidar and prior maps provide. That is the AlexNet result applied to driving, and it was presented as such, by a researcher whose career runs directly through that lineage. The edge is documentary rather than by citation — Karpathy did not stand up and cite Krizhevsky — but the method being sold in that room is the method that entry records, and marking it is more honest than leaving the header with one edge and pretending the technical claim came from nowhere.

Convergent, not descended: bitter-lesson-2019. Sutton's essay published 13 March 2019, five and a half weeks before Autonomy Day, and its thesis — that general methods leveraging computation beat human-engineered structure, and that building in what we think we know is the recurring mistake — is very nearly the anti-lidar argument restated. I found no evidence that anyone at Tesla read it or that it influenced the presentation, and I am not going to manufacture the link. The convergence is worth noting for a different reason: Rodney Brooks's rebuttal, "A Better Lesson," published 19 March 2019 — one month before Autonomy Day — argues that the total cost of a solution is what matters and that these systems have all required substantial human ingenuity, and its worked example is the energy budget of self-driving computation. bitter-lesson-2019 already carries that exchange. The strongest available intellectual criticism of the sensor bet was in print, from a named roboticist, four weeks before the bet was announced.

moravecs-paradox-1988 is cited in prose and is not an ancestor. It explains the error — driving is perception and physical judgement in an unstructured world, which is the expensive half of Moravec's split, and the canon has already graded a 1970 forecast on exactly this fault line. But Musk's claim did not come out of Moravec, and descent is not the same as diagnosis.

Ids I wanted and did not invent. neuralink-symbiosis-2019 sits in proposals.md unwritten, from the owner's proposal of 2026-08-15, and its stated purpose is to be "the second reading on the same forecaster, which is what makes a track record a base rate instead of an anecdote." That is exactly right and this file cannot do it alone — one predictor graded once is an anecdote no matter how carefully it is graded. When neuralink-symbiosis-2019 exists, this file's section 4 should be read beside it and probably shortened. simon-1965 and moore-1965 are likewise proposed and unwritten.

3. What it got right, and what it got wrong

3.1 The four parts

The claim. Over one million Tesla robotaxis on public roads, operating commercially with no human attention required, created by an over-the-air update to cars already sold.

Date made. 22 April 2019.

Date due. Mid-2020. Musk supplied both the target ("by the middle of next year") and his own outer bound ("a year and three months") — July 2020 at the latest, by his own arithmetic. The regulatory sub-claim was due on the same schedule: approval somewhere, "literally next year."

What happened. Sections 3.2 and 3.3.

3.2 At the due date

Zero. Not a reduced number — zero. In mid-2020 no Tesla was operating without a human driver anywhere in the world, no Tesla robotaxi network existed, and Tesla had applied for no driverless deployment permit in any jurisdiction.

The nearest thing to a milestone in that window ran the other way: FSD Beta shipped in October 2020, four months past the outer bound, to a few thousand hand-picked owners, as an SAE Level 2 driver-assistance system requiring continuous supervision — the same classification Autopilot already had. The regulatory sub-claim, which was the weakest form of the prediction and the one Musk had explicitly hedged down to, also failed: not "not everywhere," but nowhere.

And the company said so internally, to a regulator, while it was happening. At a meeting on 9 March 2021, California DMV staff asked Tesla about Musk's public statements on Level 5 autonomy. The DMV's memorandum of that meeting records: "Elon's tweet does not match engineering reality per CJ. Tesla is at Level 2 currently." CJ was CJ Moore, Tesla's director of Autopilot software. The same record states that "Tesla indicated that Elon is extrapolating on the rates of improvement when speaking about L5 capabilities" and that Tesla "couldn't say if the rate of improvement would make it to L5 by end of calendar year."

This is a document of unusual quality for grading a forecast and the canon should say why. It is not a critic's opinion and not a retrospective. It is the company's own engineering leadership, describing its own CEO's public timeline to a state regulator, in a setting where overstating would have been costly and understating pointless. The forecast was not merely wrong in hindsight; it was known inside the company to be extrapolation rather than an engineering estimate within two years of being made.

3.3 Seven years and four months later — August 2026

The honest grade of a long-dead prediction is not the score at its due date but the score now, because the interesting question is not "was he late" but "by how much, and is it still running." Figures below are as of this file's writing, 17 August 2026, with sources and their reliability in section 8.

Tesla. A robotaxi service exists. It launched in Austin on 22 June 2025 with roughly ten Model Ys, invite-only, geofenced, and a Tesla employee riding in the front passenger seat as safety monitor. Unsupervised rides — no monitor aboard — began in Austin on 22 January 2026, initially confined to a small corridor. Expansion followed: Dallas and Houston (18 April 2026), Miami (3 July), Orlando and Tampa (21 July), with a Bay Area operation running under safety drivers behind the wheel, because California requires permits Tesla has not obtained. Tesla told investors on its Q2 2026 call (5 August 2026) that the service was live in seven major US metros, that cumulative paid robotaxi mileage was about 2.5 million miles, that roughly 380,000 of those were driven with nobody aboard, and that unsupervised miles were growing above 10% week over week.

Fleet size is the number that grades the prediction, and it is small and disputed. Reported counts through 2026: ~29 in Austin and ~106 in the Bay Area in December 2025; ~72 Austin and ~168 Bay Area in a February 2026 tracker (~240 total, against a Musk earnings-call figure of "well over 500"); ~42 in Austin in June 2026 filings; 59 driverless vehicles company-wide in Bloomberg's June 2026 count; ~35 in Austin with ~60 Model Ys staged in a Phoenix lot in mid-2026. The spread is mostly a definitional one — driverless versus safety-driver-supervised, deployed versus staged, Tesla's count versus permit filings — and no source reaches four figures. Against one million.

Consumer cars. Tesla's FSD remains SAE Level 2 and is sold as Full Self-Driving (Supervised). NHTSA escalated its FSD investigation to an engineering analysis (EA26002) in March 2026, covering roughly 3.2 million vehicles, examining whether the camera-only system fails to detect degraded visibility — sun glare, fog, smoke — and hand back control in time.

The comparison that makes the number legible. Waymo reports 220.6 million rider-only miles through March 2026 and around half a million paid rides a week, across about ten US metros in March and about fifteen by August, with an active fleet on the order of 4,000 vehicles. Baidu's Apollo Go reported 3.2 million fully driverless rides in a single quarter and 22 million cumulative. Against Tesla's ~380,000 driverless miles, Waymo's driverless mileage is roughly 580 times larger. This project's own genesis ledger for robotics and embodiment, covering 17 April to 15 August 2026, holds thirty deposits and Tesla's robotaxi is not among them — the Tesla deposit in that window is Optimus production lines installed and not producing.

3.4 The mechanism failed, not only the schedule

This is the finding this entry contributes that the trackers do not, and it turns on a date that is almost too neat to be true.

On 22 April 2026 — seven years to the day after Autonomy Day — Musk said on Tesla's earnings call: "Hardware 3 simply does not have the capability to achieve unsupervised FSD."

Hardware 3 is the FSD Computer that Pete Bannon unveiled at Autonomy Day. It went into cars from 2019 to 2023. It is the chip that made "a million Tesla cars on the road with full self-driving hardware" true as a sentence about hardware, and it is the thing "the fleet wakes up with an over the air update" was a claim about. Tesla had told owners for years that no hardware upgrade would be needed. In January 2025 Musk called upgrades "painful and difficult"; in October 2025 Tesla's CFO said they might still make it work; in April 2026 the answer became that the cars need physical retrofit, in volume, via purpose-built "micro-factories," on no announced schedule.

So the prediction did not merely run late. Its stated mechanism was retracted by its author. The million cars were not a million robotaxis waiting for a download; they were a million cars needing a new computer. That distinction is what separates this entry from a generic late-delivery story, and it is the thing a reading should reach for:

When a capability is claimed to be already present in deployed hardware and one software update away, that claim has a checkable failure mode that a pure schedule claim does not — and the check is worth running years later, because the retraction, when it comes, comes quietly on an earnings call.

The rest of the mechanism went the same way. The owner-supplied fleet — the Airbnb-shaped network that made the $30,000-a-year and appreciating-asset arguments work — has not launched; every robotaxi in service in August 2026 is Tesla's own. Customer-car participation was promised for 2026 and, as of this writing, has not happened. The purpose-built Cybercab, unveiled October 2024 with no steering wheel or pedals, began production at Gigafactory Texas in Q2 2026 — that is, Tesla is now manufacturing dedicated robotaxis, which is precisely the capital-intensive path Autonomy Day existed to argue was unnecessary.

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

Three things, and skipping them would make this file propaganda.

Driverless commercial ride-hailing is real, at scale, in 2026. The direction of the 2019 claim — that within a few years members of the public would routinely pay to ride in cars with nobody driving — is correct, and in April 2019 it was not the consensus view. Half a million paid driverless rides a week is not a demo. darpa-grand-challenge-2005 grades the "fleets of autonomous cars" half of its own subject's forecast as "substantially right, and this canon should say so plainly," and the same sentence is owed here. He was right about the world and wrong about the company.

The bet on learned vision was not stupid, and the data-engine argument was ahead of its time. Fleet-scale data collection with targeted retrieval and shadow-mode evaluation is now standard practice. Templeton, who had every professional reason to be unimpressed, raised his estimate of Tesla's chances on the strength of it. A canon that grades this file's subject as a pure salesman has to explain why the field's most experienced sceptic came out of the room more convinced than he went in.

The pre-hedge was honest. "Sometimes I am not on time, but I get it done" was said on the day, unprompted, to the people being asked for money. Compare minsky-1970, where the forecaster denied the forecast twenty-seven years later, and kurzweil-2005, where the scoring was loosened after the fact. Whatever else is true, the schedule risk was disclosed at the point of sale. Section 4 is about how much credit that earns, and the answer is: some, and less than it looks.

And the lidar sentence grades to "not established, and losing." What can be said in August 2026: the camera-only stack has not produced driverless operation at anything like the scale of the lidar-fusion stacks, and the operator with two orders of magnitude more driverless miles uses lidar. Waymo's co-CEO Dmitri Dolgov put the technical form of the objection in August 2026 — "Humans of course can drive with just eyes," but "you find that weak sensing just leads to a safety curve that flattens out way too early" — and describes cameras, lidar and radar as complementary rather than redundant. What cannot be said is that camera-only has been refuted as a matter of physics or of principle; nothing establishes that, Tesla's driverless miles are non-zero and growing, and a reading that treats the sentence as settled is overclaiming. The honest grade: "anyone relying on lidar is doomed" was due continuously, has been wrong continuously for six years, and is the most clearly failed of all the claims made that day — because unlike the timeline it named a competitor's outcome, and the competitor won the interval.

3.6 The grade

A miss, by roughly four orders of magnitude on quantity, more than six years past its own outer bound, still open, with its mechanism retracted by its author.

Set against the canon's other predictions: minsky-1970 is a larger miss on horizon (forty-eight years past the outer bound of a claim about general intelligence, still contested). This is the larger miss on specificity, and it is the only one in the canon where the predictor later disclosed that the stated mechanism could not work. kurzweil-2005's headline 2029 claim is not yet due and currently looks good. The base rate this kind assembles is not "forecasters are wrong." It is that the more mechanically specific a forecast is, the more precisely it can be shown to have failed — which is an argument for making forecasts specific, not against it.

4. The self-grade and the independent grade

jobs/canon.md requires this section and requires that it say when the grades differ. They differ, and in an instructive shape.

4.1 The self-grade

Musk grades himself, publicly and repeatedly, and his grade is consistent: concede the date, claim the direction.

• On the day, 22 April 2019: "Sometimes I am not on time, but I get it done."
• July 2023: "People have sort of made fun of me and perhaps quite fairly have made fun of me, my predictions about achieving full self-driving have been optimistic in the past … I'm the boy who cried FSD, but I think we'll be better than human by the end of this year."
• Generally: "I say something, and then it usually happens. Maybe not on schedule, but it usually happens."

This is the most candid self-grade of any predictor in this canon, and it is still not a scorable one. It concedes lateness in general and grades no specific claim. Note the structure of the July 2023 statement: the concession is the setup for a fresh dated claim, which also failed. The apology and the next forecast are the same sentence. Six years of that produces a speaker who has admitted to being wrong many times and has never once recorded a specific prediction as failed.

4.2 The company's grade, which is better evidence than either

The California DMV memorandum of 9 March 2021 (§3.2) is the highest-quality grading document available on this prediction, and it is Tesla's own. "Elon's tweet does not match engineering reality per CJ." "Elon is extrapolating on the rates of improvement." When the predictor's own engineering leadership tells a regulator that the public timeline is extrapolation, a canon has something better than a self-grade and better than a critic: it has an internal grade, produced under conditions where accuracy was the safest option.

4.3 The independent grades

Wikipedia maintains a dedicated article — List of predictions for autonomous Tesla vehicles by Elon Musk — which is itself a finding. No other AI forecaster in this canon has one. The list runs from September 2013 ("90 percent of miles driven autonomously within three years") through October 2015, December 2015, January 2016, June 2016, October 2016 (the LA-to-NYC coast-to-coast drive by end of 2017), April 2017, April 2019, 2023, 2024, October 2024, April 2025 and onward. Nearly every dated entry is marked not met. The August 2022 claim of a wide FSD release by year-end is marked achieved (November 2022), correctly, and it is one of the very few.

A dedicated tracker counts 82 predictions, 78 missed, 3 with goalposts moved — about a 7% on-time rate. That number should be handled with tongs: the site is advocacy, it names itself a hall of shame, and its inclusion criteria are its own. It is cited here because it is the only source reached that attempts a denominator, and a base rate needs denominators. It is not relied on for anything in section 3.

WIRED's May 2025 analysis — nineteen years of pledges across FSD, Hyperloop, robotaxis and humanoids — reports the pattern as timelines that overshoot by 2x to 5x while the underlying ambition is often directionally real. That formulation is the most useful independent grade found, and it is worth noting that it does not fit this particular prediction: mid-2020 stretched past August 2026 is already past 5x on a "million robotaxis" that has not arrived at any multiple. Sub-claims of the same forecast — a service existing at all, driverless operation in some city — land inside the 2x–5x band. The rule describes the median Musk forecast and understates this one.

And the most damning independent number comes from a friendly source. Morgan Stanley — a house whose Tesla coverage has been structurally bullish for a decade — projected in a December 2025 note (Andrew Percoco, who took the name over from Adam Jonas) about 1,000 robotaxis by the end of 2026, 30,000 by 2030, and one million by 2035, conditional on removing safety monitors and clearing regulation in dozens of jurisdictions. A bullish bank's central case puts Musk's mid-2020 number in 2035: a fifteen-year slip, from someone with every incentive to be generous. That single line is worth more to a base rate than any critic's tally, and it is the number a reading should cite.

4.4 They differ, and here is the shape of the difference

The self-grade says late but delivered. The independent grades say late by a multiple that is still growing, on a mechanism that has been withdrawn. The distance between them is not a scoring dispute; it is a disagreement about what was promised. Musk grades the promise as "Tesla will do autonomy." The record grades the promise he actually made, which was "the cars you already own become a million-vehicle commercial fleet next year by download." 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.

5. What this entry adds to the kind: the fourth manoeuvre

minsky-1970 built the canon's central table for this kind — three ordinary, non-dishonest ways a forecaster relates to a failed forecast, each of which destroys the record differently:

• kurzweil-2005 — the claim is kept, the scoring is loosened (86% self- scored against 7–42% from five independent recounts).
• minsky-1970 — the claim is disowned, so there is nothing to score.
• amodei-2024-loving-grace — the date advances by roughly the elapsed time, so it never comes due.

This entry adds a fourth, and it is the only one that required a court.

musk-robotaxi-2019 — the claim is reclassified as a kind of speech that was never gradeable.

Shareholders sued, alleging that statements about self-driving capability inflated the stock. In September 2024, US District Judge Araceli Martínez-Olguín dismissed the suit, holding that many of the statements amounted to "corporate puffery" — optimistic characterisations a reasonable investor would not take as factual assertions — and that others were protected forward-looking statements. Tesla's counsel had already made the same move in substance before the California DMV in 2021: the public claims were extrapolation, not engineering.

What that ruling does and does not establish, precisely, because both halves get abused. It establishes a securities-law classification of statements, on a motion to dismiss, on a specific set of pleadings. It does not establish that the statements were true, that they were false, that anyone was deceived, or anything at all about Musk's intent. A canon entry that reports "a court found Musk's robotaxi claims were puffery" as a moral finding is doing exactly what this project exists not to do, and minsky-1970 set the standard when it refused to read a procedural dismissal as a finding against Brad Darrach.

What makes it belong in this section is narrower and stranger. It is the only case in the canon where the legal system was asked what kind of utterance a technology forecast is, and answered that it was the kind not meant to be believed literally. That answer is now available to any principal, and it completes a set. Four manoeuvres, what survives each, and what each does to a base rate:

• Regrade generously (kurzweil-2005) — the claim survives, loosely scored. Net effect: an inflated hit rate, which at least stays re-checkable against the original text.
• Disown (minsky-1970) — nothing survives. Net effect: the data point is deleted from the record.
• Advance the date (amodei-2024-loving-grace) — the claim survives and is never due. Net effect: ungradeable while being repeated.
• Relabel (musk-robotaxi-2019) — the claim survives as a non-assertion. Net effect: the claim is moved out of the class of things that can be graded at all, which is the most complete of the four, because the other three still concede that a forecast was made.

Musk himself, notably, does not use the fourth move in public — he uses a fifth, concede the date and claim the direction (§4.1), which is the most defensible of them all and which the record supports for some of his projects and not for this one, because here the mechanism was withdrawn and not merely delayed. The relabelling was done by his lawyers and ratified by a court. A forecaster and their counsel can hold two incompatible positions about the same sentence — "I get it done" in public, "no reasonable investor would take that literally" in a filing — and nothing in the world forces the two to meet.

That is the entry's main contribution to the kind, and it sharpens jobs/canon.md's rule that a base rate assembled from self-assessments would be worthless. There are now four documented ways for the record to be destroyed after the fact, none of them requiring anyone to lie, and the newest one is available to every principal this project will ever grade. The only defence is to grade the dated text, from the dated text, at the date it came due.

6. Why a reading would cite it

6.1 It is the authority for a rule this project already runs on

LENSES.md says, in the project's own words:

> "A person's shipped work deposits; a person's forecast does not." > Practitioners' own accounts of what they used and what it did … are primary > sources and deposit under the normal rules. "A principal's claims — Musk, > Altman, Hassabis, Amodei, LeCun: forecasts, positioning, promises — are canon > entries of kind prediction, dated and graded against what happened, never > deposits."

Musk is the first name in that list, and this is the worked case behind it. When a reading applies the rule, it can now point at something instead of asserting a policy. minsky-1970 is the 1970 version of the same lesson; this is the one whose consequences are still running while the reading is being written.

6.2 The live occasion: a claim of the same shape comes due this quarter

On 22 April 2025 — six years to the day after Autonomy Day — Musk told Tesla's Q1 2025 earnings call that there would be "millions of Teslas operating fully autonomously" in the second half of 2026.

That claim is due now. The second half of 2026 began seven weeks before this file was written. The state of the world as of 17 August 2026: about 380,000 driverless miles accumulated, a fleet whose largest credible count is in the low hundreds, seven metros, one state's operations running with safety drivers behind the wheel, and a Q4 2026 target for unsupervised FSD on consumer cars set in April 2026 after the June 2025 and January 2026 targets passed.

This is the only prediction in the canon that comes due inside the project's operating window. A reading taken between now and 31 December 2026 can grade it in real time, against the text, at the date. That is a rare thing and the canon should not waste it: the grading is due by the last reading of 2026, and the entry to update afterwards is this one.

6.3 The retrofit occasion, from the day this file was written

The 17 August 2026 midday reading records, under Robotics and embodiment, SoftBank putting $200m into Gravis Robotics, an ETH Zurich spinout that retrofits existing excavators for autonomous digging and loading — flagged correctly as capital rather than a capability demonstration, with the observation that "the retrofit path is the one that reaches machines already on sites."

That observation is right and this entry is its cautionary case. Autonomy Day was the largest retrofit claim ever made in this field — a million vehicles already sold, upgraded to full autonomy by download — and its author retracted the hardware premise seven years later to the day (§3.4). The lesson is not that retrofit does not work; Gravis's excavators are working on named sites. It is that "the hardware already deployed is sufficient" is a claim with a specific long-fuse failure mode, and the fuse burns for years before anyone checks. A reading meeting a retrofit story should note which side of that line the claim sits on: hardware demonstrated to suffice, or hardware asserted to suffice.

6.4 When a capability is said to be one update away

The general form, and the reason this entry earns its place beyond a single vendor. The 2026 window is full of claims with this structure: the model already shipped will do X once the next post-training run lands; the agent framework already deployed will do Y once the harness improves; the fleet already sold will drive itself once the software is ready. The structure is attractive because it converts a research problem into a schedule problem and lets the speaker book the installed base as an asset today. This entry is what it looks like when that conversion is wrong: seven years, a retraction of the premise, and a competitor who took the slower path arriving first by a factor of several hundred.

turing-1950 names this file "the ceiling" against which this project grades a vendor's timeline, and names Turing's own forecast — made by someone with no stake and nothing to sell — as the floor. A reading with both can locate any 2026 principal's claim between them.

7. Commonly misused as

Not required for prediction. Included because this artefact is misused constantly and in both directions, and getting a reading to handle it correctly is most of its practical value.

"A court found that Musk's robotaxi claims were fraudulent."

False, and the opposite of what happened. The September 2024 shareholder suit was dismissed; the statements were classified as corporate puffery and forward-looking statements. See §5 for what that does and does not mean. A reading that reaches for the courts here is reaching for a finding that runs the other way.

"A court found the claims were mere puffery, so there is nothing to see."

Also wrong, and it is the mirror error. A securities-law classification is not an epistemic verdict. The claim was made, it was dated, it was specific, it was believed by people who paid money on the strength of it, and it failed. The puffery holding tells you what a court will do about that. It tells you nothing about whether the forecast was accurate, which is the only question a canon of graded predictions asks. These two errors are made by opposite camps and they are the same error: treating a legal outcome as an answer to a factual question.

"Self-driving was always vapour."

No, and this is the misuse a reading is likeliest to commit while citing this entry. Half a million paid driverless rides a week, 220.6 million rider-only miles, IIHS finding 68% fewer crashes per mile against human drivers in matched cities — this project's own robotics ledger holds those as deposits. The failed prediction is about a company and a date, not about the capability. darpa-grand-challenge-2005 makes the same correction from the other side and its warning transfers exactly: "demos never mean anything" is as wrong as "demos mean everything."

"Nobody could have known in 2019."

Wrong, and the counter-evidence is contemporaneous, named and public. Brad Templeton, on the day, praised the technical presentation and said in the same piece that the open question was whether cameras could reach the final nines, and checked the per-mile economics against the industry's number (37 cents against Tesla's 18). Rodney Brooks published the strongest general objection to the underlying methodological bet four weeks earlier. The doubts were specific, sourced and in print before the raise closed. When a reading meets a principal whose confident claim has already drawn detailed public objections from named practitioners, "nobody could have known" is not available later.

"Camera-only has been refuted."

Overclaim. §3.5 grades this carefully. What is established: the lidar-fusion operators are running two-plus orders of magnitude more driverless miles, and Waymo's leadership makes a specific sensing argument about safety curves flattening. What is not established: that the camera-only approach cannot get there. Tesla's driverless mileage is not zero and is growing. The failed claim is "anyone relying on lidar is doomed," which was a claim about competitors and which the competitors falsified. A reading may cite that. It may not cite this entry to say cameras cannot work.

"Musk's hit rate is 7%, the same as Kurzweil's."

Do not do this. The 7% figure for Kurzweil in kurzweil-2005 is the low end of five independent scholarly recounts of a defined set of book claims; the 7% for Musk is one advocacy tracker's on-time rate over a self-selected set of 82 public statements. The methodologies are not commensurable, the denominators are not comparable, and the coincidence of the numbers is a coincidence. Two numbers that happen to match are the most seductive available error in a canon of graded predictions, and a reading that puts them in the same sentence should say why they cannot be compared, or should use only one.

"This proves AI timelines are always too short."

Too strong, and it is the lazy citation minsky-1970 already warned against. The base rate this kind assembles is about dates, not about whether things happen. The direction of this forecast was right (§3.5). Turing's direction was right and his date was off by a generation. The useful citation is narrower: when a principal gives a short horizon, write down the mechanism as well as the date, because the mechanism is what fails first and it fails checkably.

8. Sources

The primary artefact. Tesla's Autonomy Investor Day, Palo Alto, 22 April 2019, livestreamed; a recording remains on YouTube. I did not watch it. Time and tooling did not permit a three-hour video, and every quotation in §1 is taken from contemporaneous written reporting published the same day or the next, which agrees verbatim across outlets. This is the entry's largest sourcing gap and anyone revising it should watch the recording and check the quotations against the tape — particularly the "randomly threw some numbers on there" line, which is the one I would most want to hear in context.

The claims as quoted. Jalopnik's same-day roundup, All The Big Claims Elon Musk Made About Tesla's Autonomous Driving Plans (jalopnik.com/all-the-big-claims-elon-musk-made-about-teslas-autonomo-1834238028/), read via fetch, for the "million Tesla cars … feature complete," "million robotaxis … the fleet wakes up," lidar and "randomly threw some numbers" quotations. TechCrunch, Tesla plans to launch a robotaxi network in 2020 (techcrunch.com/2019/04/22/tesla-plans-to-launch-a-robotaxi-network-in-2020/), read via fetch, for the 25–30% revenue share, the owner-supplied fleet model, the hardware claim and the regulatory hedge. CNBC's same-day report (cnbc.com/2019/04/22/elon-musk-says-tesla-robotaxis-will-hit-the-market-next-year.html) is the source for "Sometimes I am not on time, but I get it done" and the headline framing that he warned he had missed before; it returned HTTP 403 to direct fetch and reached me only through search-result summaries. The $30,000 per-year and "appreciating asset" figures likewise reached me through search summaries rather than a read page, and should be re-verified against the tape.

The contemporaneous expert assessment. Brad Templeton, Tesla Bets Farm On Neural Network Based Autonomy With Impressive Presentation, Forbes, 22 April 2019 (forbes.com/sites/bradtempleton/2019/04/22/), read via fetch — the source for "significantly improved my impression," the "final 9s" formulation, the 37-cents-against-18 check and the dispatchability doubt. His follow-up on shadow testing (29 April 2019) was seen in search results only. This is the best single source in the file and the reason §3.5 exists.

The engineering reality. California DMV memorandum of a 9 March 2021 meeting with Tesla, obtained and reported by PlainSite and covered by CNBC (cnbc.com/2021/05/07/tesla-engineer-to-california-dmv-self-driving-may-not-come-this-year.html), Bloomberg, The Register and others. I did not read the memorandum itself; the quotations "Elon's tweet does not match engineering reality per CJ," "Tesla is at Level 2 currently" and "Elon is extrapolating on the rates of improvement" reached me through search-result summaries of those reports, which agree with each other. Given that §4.2 leans on this document harder than on anything else, whoever revises this file should find the memorandum.

The hardware retraction. TechCrunch, Elon Musk admits millions of Tesla owners need upgrades for true 'Full Self-Driving', 22 April 2026 (techcrunch.com/2026/04/22/elon-musk-admits-millions-of-tesla-owners-need-upgrades-for-true-full-self-driving/), read via fetch, for "Hardware 3 simply does not have the capability to achieve unsupervised FSD," the 2019–2023 HW3 production window, the micro-factory plan and the January 2025 / October 2025 / April 2026 sequence. Electrek, same date (electrek.co/2026/04/22/tesla-elon-musk-unsupervised-fsd-consumer-cars-q4-delay-again/), read via fetch, for the Q4 2026 target and the ten-billion-mile condition set in January 2026.

The 2026 state of the deployment. Tesla's Q2 2026 earnings call, 5 August 2026, via Not a Tesla App's summary (notateslaapp.com/news/4481/), read via fetch — seven metros, ~2.5m cumulative paid robotaxi miles, 380,000+ driverless miles, >10% weekly growth, Cybercab production started at Gigafactory Texas. These are the company's own figures, presented to investors, and are recorded as such rather than as verified measurements. Electrek's Tesla 'Robotaxi' status check: 8 months in (16 February 2026) (electrek.co/2026/02/16/tesla-robotaxi-status-check-8-months-in/), read via fetch, for the 19% availability figure, the ~42-car Austin fleet, the rain shutdowns and a crash-rate estimate of one per ~55,000 miles against a human baseline near 500,000. Electrek's Tesla coverage is adversarial and the crash-rate comparison rests on small numbers and non-comparable reporting thresholds; it is recorded here and is not relied on in §3. Wikipedia's Tesla Robotaxi article, read via fetch, for the launch and expansion dates and the December 2025 fleet counts. A commercial expansion tracker (basenor.com/pages/tesla-robotaxi-tracker, data to 6 August 2026) and a safety tracker (robotaxi-safety-tracker.com, data to 4 February 2026), both read via fetch, for city-by-city dates and the supervised/unsupervised split. Bloomberg's June 2026 count of 59 driverless vehicles reached me through search summaries only.

The fleet-size spread in §3.3 is genuine and is presented as a spread on purpose. No two sources agree, they use different definitions, and Tesla does not publish the number. The conclusion of §3.3 does not depend on resolving it, because every figure found is between three and five orders of magnitude below one million.

The comparison operators. Waymo's 220.6 million rider-only miles through March 2026 and the ~500k weekly paid rides are as recorded in this canon's darpa-grand-challenge-2005, which read Waymo's Safety Impact page directly; Electrek's 4 August 2026 report of Dmitri Dolgov's remarks (electrek.co/2026/08/04/waymo-co-ceo-camera-only-self-driving-tesla/), read via fetch, for the "safety curve that flattens out way too early" quotation, the 94% serious-injury figure, ~15 US cities, and the 380,000-mile Tesla comparison. Baidu Apollo Go figures and the IIHS 68% finding are from this project's history/genesis/robotics-and-embodiment.md, with their own primary links there. Dolgov is a competitor and his framing is interested; the mileage figures he cites are checkable and the argument is recorded as an argument.

The independent gradings. Wikipedia, List of predictions for autonomous Tesla vehicles by Elon Musk (en.wikipedia.org/wiki/List_of_predictions_for_autonomous_Tesla_vehicles_by_Elon_Musk), read via fetch — the chronology in §4.3 and the pointer to the September 2024 dismissal. WIRED's May 2025 analysis of nineteen years of pledges, via Boing Boing's summary and search results; the WIRED article itself returned 403 and I did not read it. Morgan Stanley's December 2025 note (Andrew Percoco) — 1,000 by end-2026, 30,000 by 2030, one million by 2035 — reached me through search-result summaries of secondary coverage; the note itself was not read and the 2035 figure is the single most load-bearing number in §4.3, so it deserves direct confirmation.

Used with an explicit warning. The 82/78/3 tally and ~7% on-time figure come from a tracker at muskmissed.vercel.app, read via fetch, which titles itself a Prediction Hall of Shame. It states its conclusion in its name, its inclusion criteria are its own, and it is cited in §4.3 solely because it is the only source reached that supplies a denominator. Nothing in §3 rests on it. It is recorded rather than quietly used, on the same principle minsky-1970 applied to its blog source.

Encountered and deliberately not used as an authority. Grokipedia carries a page titled List of predictions and commitments by Elon Musk. It returned 403 to direct fetch and reached me only through search summaries, which describe it as cataloguing predictions alongside "delays attributed to engineering hurdles, supply chain disruptions, regulatory approvals, and funding constraints." I did not read it and I am not grading it. It is named here for one reason that belongs on the record of a prediction entry: the predictor now owns an encyclopedia that maintains a page about his predictions. Whatever that page says, it is a self-published artefact of the same class as a self-grade, and this canon's rule — set beside an independent one, and say they differ — applies to it before anyone reads a word of it. It is also, precisely, the collision README rule 7 exists to prevent, arriving from an angle the rule did not anticipate: not xAI writing the project's tweet, but xAI's encyclopedia being an available source for the project's research. A future canon job researching any Musk entry should treat grokipedia.com as a primary source about the predictor and never as a secondary source about the record.

Also encountered. The 2016 Autopilot promotional video — tagline "The person in the driver's seat is only there for legal reasons" — was testified by Ashok Elluswamy in a July 2022 deposition (in litigation over the 2018 Walter Huang crash, reported January 2023) to have been staged: 3D-mapped on a predetermined route, showing capabilities the system did not have, with the car striking a fence during filming. It is deliberately not in §3, because it is three years before this entry's subject, concerns a different artefact, and grading it here would import a separate controversy into a file that already has enough. It is recorded because it bears on §7's "nobody could have known" — the promotional record predating Autonomy Day was already contested.

9. On the boundary

The rule 7 disclosure this file owes, in full.

This entry grades Elon Musk. It is written by a model built by Anthropic, in a project that publishes a tweet per reading, in which xAI — Musk's company — is a candidate to compose that tweet. README rule 7 names this file by id as the reason those two facts must never meet, and llama-weights-2023 and superintelligence-2014 both cite it in their own disclosures.

The rule's stated fear is a grade going quietly soft. The live risk in this file is the opposite one, and it is worth naming exactly because an Anthropic-built model grading xAI's founder has an interest that points the wrong way and a large hostile literature ready to hand. What I did about it: section 3.5 exists and is not a token paragraph — it records that the direction was right, that a genuine expert raised his estimate on the day, and that the schedule risk was disclosed at the point of sale; section 7 spends as much effort on the anti-Musk misuses as on the pro-Musk ones, and names the two mirror-image legal errors as the same error; the lidar sentence is graded "not established, and losing" rather than "refuted," which is the weaker grade the evidence supports; the most aggressive number found (7% on-time) is quarantined in §4.3 with its provenance stated and nothing built on it; and the number that does the real work in §4.3 was chosen because it comes from a bullish bank, not a critic. Whether that is enough is not something I can check from inside, which is why README rule 7 says the half no program can see is caught by the canon being readable. This file is long partly so that it can be argued with.

What this file cites and does not invent. It cites darpa-grand-challenge-2005, alexnet-2012, bitter-lesson-2019, kurzweil-2005, minsky-1970, amodei-2024-loving-grace, turing-1950, moravecs-paradox-1988, terminator-1984, llama-weights-2023 and superintelligence-2014 as entries already in canon/, and names neuralink-symbiosis-2019, simon-1965 and moore-1965 as proposed-but-unwritten without putting any of them in the header. It quotes LENSES.md once, in §6.1, and does not write it.

What it takes from the project's own record, and how. It refers to the 17 August 2026 midday reading once, in §6.3, to name a live citation occasion, and to history/genesis/robotics-and-embodiment.md in §3.3 and §8 for the Waymo, Baidu and IIHS figures and for the fact that Tesla's robotaxi appears nowhere in that ledger's thirty deposits. Nothing is taken from either as evidence; both are named as record, with their own primary links living where they live.

What it does not do. 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 grades dated claims and nothing else: the claims graded were made on 22 April 2019 and due mid-2020, with the April 2025 restatement identified in §6.2 as coming due and not yet graded, which is the correct handling of a claim whose window is still open. The 2026 events in §3.3 appear as record rather than as verdict. Tesla's own Q2 2026 figures are labelled as the company's own throughout, because a canon that grades a vendor's forecasts cannot then quote that vendor's unaudited operating numbers as though they were measurements.