Musk Wiki

Autonomous driving

NextChildhood adversity

Autonomous driving

Musk first set out his case for autonomy formally in the 2016 master plan, and it stands on two legs. One is a safety argument with a number attached: self-driving can be an order of magnitude safer than a human, and fleet-scale data will prove it. The other is economic: an autonomous car can be shared, which turns a depreciating asset into one that earns money. By 2025 the whole vision has folded into the wider physical-AI bet of Sustainable abundance. And in the May 2025 CNBC interview it finally became a concrete robotaxi rollout with a date on it.

The two pillars

  1. Safety through data. Autonomy isn’t a gadget; it’s a way to save lives, and the argument is flatly statistical. Prove it with billions of fleet miles before you call it safe.
  2. Shared fleets. That same autonomy lets owners rent their cars into a network, which Musk calls “exactly why” the self-driving system is being built at all. The robotaxi grows straight out of this.

The 2016 plan is unusually honest on one point: “self-driving ready” hardware is not the same thing as “self-driving capable” software, and a great deal of that software is still unsolved. That kind of hedge is rare in an otherwise supremely confident document. The sentence that says it is quoted below.

Evidence

The quantified safety ambition:

“We are currently developing the technology to enable self-driving cars that can be 10x safer than manually driven cars.” ↗

The stated purpose of building autonomy — shared fleets:

“Let me emphasize this point: autonomous cars can be shared. We believe that owners of Tesla vehicles will be empowered with the option to share their car with a network of other drivers, thus significantly offsetting the monthly loan or lease cost. This is exactly why we’re building the self-driving system.” ↗

The honest caveat about how much remains unsolved:

“It is important to note that self-driving ready hardware does not mean that the car is self-driving capable. The software required to make the car fully self-driving is complex.” ↗

The 2025 framing of autonomy as a benefit to all:

“Autonomous vehicles have the capacity to dramatically improve the affordability, availability and safety of transportation while reducing pollution, particularly in our increasingly dense global cities.” ↗

2014–2015 — Autopilot arrives, and the timeline starts sharpening (Tesla earnings)

His earliest dated autonomy forecasts are spoken, not written: the 2014–2015 Tesla earnings calls, where Autopilot shows up as a product and Musk first commits to a full-autonomy timeline on the record. Then, quarter by quarter, you can watch the timeline tighten. The first call, Q3 2014, is cautious and hedged on the regulators:

“it’s probably no sooner than 7 years from now and could be up to 10, I think.” ↗

He pairs that caution with full confidence that Tesla will lead the race (“it’s quite likely that Tesla will be the leader in making cars like that”). By Q2 2015, with Autopilot about to ship, two design convictions surface. One is the bet to own the data set (“we don’t really have much choice but to create our own data set for driving”). The other is the aviation analogy he uses to put responsibility on the driver at launch, a view he revises sharply in later years:

“it’s much like the Autopilot in a plane where you turn the Autopilot on in a plane, but there’s still an expectation that the pilot will pay attention to what the plane is doing, and, you know, won’t sort of go to sleep or just disappear from the cockpit.” ↗

By Q3 2015 he restates the timeline for the whole industry, and now the conviction is total. A car without autonomy becomes “like owning a horse,” the same image he reaches for four years later on Lex #18:

“I think it will be quite unusual to see cars that don’t have full autonomy, let’s say, in for new car production in the 15 to 20 year time frame. And for Tesla, it will be a lot sooner than that.” ↗

“any cars that are being made that don’t have full autonomy full autonomy will have negative value. It will be like owning a horse.” ↗

So the autonomy thread is born here, out loud. The habit of dated predictions, the “horse” metaphor, the own-the-whole-stack strategy: all of it is in place two and three years before the 2017 and 2019 restatements that follow. The 2014 calls carry no speaker labels; the Q2 2015 call labels Musk explicitly. More on both in Tesla Earnings Calls 2013-2015.

2016–2018 — the timeline peaks, then slips: coast-to-coast and the data moat (Tesla earnings)

The 2016–2018 earnings calls are the most falsifiable stretch of the whole thread. The optimism crests and then walks itself back, quarter after quarter, on the same dated record. It opens in 2016 with pure timeline-optimism (“full autonomy is gonna come a hell of a lot faster than anyone thinks it will,” Q2 2016) and the moral framing he keeps for years (“it would be… morally wrong not to allow autonomous driving,” Q3 2016). Q1 2017 is the most aggressive claim in the record: the hardware “for at least Level 4, current Level 5 autonomy has been in every Tesla”, so “it’s a matter of upgrading the software and we can achieve Level 5.”

Then comes the coast-to-coast demo deadline, and its slow-motion collapse. Each new date arrives wrapped in self-aware humor: “I may have egg on my face on that front” (Q2 2017), then “three months, six months at the outside” (Q4 2017), then the admission that doing it on a fixed route “would be kind of gaming the system” (Q2 2018). The Q4 2017 call is also where the anti-lidar belief first surfaces on an earnings call. Lidar is “a crutch that will drive companies to a local maximum,” vision the path to “the global maximum”, two years ahead of the 2019 “fool’s errand” version. The same call reaches for AlphaGo to explain why he expects sudden exponential progress, and lands his trademark hedge-then-conviction: “perhaps I am wrong… I am quite certain that I’m not”. By 2018 the argument moves off the technology and onto the moat. Tesla has “millions of cars in the field with full autonomy capability”, rivals have “5%… of the miles that Tesla has”, and the only thing left in the way is “regulatory approval”. Then it lands right back where it started, at “towards the end of this year” (Q4 2018), a near-perfect bookend on the optimism it opened with. Ten of the twelve transcripts label Musk; Q3 2017 and Q2 2018 don’t. See Tesla Earnings Calls 2016-2018.

2019–2021 — the same prediction, every quarter, never resolving (Tesla earnings)

Across the 2019–2021 earnings calls the autonomy prediction does what it does best on this medium. It comes back almost every quarter, and it never resolves. 2019 opens with the scaffolding of the idea: a three-level account of what “feature complete” actually means (“we think the car is safe enough to be driven without supervision… Then the 3rd level would be that regulators are also convinced,” Q3 2019). Alongside it, the value gets pitched as “the biggest step change increase in asset value in history by far” (Q3 2019). Then the date gets restated at peak confidence, almost every quarter. The car will drive “from your home to your office… with no interventions by the end of the year” (Q1 2020); “I’m highly confident the [car] will drive itself with the liability in excess of humans this year” (Q4 2020); “I am highly confident that we will get this done” (Q1 2021).

The Q4 2021 call delivers the strongest lines in the entire record: “my personal guess is that we’ll achieve Full Self-Driving this year”, and “I would be shocked if we do not achieve Full Self-Driving safer than a human this year. I would be shocked.” The reasoning gets sharper alongside the date, but the optimism never bends. Self-driving becomes “a pretty significant Part of artificial intelligence, specifically real world artificial intelligence” (Q1 2021). He says “being safer than a human is a low standard, not a high standard”, because people are “very, very lossy” (Q4 2021). The elevator-operator inversion arrives in full: “Autonomy will become so safe that it will be unsafe to manually operate the car” (Q2 2021), with the long-horizon bet that “all transport will go autonomous” (Q3 2020). And the value lever gets pushed to its limit, “one of the most valuable things that is ever done in the history of civilization” (Q2 2021), “the biggest increase in asset value of any asset class in history” (Q4 2021). This stretch is the cleanest demonstration that the autonomy timeline is a recurring posture of confidence rather than a converging estimate. His own explanation for why sits on that same Q1 2020 call, in the 50th-percentile self-model. Some quarters label Musk; the 2019 and Q2 2021 calls are label-light, with Musk identified by his opener and Q&A position. See Tesla Earnings Calls 2019-2021.

2017 — “in about 10 years,” and the elevator (World Government Summit)

Two months before TED2017, the February 2017 World Government Summit conversation leads not with safety or vision but with a timing call. His forecast:

“it will be very unusual for cars to be built that are not fully autonomous.” ↗

He puts the horizon at “about 10 years”, and reaches for the same picture he uses on Lex in 2019. Getting into a car becomes getting into an elevator, the elevator operator now long obsolete:

“So, getting in a car will be like getting in an elevator.” ↗

Behind it sits the narrow-vs-general distinction the rest of his AI thinking leans on. Vehicle autonomy is “narrow AI … narrowly trying to achieve a certain function,” a different category entirely from the “deep” AI he warns about (paraphrased; the distinction runs across several cues). In 2017 he still sells autonomy as a near-term convenience, “extreme levels of safety,” elevators “taken for granted”, and not yet the safety crusade the 2019 “two-ton death machine” line turns it into. But the timing call and the elevator are already here, two years before #18 and eight before the 2025 robotaxi date. See World Government Summit 2017.

2017 — vision only, and autonomy makes traffic worse (TED)

The 2017 TED conversation is his earliest spoken statement of the camera-first thesis, eight years before the 2025 “the road system is designed for AI” line says the same thing. The argument is biological. The road is built to be read with eyes, so the matching machine sensor is a camera, and the whole problem collapses onto vision:

“This is just using passive optical, which is essentially what a person uses.” ↗

“and so once you solve cameras or vision, then autonomy is solved. If you don’t solve vision, it’s not solved.” ↗

“You can absolutely be superhuman with just cameras.” ↗

This is the 2019 vector-space diagnosis (“the hardest thing is having accurate representation … in vector space”) two years early and in seed form, with the entire difficulty placed in perception-from-vision. Here he states it as a flat reduction: solve vision, and autonomy follows.

His safety model is the probabilistic one he uses everywhere. There is no zero-risk baseline, only a human accident rate to beat:

“The thing to appreciate about vehicle safety is this is probabilistic.” ↗

“So really the key threshold for autonomy is how much better does autonomy need to be than a person before you can rely on it?” ↗

And the shared-autonomy fleet, the seed of the later robotaxi, is stated as a sure thing with only its timing left open:

“That’s 100 percent what will occur. It’s just a question of when.” ↗

Autonomy worsens traffic — the induced-demand twist

The most distinctive move of the 2017 talk is counter-intuitive. Many people assume self-driving eases congestion. Musk argues the opposite. Cheap shared autonomy is cheaper than a bus, so it induces far more driving, and surface traffic gets worse. Which is exactly why he needs tunnels:

“A lot of people think that when you make cars autonomous, they’ll be able to go faster and that will alleviate congestion.” ↗

“the affordability of going in a car will be better than that of a bus.” ↗

“So the amount of driving that will occur will be much greater with shared autonomy, and actually traffic will get far worse.” ↗

It’s a small but telling case of his first-principles reflex working on a second-order effect. Instead of assuming autonomy is plainly good for cities, he reasons through the demand response and lands on the opposite of the popular answer.

2019 — a car without autonomy is “a horse” (Lex Fridman #18)

The 2019 Lex Fridman conversation states the belief at its purest. Not as a feature, but as a claim about obsolescence. Both written master-plan pillars are here, safety and shared fleets, but the 2019 version is sharper. Autonomy is now what decides whether a car is useful at all.

The obsolescence thesis, stated flatly:

“in the future, any car that does not have autonomy would be about as useful as a horse.” ↗

His mental model for how the fleet learns: every human intervention is, by definition, a fault signal to learn from.

“the way to look at this is view all input as error.” ↗

Moments later he puts it bluntly. If the user had to touch the controls, something is already wrong, so all input is error (paraphrased; the line breaks across two caption cues, ↗).

And then the inversion he’s best known for on this subject: that the dangerous, soon-to-be-archaic default is manual driving, not autonomy.

“Frankly it’s pretty crazy letting people drive a two-ton death machine manually.” ↗

“it’s gonna seem like a mad thing in the future that people were driving cars.” ↗

He drives it home with an elevator analogy he comes back to twice. Nobody wants an elevator operator pulling a lever between floors anymore, because the automated elevator is safer. And once a system is dramatically safer than a person, putting a human back in can actually decrease safety (paraphrased; the elevator passage breaks across many caption cues). His one-line reason for trusting the software to get there is the same nonlinear-progress intuition the 2025 interview phrases as “slowly but then all at once”:

“The rate of improvement is exponential.” ↗

This 2019 baseline matters for the evolution-of-views story. Six years before he would put a robotaxi date on the calendar, the belief is already fully formed and stated in nearly the same words: autonomy as the line between a useful car and a horse, safety proven by treating every human input as error.

2019 — the case against lidar: “a fool’s errand” (Tesla Autonomy Day)

Tesla’s first investor event on self-driving, Autonomy Day in April 2019, came two months after Lex #18, and it’s where the anti-lidar position that anchors the camera-first thesis gets said out loud. Asked whether the new full-self-driving computer could also handle lidar input, Musk doesn’t reach for an engineering trade-off. He dismisses the sensor outright, on principle:

“lidar is is a fool’s errand” ↗

He adds that anyone relying on it is “doomed,” and that the sensing is expensive and unnecessary (the caption garbles the clause as “anyone luck relying on orb lidar is doomed doomed expensive expensive sensors”, so it’s paraphrased; “doomed” is the word widely reported from the event, ↗). Then he reaches for the image he’s remembered for, extra sensors as surgically useless organs:

“it’s like having a whole bunch of a expensive appendices that compact one appendix is bad well now don’t put a whole bunch of them that’s ridiculous” ↗

This is the first-principles move that runs under the whole camera-only stance: build the design out of the structure of the problem, a road meant to be read with eyes, instead of piling on hardware. It’s the same argument as the 2017 TED “solve vision and you solve autonomy” and the 2025 “the road system is designed for AI”, here in its sharpest and most-quoted form. The deeper anti-lidar engineering at this event, the sensor-fusion “confusion,” depth-from-vision, “lidar is a crutch”, is Karpathy’s vision talk, not Musk, and isn’t attributed to him.

The public robotaxi prediction — “over a million robotaxis … next year for sure”

Autonomy Day is also where Musk first puts a public date on the shared-fleet idea. Tesla will have autonomous robotaxis running the following year, and within roughly a year more than a million of them on the road, switched on by an over-the-air update.

“we expect to have the first operating Robo taxis next year with no one in them next year” ↗

“but next year for sure we will have over a million Robo taxis on the road” ↗

His reason is the same nonlinear-progress intuition he returns to again and again (the #18 “rate of improvement is exponential,” the 2025 “slowly but then all at once”). Data and software improve exponentially while people extrapolate in straight lines (paraphrased; multi-cue, ↗).

Tone note (timeline did not hold). Recorded as Musk’s stated framing, not as fact. The “over a million robotaxis … next year [2020]” prediction did not happen on schedule. No driverless Tesla robotaxi fleet launched in 2020; a small, geofenced, safety-monitored pilot began in Austin only in 2025. It’s one of the clearest optimism-and-timeline datapoints in his record: the belief stated years ahead of the capability, the schedule slipping by a wide margin. And he attaches his own caveat in the same breath (“sometimes I’m not on time but I get it done”). The framing and the slip are what stand on the record, with no verdict on the prediction itself.

2019 — autonomy as a vector-space problem (Lex Fridman #49)

Seven months after #18, the November 2019 Lex Fridman conversation (#49) supplies the part #18 left implicit: which part of self-driving is the hard part. His answer is perception as representation. Building an accurate model of the world, not deciding what to do with it:

“The hardest thing is having accurate representation of the physical objects in vector space.” ↗

“Once you have an accurate vector space representation, the planning and control is relatively easier.” ↗

This is the first-principles breakdown that sits under the later camera-only stance. If the binding constraint is reconstructing the scene in vector space from vision, then the engineering problem is perception, and planning comes downstream and comparatively easy. It’s the 2019 seed of the 2025 “the road system is designed for AI” argument: both put the whole difficulty in turning images into an accurate world-model. See Lex Fridman #49 (2019).

2021 — the objective in one rule: “don’t crash” (Tesla AI Day)

In the Q&A at Tesla AI Day (August 2021), Musk squeezes the entire autonomy objective into a single rule, and folds a first-principles claim into it about what perception must not require. The prime directive:

“the prime directive is ‘don’t crash’” ↗

And the consequence he draws from it. Collision-avoidance can’t depend on classification; the car has to dodge an object whether or not it can name it:

“Even if it’s a UFO that crash-landed on the highway – still don’t hit it. It should not need to recognize it in order to not hit it” ↗

This is the same build-the-design-from-the-problem reasoning as the 2017 “solve vision and autonomy is solved” and the 2019 “lidar is a fool’s errand” dismissal, now pointed at the objective function. Define safety as “don’t hit it,” and keep that rule logically prior to, and independent of, whatever the perception system can label. The vision and planning engineering at this event was Karpathy’s and Ashok’s; Musk’s contribution here is the objective. See Tesla AI Day 2021.

2021 — biological vision in silicon: “recreate what humans do” (Lex Fridman #252)

The 2021 Lex Fridman conversation (#252) pushes the vector-space diagnosis into its most explicit biological-mimicry form, and it’s as much a window into how Musk thinks about the human brain as about cars. The whole problem, in his telling, is rebuilding human perception in silicon. The road was made for eyes and neural nets, so the answer has to be cameras and neural nets:

“to solve self-driving you have to solve. You basically need to recreate what humans do to drive, which is humans drive with optical senses, eyes, and biological neural nets.” ↗

What makes the conversation unusual is the long stretch where he reasons about the brain itself. The eye paints in the color and the blind spot. The brain is “constantly trying to forget as much as possible” because memory is expensive, boiling perception down to “the smallest possible vector space of only relevant objects,” and safe driving needs object permanence, remembering the child hidden behind the truck, the same “memory across both time and space” the cars need. These are paraphrased, since they run across long multi-cue exchanges, but they show his recurring habit: treat the mind as hardware to be reverse-engineered, the brain-as-machine reflex aimed at perception. He also calls the deployment bar, not parity with humans but “two or three times” safer, and predicts the car will eventually maneuver “far more than what … James Bond could do” (paraphrased). See Lex Fridman #252 (2021).

2023 — “the single biggest asset value increase in history” (Tesla Shareholder Meeting)

The 2023 Tesla shareholder meeting is where the shared-fleet economic case gets its biggest dated statement yet. This is the second pillar, written into the 2016 plan and reframed at Autonomy Day 2019 as a “factor of five” jump in utility. His logic is the idle-asset one. A private car sits unused about 90% of the week, so making it autonomous multiplies how much it’s used, and flipping the existing fleet on by software is, in his telling, historic:

“I think will be the single biggest asset value increase in history.” ↗

Back on it in the Q&A, he pushes the claim a notch further. A car that costs the same but is used about five times as much could run “80% margins,” so the moment of the flip is:

“it’s probably going to be the biggest asset value step change in in history of Earth.” ↗

Same shared-fleet idea, autonomy turning a depreciating asset into one that earns, restated in May 2023 at full volume, the belief held and amplified while the capability is still unshipped. It’s a clean optimism-and-timeline datapoint. By 2023 the 2019 “fivefold” framing and “over a million robotaxis next year” prediction had not materialized, yet the conviction is stated larger here, not smaller, the same pattern the Autonomy Day Tone note records. The FSD “march of nines,” the V12 end-to-end neural net, and the “local maximum / logarithmic curve” reasoning from the same Q&A are autonomy engineering, covered on the source page. See Tesla Shareholder Meeting 2023.

2024 — the robotaxi becomes a product: “individualized mass transit” (We, Robot)

The October 2024 “We, Robot” event is where the long-abstract robotaxi finally gets a physical object, the two-seat Cybercab with no wheel and no pedals, and where Musk lands on the phrase he uses for the economics. The shared-fleet idea from the 2016 plan, amplified at the 2023 meeting, comes back here as a category name:

“The cost of autonomous transport would be so low, you can think of it as individualized mass transit.” ↗

The payoff he leads with is the one the 2025 master plan later turns into the optimization target. Not speed, not cost, but reclaimed time:

“With autonomy, you’ll get your time back.” ↗

And the safety case arrives flat and absolute, the 2016 “10x safer” target stated as plain fact:

“Autonomous cars will be ten times safer than human drivers.” ↗

The reasoning hasn’t budged. The idle-asset math (a car used about 10 of 168 hours, idle the rest, worth “five times—ten times—more” if autonomous) restates the 2023 “biggest asset value increase” argument, and the elevator-operator analogy returns yet again from #18. What 2024 adds is the product. The Cybercab’s pricing (below $30,000), production timing (“2026 … before 2027”), roughly 20¢/mile operating cost, the Robovan, inductive charging, and unsupervised FSD “in Texas and California next year” are product-and-timeline spec, covered on the source page. See We, Robot (2024).

2025 — the robotaxi rollout, and the vision-not-lidar case (CNBC)

The May 2025 CNBC interview is where the long-promised autonomy meets an actual launch, with paid Austin robotaxis the following month. Two things stand out: his read on timing, and his philosophy of sensors.

The nonlinear view of progress, his reason for being confident now after years of missed dates:

“these things happen slowly but then all at once” ↗

A genuinely cautious rollout plan, start with about 10 cars and scale to 20, 30, 40, framed as deliberate restraint rather than a land-grab:

“we want to deliberately take it slow” ↗

And his first-principles defense of Tesla’s camera-only approach against Waymo’s lidar and radar. The road was built for biological vision, so the machine answer is vision, not lasers:

“the way that the road system is designed is for AI” ↗

He pushes it into a redundancy argument. With several sensor types running at once (camera plus lidar and radar), the systems “tend to get confused” about which one to believe, so Tesla “turned off the radars”. In his telling, the disagreement between sensors is itself a cause of accidents. It’s the same First principles move that runs through the rest of his thinking: build the design out of the structure of the problem, a road made for eyes, rather than out of stacked-up hardware.

2022–2026 — the boy who cried FSD, in his own words (Tesla earnings)

The 2022-2026 earnings calls keep the near-quarterly FSD prediction running at peak confidence: “I think we will achieve that this year” (Q1 2022), “Our probability of that occurring is 100%. I think you know, we’re almost there” (Q3 2022), “Our internal estimate is Q2 next year to be safer than human” (Q3 2024), “millions of Teslas operating fully autonomously in the second half of next year” (Q1 2025). What’s new is that the recurrence-without-resolution becomes visible to Musk himself. He says so on the record: “I know I’m the boy who cried FSD” (Q2 2023), “I know people have said, well, Elon’s the boy who cried wolf… there’s a damn wolf this time” (Q4 2024), “my predictions on this have been overly optimistic in the past” (Q2 2024). He even names the mechanism for why the dates slip: “it will curve over logarithmically”, “a series of stacked log curves” (2023). And he keeps moving the bar down (“march of nines”; “stopping autonomy means killing people,” Q1 2024; “it is only a matter of time before we exceed the reliability of humans,” Q1 2024).

The clearest evolution-of-views marker is the Q1 2026 walk-back of an old hardware promise: “Hardware 3 simply does not have the capability to achieve unsupervised FSD. We did think at one point it would have that”. That’s a flat admission that an earlier confident claim was wrong. The same call recasts the robotaxi bottleneck as the car’s own timidity: “the car basically gets paranoid and gets stuck… it sometimes gets scared to do things.” And the horse-and-flip-phone inevitability gets self-cited year after year (“I said this many years ago…,” Q1 2025), ending at “the only manually driven car will be the new Tesla Roadster” (Q1 2026). Every quarter here labels Musk explicitly. See Tesla Earnings Calls 2022-2026.

2015 — the goal and the prediction (tweets)

The 2015-2017 tweets add two early dated markers. In November 2015 he names the goal that organizes the whole program: “Ramping up the Autopilot software team at Tesla to achieve generalized full autonomy.” And in March 2015 he states the value-and-prediction pair he keeps coming back to. First the value, on human freedom, “Tesla is strongly in favor of people being allowed to drive their cars and always will be”. Then, right behind it, the social forecast that safe autonomy may one day outlaw exactly that, “when self-driving cars become safer than human-driven cars, the public may outlaw the latter. Hopefully not”. See Elon Musk Tweets 2015-2017.

2021 — the vision-only pivot, in real time (tweets)

The 2021-2022 tweets catch the turning point of his self-driving philosophy as it happens: the 2021 decision to pull radar and bet everything on cameras, argued from first principles. The thesis is that the road system is built for vision, so vision is the thing that has to be solved:

“What has become absolutely clear is that the plethora of self-driving corner cases can only be solved with real-world optical intelligence. This is how humans designed the road system to work. Once you have that in silicon form, everything else is just icing on the cake.” ↗

The sensor-fusion rejection that justified deleting radar: when the two sensors disagree, trust the more precise one instead of fusing them.

“When radar and vision disagree, which one do you believe? Vision has much more precision, so better to double down on vision than do sensor fusion.” ↗

And the line that ties FSD straight to his general-AI thinking, that solving driving means solving real-world AI, the thesis behind Tesla-as-an-AI-company:

“A major part of real-world AI has to be solved to make unsupervised, generalized full self-driving work, as the entire road system is designed for biological neural nets with optical imagers” ↗

“Tesla is solving a major real-world AI problem.” ↗

This is a clean evolution-of-views datapoint: the same biological-vision-in-silicon logic he gives Lex Fridman that same year, here in his own dated tweets at the exact moment he decided to pull the radar. See Elon Musk Tweets 2021-2022.