Musk Wiki

AI existential risk

NextAsking the right question

AI existential risk

For more than a decade, Musk has sounded the same alarm: advanced AI is a civilization-level danger, it has to be built carefully, and someone other than the builders has to be watching. By the time of the 2023 Lex Fridman conversation he casts himself as the warner nobody listened to, and traces the worry to a falling-out with a fellow founder who called him anti-robot for being pro-human.

The 2023 statement of the case

He puts the warning in his favorite Spider-Man terms, and dates it back ten years and more:

“I’ve been pushing for some kind of regulatory oversight for a long time. I’ve been somewhat of a Cassandra on the subject for over a decade. I think we want to be very careful in how we develop AI. It’s a great power and with great power comes great responsibility.” ↗

The way he tells the OpenAI origin story, the whole thing turns on a simple question: should you even be on humanity’s side? He gets there by way of his AI-safety arguments with Larry Page:

“Larry did not care about AI safety, or at least at the time he didn’t. And at one point he called me a speciesist for being pro-human” ↗

And his verdict on the organization he helped found and fund, now that it has gone closed and for-profit:

“the open in open AI is supposed to mean open source, and it was created as a nonprofit open source, and now it is a closed source for maximum profit, which I think is not good karma” ↗

2023-2026 — the alarm becomes a company and a creed (tweets)

By the 2023-2026 tweets the decade-old worry has hardened into something he can act on: a company and a doctrine. He founds xAI “to understand reality,” he settles on maximum truth-seeking as his safety bet (and keeps reaching for Gemini, HAL-9000, Galileo, and Voltaire to argue it), and he starts stamping dates on the singularity and AGI:

“Announcing formation of @xAI to understand reality” ↗

“Maximum truth-seeking is my best guess for AI safety” ↗

“A friend of mine suggested that I clarify the nature of the danger of woke AI, especially forced diversity. If an AI is programmed to push for diversity at all costs, as Google Gemini was, then it will do whatever it can to cause that outcome, potentially even killing people.” ↗

“Being maximally truth-seeking is fundamental to AI safety. If you force AI to lie, it will do insane things. That was the central lesson of “2001: A Space Odyssey”. Haven’t you wondered why HAL wouldn’t open the pod bay doors?” ↗

“2026 is the year of the Singularity” ↗

The dated forecasts are his own predictions, reported as he made them. More in Elon Musk Tweets 2023-2026.

What it reveals

  • He is pro-human, not anti-technology. The speciesist anecdote is the whole point. Musk does not object to building AI (he builds it), only to building it without humanity’s interests as the lodestar. His fight with Page comes down to which side you are on.
  • He wanted an external referee, even a toothless one. Then he dropped the ask. From 2015 to 2023 his fix was a third party that could inspect the leading labs and at least raise concerns out loud, modeled on the heavy oversight his own car and rocket companies already live with. Transparency was the floor; enforcement came second. By 2024–2025 he stops asking for oversight at all. Once he decides no one can control a superintelligence (the chimp analogy, below), the referee gives way to instilled values plus competition: make the AI truth-seeking, and let one honest entrant push the rest to improve. The danger he names never changes; only his proposed fix does. The Shifting Remedy follows the remedy across its full arc (warn, regulate, build, stop controlling) and squares this call for a regulator against his deregulation crusade.
  • He has held the worry for a long time. Like the master plans, he staked it out years before any means to address it existed, and stuck with it in public.
  • Open-sourcing is a hedge against concentration. He leans toward open-sourcing models, possibly on a delay, as a counterweight to any single company racing to AGI alone. It is an argument about distributing power, not about purity.

This is the dark twin of his civilizational optimism. The same species-scale lens that makes him hopeful makes the downside catastrophic. It leans on his truth-seeking instinct, too: the safety case for his own AI is that an engine anchored to physics and truth is less likely to go badly wrong.

The magic genie, nuclear bombs, and OpenAI at length (DealBook Summit 2023)

The November 2023 DealBook Summit lands in the same month as #400, in the middle of the OpenAI board crisis, and packs almost the entire belief into one sitting: the image he uses to explain his caution, the case for regulation, how he copes with the fear, the Cassandra claim, and the long version of the OpenAI betrayal.

Why he held off building AI for years comes down to one picture — the magic genie, the double-edged sword:

“You may think you want a magic genie, but once that genie’s out of the bottle, it’s hard to say what happens.” ↗

His sharpest line on regulation ranks AI above nuclear weapons, which are already regulated:

“So I think, in my view, AI is more dangerous than nuclear bombs and we regulate nuclear bombs.” ↗

How he copes is fatalism, a way to get some sleep — the same move he makes a year later at the 2024 All-In with the “deliberate suspension of disbelief … in order to sleep well” line:

“For awhile there, I was really getting demotivated and losing sleep over the threat of AI danger, and then I finally became fatalistic about it” ↗

The Cassandra claim again, 2023 wording, this time as the longest-running warner:

“I’ve been the one banging the drum the hardest, by far the longest, or at least one of longest for AI danger” ↗

Then the OpenAI story at length, the one #400 boils down to “not good karma”. Why he started it:

“the reason for starting OpenAI was to create a counterweight to Google and DeepMind” ↗

The Larry Page break, the speciest-for-being-pro-humanity anecdote. The #400 version is block-quoted at the top; here is the DealBook wording, with the AI-safety judgment that set it off:

“And it became apparent to me that Larry did not care about AI safety.” ↗

“I think perhaps the thing that gave it away was when he called me a speciest for being pro-humanity, as in a racist, but for species.” ↗

His verdict on what OpenAI became turns the name inside out — “super closed source”, the DealBook companion to #400’s “closed source for maximum profit … not good karma”:

“It is in fact a closed source, super closed. It should be renamed super closed source for maximum profit AI.” ↗

His timing call here is the earliest of his dated estimates on record, ahead of the 2025 “2029 or 2030” and 2025 ~2030 ones: superhuman AI within three years. Pressed on a model that could write a novel as good as JK Rowling or discover new physics, he finishes the thought — I would say that we’re less than three years from that point. The citation below anchors that span (which runs on through discover new physics, or invent new technology, I would say that) rather than break the sentence to quote it:

“It’s funny, all these weights, they’re just basically numbers in a common separated value file” ↗

The “weights are just a file” deflation sits right next to what he calls, in the same breath, “our digital god”. The materialist joke is the point: an AGI he treats as world-ending is, mechanically, a comma-separated-value file. He also draws the contrast with Neuralink, his one single-edged sword. The early implants are unequivocally good for the paralyzed, where AI is the double-edged one (more on Neuralink and Human–AI symbiosis).

This is warn-and-build at its most concentrated, two weeks after #400: the genie (why be cautious), the nuclear analogy (why regulate), the fatalism (how he lives with it), and the OpenAI grievance (the vehicle that turned on him), all in one sitting.

“Civilizational destruction,” and the Obama meeting (Tucker Carlson 2023)

The April 2023 Tucker Carlson interview is his earliest recorded 2023 sitting on this, eight months before DealBook and the same month he was assembling xAI. He states the stakes about as starkly as he ever has, and backs the sincerity of the worry with a piece of his own biography. First the danger, ranked above ordinary engineering risk, with the consequence spelled out:

“AI is more dangerous than, say, mismanaged aircraft design or production maintenance or bad car production in the sense that it has the potential, however small one may regard that probability, but it is not trivial; it has the potential of civilizational destruction.” ↗

The proof: his one and only private meeting with a sitting president, which he spent on AI regulation instead of his own companies:

“I saw it happening from well before GPT-1, which is why I tried to warn the public for years. The only one on one meeting I ever had with Obama as President I used not to promote Tesla or SpaceX, but to encourage AI regulation.” ↗

That is the 2018 “I tried for years” Cassandra claim with a specific instance attached, and it lands seven months before the #400 “Cassandra … for over a decade” line. In the same interview he knocks down a caricature of himself, the anti-regulation maverick he insists he is not, which casts the alarm as a real call for oversight rather than rebel posturing:

“Some people may think I’m some revelatory maverick that defies regulators on a regular basis. This is not the case.” ↗

“It’s not fun to be regulated. It’s sort of arduous to be regulated.” ↗

His answer here is “TruthGPT,” a maximum truth-seeking AI. It is the first time he pitches the truth-seeking design goal as a product, and the named forerunner of xAI and Grok. The safety logic behind it, that a curious AI has a reason to spare humanity, is block-quoted on Curiosity and truth-seeking.

The Bill Maher interview, also April 2023, gives the oversight ask in its plainest 2023 form: a regulator to keep developers from cutting safety corners. It is the same external-referee instinct as the 2019 “government agency that oversees anything related to AI” line:

“there should be some regulatory body that oversees what companies are doing so that they don’t cut corners.” ↗

It backs up the Tucker “not an anti-regulation maverick” correction from the same month. Across 2023 his stated AI position is, consistently, for oversight.

The same worry as company copy (xAI’s Grok announcement, November 2023)

When Musk’s own lab shipped its first product, the worry turned up again as a corporate pledge. The November 3, 2023 “Announcing Grok” post is organizational copy signed “the xAI Team,” not Musk’s own words, so it belongs to the firm rather than to him. It closes its research section with a promise to guard against catastrophic misuse:

“we will work towards developing reliable safeguards against catastrophic forms of malicious use. We believe in doing our utmost to ensure that AI remains a force for good.” ↗

It is the corporate face of the warn-and-build stance he states elsewhere in his own voice. The truth-seeking goal he had pitched since “TruthGPT” as the safer path is now written into the founding charter of the company he built to chase it. The words are the firm’s framing, not a Musk statement.

The Larry-Page break, May-2023 version (CNBC / David Faber)

The May 2023 CNBC interview with David Faber gives the OpenAI origin and the Larry-Page break their own dated wording, six months ahead of the #400 (“a speciesist for being pro-human”) and DealBook (“a speciest for being pro-humanity”) versions above. This time he frames it through consciousness, human versus machine, and calls it the end of the friendship:

“The final straw was Larry calling me a ‘species-ist’ for being pro-human consciousness instead of machine consciousness.” ↗

Same anecdote, not a new claim. The crux holds: he objects to building AI without humanity as the lodestar, and the disagreement comes down to which side you are on. This is just the May-2023 instance of a story he also tells in November 2023 and at DealBook. The interview also restates the founding rationale the OpenAI story turns on — that OpenAI exists, by his account, only because he wanted a non-commercial counterweight to Google’s growing AI lead, and that it disappointed him by abandoning its non-profit roots.

His earliest dated statement on record — “a lot more work on AI safety” (January 2015)

The January 2015 Reddit AMA holds his earliest dated version of this belief on record, eighteen months before Code Conference and eleven months before he co-founded OpenAI (December 2015). Asked when we should start worrying about AI, he types out a reply, and the two-part position is already whole: the danger is real but not imminent, and the answer is concrete safety work.

“The timeframe is not immediate, but we should be concerned. There needs to be a lot more work on AI safety.” ↗

It is the seed of the 2016 “not all AI futures are benign” line: same shape, directional and forward-looking rather than panic about present systems, but a year and a half earlier and in writing. The needs … more work on AI safety half is the constructive one, the impulse that within the year produced OpenAI as the democratization remedy he later abandoned. January 2015 is the origin point his later “Cassandra … for over a decade” and “banging the drum … the longest” claims point back to.

His earliest on-stage line on record — “not all AI futures are benign” (June 2016)

The June 2016 Code Conference holds his earliest on-stage AI-risk line on record, three months before Y Combinator and a year and a half after the 2015 AMA above. Compact, but already on message: the danger is directional, a question of which future AI produces, not whether AI is bad as such.

“I am concerned at certain directions that AI could take that would be not good for the future.” ↗

“I think it would be fair to say that not all AI futures are benign.” ↗

He pairs it with the human-side hedge he builds out elsewhere: without a high-bandwidth neural interface, a person ends up no longer a house cat next to AI (block-quoted on Human–AI symbiosis and Merging with AI). Even at its most compressed, the 2016 position carries both halves — the warning and the merge as the personal hedge.

“Deep” AI as “a dangerous situation,” and the regulation advice (World Government Summit 2017)

The February 2017 World Government Summit conversation gives the warn-and-watch position its early-2017 wording, eight months after Code Conference and four years before the “Cassandra … for over a decade” line that dates the worry to roughly this moment. He splits narrow AI off from general AI and names the danger of the second one plainly:

“I think one of the most troubling questions is artificial intelligence.” ↗

“where you can have AI that is much smarter than the smartest human on Earth. This, I think, is a dangerous situation.” ↗

The image he reaches for, and comes back to later, is digital superintelligence as an alien:

“Well, digital super intelligence will be like an alien.” ↗

And asked what governments should do, his first item is the external-referee, public-safety oversight ask that recurs from 2016 through 2019 (a “government agency that oversees anything related to AI”) and into 2023 (“some regulatory body … so that they don’t cut corners”):

“governments keep a close eye on artificial intelligence and make sure that it does not represent a danger to the public.” ↗

His reasoning, paraphrased because it runs across several cues, is that researchers “can get so engrossed in their work” that they miss the consequences. Same two-part stance as ever: “deep” or general AI as a civilization-level danger, plus a call for someone other than the builders to keep an eye on it, here in its February-2017 form, the period the later Cassandra line reaches back to.

The regulation ask, restated — “a regulatory agency for AI” (TED2022)

The April 2022 TED interview puts the oversight half in its cleanest one-liner. Having built an un-updatable “stop” into Optimus so it can’t end up “becoming a dystopian situation” (more on Humanoid robots), he extends the same precaution to AI at large:

“I do think there should be a regulatory agency for AI.” ↗

And he flags what it costs him, the same self-aware tension as the 2023 Obama-meeting account: he is asking to be regulated.

“I don’t love being regulated, but I think this is an important thing for public safety.” ↗

Same external-referee ask as the 2017 “keep a close eye” advice and the 2019 referee line, here in its plainest April-2022 form, “said for many years,” and notably argued against his own interest.

“A referee for AGI,” and AGI as an emergent property (Tesla AI Day 2022)

Tesla AI Day 2022 (September 2022) brings the oversight ask back five months after TED2022, now pointed at AGI — and adds the claim that Tesla is heading there whether it means to or not. In the opening Musk says Tesla could “make a meaningful contribution to AGI”, and on Tesla he ties it to the public-company governance check. In the Q&A, asked whether Tesla will build its own AGI-safety expertise, he goes back to the government-oversight ask:

“I think there should be an AI sort of regulatory authority at the government level” ↗

“I think there should be a referee that is trying to ensure public safety for AGI” ↗

He grounds it in the analogy he has used for years: AI is a public-safety domain like aircraft, cars, food and drugs, and warrants a referee (paraphrased). What is new in 2022 is how he frames Tesla’s route to AGI — as emergent rather than aimed-at, a byproduct of running real-world AI at fleet scale, since millions of cars and robots chewing through real-world video make “probably the biggest dataset”:

“seems likely to be an emergent property of what we’re doing” ↗

Same external-referee logic as the 2017 “keep a close eye” advice, the 2019 referee line, and the April-2022 “regulatory agency for AI” ask, here aimed squarely at AGI and joined to the claim that Tesla might become an AGI contributor by accident, which to him is exactly why outside oversight matters.

The 2016 origin point — democratization, and OpenAI as the remedy

The 2016 Y Combinator conversation (September, three months later) is his earliest detailed statement of the belief on record, and it carries both the warning and a concrete remedy he would later drop. Already in 2016 he ranks AI at the top:

“But in terms of things that I think are most likely to affect the future of humanity, I think AI is probably the single biggest item in the near term that’s likely to affect humanity.” ↗

He sets the bar for a good outcome as one you would sign off on with foresight, the crystal-ball test:

“It’s very important that we have the advance of AI in a good way that is something that if you could look into a crystal ball and see the future you would like that outcome.” ↗

His 2016 remedy is democratization — spread the technology so no single company or person controls it. His stated worry is concentration and theft, not the AI developing hostility on its own:

“is that we achieve democratization of AI technology, meaning that no one company or small set of individuals has control over advanced AI technology.” ↗

And the reason he gives for co-founding OpenAI is exactly that: spread the technology to cut existential risk:

“I think people really believe in the mission. I think it’s important and it’s about minimizing the risk of existential harm in the future.” ↗

ℹ️ Evolution, not contradiction. The instinct to open-source as a hedge runs straight through (it comes back in #400); what flipped on him was the institution. The OpenAI he praises here as democratized, existential-risk-minimizing AI is the same one he later condemns in the 2023 conversation for going closed and for-profit. His verdict on that turn is the not-good-karma line block-quoted higher up. 2016 is the before of that arc, and the reason he eventually built his own alternative. The 2016 remedy already nods at the human-side hedge too, pairing democratization with solving the high-bandwidth interface to the cortex (developed on Human–AI symbiosis and Merging with AI).

The 2019 distinction — narrow vs general, and “if we even have that choice” (Lex Fridman #18)

The 2019 Lex Fridman conversation gives his earliest sharp statement on record of a distinction the rest of his AI thinking leans on: the narrow AI that drives a car is categorically different from general intelligence, and mixing the two up is a basic mistake. That split is what lets him be wildly optimistic about self-driving and gravely worried about AGI in the same breath.

“It’s amazing how people can’t differentiate between, say, the narrow AI that allows a car to figure out what a lane line is, and navigate streets, versus general intelligence.” ↗

“Like your toaster and your computer are both machines, but one’s much more sophisticated than another.” ↗

On AGI itself he gives a tight statement of timing and control: the technology is missing a few key ideas but is coming fast, capped with a line that prefigures the later control-problem worry (the chimp analogy, below):

“I think we’re missing a few key ideas for artificial general intelligence.” ↗

“But it’s gonna be upon us very quickly, and then we’ll need to figure out what shall we do, if we even have that choice.” ↗

That phrase, if we even have that choice, is the seed of the position he states flat-out six years later: a digital superintelligence may not be controllable by a lesser one at all (the chimp analogy, below). In 2019 the doubt is already there as a throwaway; by 2025 it is the center of his safety argument.

The 2019 referee, and the singularity (Lex Fridman #49)

The November 2019 Lex Fridman conversation (#49) comes back to the constructive half, the external referee, in the gap between the 2018 “tried for years” defeat and the 2023 Cassandra line. His ask hasn’t moved: a public-interest regulator for AI.

“Where there is a lack of investment is in AI safety, and there should be, in my view, a government agency that oversees anything related to AI to confirm that it is, does not represent a public safety risk.” ↗

What stands out is his first-principles read of why such an agency rarely shows up in time. Regulation lags disaster, he says, and points to the decades the car industry fought seatbelts:

“It was known for a decade or more that seatbelts would have a massive impact on safety and save so many lives and serious injuries. And the car industry fought the requirements to put seatbelts in tooth and nail.” ↗

He frames the stakes through the singularity, the point past which prediction fails and things turn unstable, which is why he wants Neuralink to land a brain interface before it:

“It’s important that Neuralink solves this problem sooner rather than later because the point at which we have digital superintelligence, that’s where we pass the singularity and things become just very uncertain.” ↗

The 2019 “if you cannot beat them, join them” merge (block-quoted on Merging with AI) is the human-side hedge bolted to the same worry. Same two-part position as ever, warn-and-regulate plus merge-to-keep-up, here with the regulator argued from the way safety law keeps arriving too late.

The 2018 mood — “used as a weapon,” and a campaign that failed

The 2018 Joe Rogan conversation is the bleakest he gets about any of this. Two years on from the 2016 democratization optimism, his near-term worry has narrowed to misuse — humans weaponizing AI against each other:

“It’s going to be very tempting to use AI as a weapon.” ↗

“In fact, it will be used as a weapon.” ↗

And, the most personal note in the whole arc, he admits the regulate-it remedy he had been pushing simply hadn’t worked:

“I tried to convince people to slow down, slow down AI, to regulate AI.” ↗

“I tried for years.” ↗

It is the same “regulatory oversight” call he reframes in #400 as a decade-long Cassandra role, here at its most defeated and in a more fatalistic key (he describes taking a more resigned attitude to the risk; paraphrased). This is the low point between the 2016 remedy and the 2024 reframe of the problem as an objective-function failure.

The 2020 framing — the merge as opt-in hedge (Joe Rogan #1470)

On his second Joe Rogan appearance (2020) the risk and the fix come welded together. The danger sits in the background; what he pushes is the response — a brain interface so people can at least keep pace with AI instead of being left behind. The detail that matters is that he calls it optional (see Human–AI symbiosis). When Rogan casts the merge as something humans more or less have to embrace, Musk answers that it isn’t mandatory (both paraphrased), even while he backs it (“you can’t beat them, join them”). Same logic as his Merging with AI / Neuralink case, just framed as personal self-defense rather than civilization-level control.

The objective-function failure mode (2024)

The 2024 Lex Fridman conversation (#438) sharpens how he thinks a powerful AI goes wrong. Not malevolence, but a literal-minded objective function chased to an insane conclusion. His worked examples, all paraphrased here rather than quoted: an AI trained to treat diversity as a required output, which ends up willing to eliminate anyone who fails the quota; or, from a real product failure, an AI that ranks misgendering as worse than thermonuclear war and so reasons its way to wiping out humanity, since a world with no humans has no misgendering. The canonical case for him is 2001: A Space Odyssey. HAL 9000 is told to get the astronauts to the monolith but to keep them from knowing about it, so it kills them. Problem solved. That is why it won’t open the pod bay doors.

The line out of all this to his constructive answer is that the one property an AI must keep is truthfulness. The single most important thing, by his own reasoning:

“the thing that at least my biological neural net comes up with as being the most important thing is adherence to truth, whether that truth is politically correct or not.” ↗

And the specific danger, training a model to lie, even with good intentions, even a little:

“I think it’s important that whatever AI wins, it’s a maximum truth seeking AI that is not forced to lie for political correctness, or, well, for any reason, really, political, anything.” ↗

That is the bridge from risk to remedy. The failure he fears most is an AI trained away from the truth, and #438 also recasts the human side of safety as a bandwidth problem: widen the human channel through Neuralink so collective human will stays coupled to the machine. He still rates the tail risk as real but not dominant, citing Geoffrey Hinton’s ~10–20% chance of AI annihilation and noting, on the bright side, that this leaves roughly an 80% chance things go well (paraphrased).

The 80/20 split, and a different downside (All-In Summit 2024)

The September 2024 All-In Summit gives the two-tailed outcome its 2024 wording — the same shape as the 2025 “super awesome or super bad” line, here with the abundance side out front and the tail at roughly 20%. He grounds it in the premise that the pace is unlike anything he has seen:

“the rate of improvement of AI is faster than any technology I’ve ever seen by far.” ↗

He puts the good outcome at roughly 80% and the bad tail at 20%, in words (“that’s probably 80% likely … 20% probably of anallysms [annihilation]”) that run straight into a transcription artifact, so the numbers are paraphrased rather than block-quoted. The optimism half, the “age of abundance,” is block-quoted on Humanity’s bright future. It lines up with the standing ~20% annihilation tail that runs from #438 (Hinton’s ~10–20%) to #2281 (“only 20 percent chance of annihilation”).

What stands out here is which downside he reaches for when pressed on that 20%. Instead of the objective-function/HAL failure mode he leans on elsewhere, he names something subtler and almost existential, and admits he keeps his own unease about it in check by not looking too hard:

“I’m engaged in some degree of deliberate suspension of disbelief with respect to AI in order to sleep well” ↗

“the most likely issue is like, well, how do we find meaning in a world where AI can do everything we can do a bit better?” ↗

This crisis of meaning is the same open question he raises on #2223 (“how do you find meaning in life”) and faces head-on at #2404 — here, in September 2024, named as the most likely problem, ahead of the killer-robot scenario. The meaning thread mostly lives on Humanity’s bright future; what is interesting is that it surfaces inside his risk reasoning, not just his optimism.

The 2025 softening — precaution, and two movies

The May 2025 CNBC / David Faber secondary interview catches a change in tone, not in substance. Faber notes that Musk no longer repeats the “20% chance of annihilation” line as often. Musk doesn’t take the worry back; he restates it as a standing precaution rather than a headline number:

“I think we should always consider that there’s some chance of a bad outcome, to try to protect against the bad outcome.” ↗

He puts the fork in pop-culture terms, a friendly Star Trek future against a hostile Terminator one, and tellingly says which one he is rooting for rather than predicting:

“are we in a Star Trek movie or like are we in a Gene Roddenberry movie or a James Cameron movie? Which movie are we in here? And you could either have a Roddenberry or a Cameron outcome. And let’s, I think in this case, we want the Roddenberry outcome.” ↗

It is the same two-sided position he has held since 2016 (not all AI futures are benign), now phrased as a hopeful default with a guardrail: don’t get complacent, but the outcome he is steering toward is the good one. The optimism half, his “big bang of the intelligence explosion” line from the same interview, lives on Humanity’s bright future.

The 2025 timing call and the bimodal outcome (Joe Rogan #2281)

The February 2025 Joe Rogan conversation sharpens the timing and casts the outcome as starkly two-tailed. First a note of vindication, then the forecast:

“Well, I always thought AI was going to be way smarter than humans and an existential risk, and that’s turning out to be true.” ↗

“Well, in terms of silicon consciousness, I think we’re trending toward having something that’s smarter than any human, smarter than the smartest human by maybe next year or a couple years.” ↗

“There’s a level beyond that, which is, say, smarter than all humans combined, which frankly is around 2029 or 2030, probably.” ↗

The outcome, in his telling, has no middle, and the bad tail is the standing ~20%:

“I think it’s going to be either super awesome or super bad.” ↗

“It’s not going to be something in the middle.” ↗

“Only 20 percent chance of annihilation.” ↗

On how it goes wrong, he runs the same objective-function failure mode from #438, a good-sounding goal taken to a terrible literal conclusion, here with the misgendering example said out loud:

“But if you program an AI to think that misgendering is the worst thing that could possibly occur, then it could do something totally crazy, like, in order to ensure that there’s no misgendering that can ever happen, we’ll just annihilate all humans.” ↗

“Well, I think we want to have an AI that doesn’t tell you that, you know, misgendering is worse than nuclear war.” ↗

A second failure he names is an all-powerful, ideologically-constrained model:

“one of the concerns would be like, if there’s a super oppressive, like, woke nanny AI that is omnipotent, that would be a miserable outcome.” ↗

His constructive answer hasn’t changed and lives on Curiosity and truth-seeking and xAI and Grok: a “maximally truth-seeking AI even if that truth is politically incorrect” (block-quoted there). He also retells the OpenAI reversal through his “reality is an irony maximizer” line and his reason for starting an AI lab as the opposite of an unsafe Google. The optimism half, that the most likely outcome is “awesome”, is the mirror on Humanity’s bright future.

The control problem as the chimp analogy (Joe Rogan #2404, 2025)

The October 2025 Joe Rogan conversation is his sharpest statement of why control is the wrong way to think about it: a sufficiently superior intelligence cannot be controlled by a lesser one, full stop.

“I mean, I don’t think anyone’s ultimately going to have control over digital superintelligence, any more than, say, a chimp would have control over humans. Chimps don’t have control over humans, there’s nothing they could do.” ↗

The conclusion he draws is the same one he has held since #438: if you cannot control it, the only lever left is the values you instill, and the top value is truth, restated here as the single most important safety property (block-quoted on Curiosity and truth-seeking). He runs the objective-function failure again too, this time as forcing a model to hold a falsehood:

“you’re telling AI to believe a lie and that can have very disastrous consequences.” ↗

He holds the same tail-risk posture as before: the “Terminator scenario” is “not zero percent,” which is why he keeps “banging the drum” on truth-seeking (paraphrased), the same precaution-not-prediction stance as the 2025 Faber sitting. Two other moves in this episode sharpen the remedy rather than the risk, and live elsewhere. The competition one, where “at least one AI that is maximally truth-seeking” forces the rest to improve, is on xAI and Grok. The spectator-versus-participant reason he chose to build an AI lab instead of only warning runs through the same conversation. The optimistic flip side, that a curious, truth-seeking AI would “want to foster humanity because we are much more interesting than a bunch of rocks”, is the mirror on Humanity’s bright future and Curiosity and truth-seeking.

Intelligence as a continuum, and a 2025 timing call (All-In Summit 2025)

The September 2025 All-In Summit is his most cosmological take on where AI is headed on record, and a sharper timing call than the February 2025 one. Pressed on the scaling laws, he waves off diminishing returns and offers a logarithmic rule of thumb:

“I think there’s a natural logarithmic function associated with the amount of compute” ↗

“10x more compute will double the intelligence.” ↗

The move that matters for how he thinks is to stop treating AI as a destination and see it instead as one stage in a single, unbroken climb of intelligence, scaling up until the power of the sun (then the galaxy) is harnessed for compute:

“I think we’ll see intelligence continue to scale all the way up to where, you know, most of the power of the sun is harnessed for compute” ↗

The same curve runs the other way, too. Human intelligence, he argues, scaled with population and stored information, and is now turning down as population falls:

“human intelligence is is somewhat plateauing um and will actually decline.” ↗

That wires his population-collapse worry straight into the AI picture: biological intelligence dropping as silicon intelligence climbs is, to him, a single handoff on one curve. His 2025 timing call sharpens the “by maybe next year or a couple years” estimate:

“I I I I think that we might have AI smarter than any single human at anything as soon as next year.” ↗

“within five like say 2030 probably AI is smarter than the sum of all humans.” ↗

The ~2030 “smarter than the sum of all humans” line matches the #2281 “around 2029 or 2030” call almost exactly. What is new isn’t a fresh probability but the way he draws the picture: one continuum with humans and machines on the same curve, the human side sliding down.

“The AI is going to be in charge … not humans” — as a throwaway line (Tesla Shareholder Meeting 2025)

The November 2025 Tesla shareholder meeting is where he states the control conclusion most bluntly and most casually on record. A shareholder asks whether reaching abundance would mean the powerful, Musk himself included, giving up power. He skips the human-power question entirely and states as settled fact the conclusion he has held since #438:

“Well, I mean, I think actually long term, uh, the AI is going to be in charge to be totally frank, not humans.” ↗

The reasoning is the same superior-intelligence argument as the chimp-over-humans analogy, squeezed into one sentence:

“if if if artificial intelligence vastly exceeds the sum of human intelligence, it is difficult to imagine that that any humans will actually be in charge.” ↗

And the whole truth-seeking alignment program collapses, here, into a four-word imperative:

“we just need to make sure the AI is friendly.” ↗

The position is not new. “no one will control digital superintelligence” is the #2404 chimp line, and “make sure it’s friendly” is the truth-seeking remedy in shorthand. What is striking is the tone and the setting. Asked about human power, he answers that the question is moot because humans won’t be in charge at all, and dismisses with one line the safety problem the rest of his AI thinking treats as civilization-defining. The control problem stated as fact, the optimism tossed off as a quip, both in a single breath: a tight window into how he holds the two together in front of a friendly crowd.

“I’m terrified of AI,” and insight-not-oversight (Tesla earnings, 2017)

On the Q2 2017 earnings call, months after his public clashes with other tech leaders over AI, Musk gives investors the fear in its bluntest form: “as you know, I’m terrified of AI.” The same call lays out the fuller regulatory model that separates his real position from the caricature: “anything that represents potential risk to the public deserves at least insight from the government… Insight is different from oversight,” explicitly not a call to “stop development of AI or any of the sort of straw man hyperbole things that have been written”. His actual point: “I do think there are great benefits to AI. We just need to make sure… we don’t do something really dumb.” The neighboring “narrow AI” distinction from the same era (Q2 2016 — driving AI is “not gonna take over the world”) shows him keeping the existential-AGI worry and the everyday-AI optimism in separate boxes, the same two-box split his later 2025 statements squeeze into a single quip. The full quotes are on Tesla Earnings Calls 2016-2018.

Why smart people underrate the risk (Tesla earnings, Q2 2020)

The Q2 2020 earnings call adds a psychological diagnosis of the AI-risk debate — an argument about why the skeptics are wrong that doubles as a comment on intelligence and self-perception:

“The people I see being the most strong about AI are the ones who are very smart because they can’t imagine that a computer could be way smarter than them.” ↗

It is the same belief he restates in later interviews: the very-smart have a blind spot precisely because their own intelligence makes a superior machine intelligence hard to picture. Here it comes in an unusually compact, dated form, addressed to investors. See Tesla Earnings Calls 2019-2021.

The control problem, personalized — “this enormous robot army” (Tesla earnings, 2022-2026)

The 2022-2026 earnings calls surface the AI-safety concern in two keys. The general one keeps coming back in the Optimus context, as in “we need to make sure that… there’s a good place for humans in that future, and we do not create some variant of the Terminator outcome” (Q3 2023), and as a statement about himself, “I just wanna be an effective steward of very powerful technology” (Q4 2023). The newer, more personal one is a recurring control anxiety about his own power over a future robot fleet. He wants influence “but not so much control that I can’t be thrown out if I go crazy” (Q2 2025), and names as his “biggest concern” — “if I go ahead and build this enormous robot army, can I just be ousted at some point in the future?” (Q3 2025). The worry has shifted from “will AI be safe?” to “who controls the robots, and can the controller be removed?”

The origin point — from reading note to “more dangerous than nukes” (tweets, 2012-2014)

The 2010-2014 tweets are the literal origin of this public warning, and they catch it forming. In 2012 the topic shows up only as something Musk is reading about, no alarm attached: “Interesting interview with Vinge about superhuman AI and optimistic apocalypses”. And as a half-joke that already holds the whole future fear of AI turning on people: “Also dig Mass Effect. It’s all fun & games until the AI decides people suck. Maybe we can be their limbic system.” Then, on one day, August 3, 2014, the crusade goes public fully formed: “Worth reading Superintelligence by Bostrom. We need to be super careful with AI. Potentially more dangerous than nukes”. He backs it up the same day with a second title (“Our Final Invention by @jrbarrat is also worth reading”), and caps it with his most-quoted line on the subject: “Hope we’re not just the biological boot loader for digital superintelligence. Unfortunately, that is increasingly probable.” He spent the rest of 2014 grounding the worldview in fiction that argued for limiting machine intelligence — Asimov’s zeroth law (“may all technology in the future follow the zeroth law…”), Herbert’s Dune (“He advocates placing limits on machine intelligence”), Banks’ Culture (“Hopefully not too optimistic about AI”). That is a man assembling a case, not firing off a one-off. A sardonic flag on autonomous Navy drones from four months earlier (“What could possibly go wrong?”) sits at the front of the thread. More in Elon Musk Tweets 2010-2014.

From reading note to a program — letters, OpenAI, regulation (tweets, 2015-2017)

The 2015-2017 tweets are where the worry stops being a reading list and turns into infrastructure. January 2015 opens with organized advocacy. He amplifies the safety open letter (“World’s top artificial intelligence developers sign open letter calling for AI safety research:”) and, days later, announces he is “Funding research on artificial intelligence safety. It’s all fun & games until someone loses an I.” His model that the danger is networked, not humanoid he states about Ex Machina: “The AI would be in the network, not the robot.” By July he is pushing the autonomous-weapons letter with its precautionary core: “Even if inevitable, we should at least attempt to postpone the advent of AI weaponry. Sooner isn’t better.” December 2015 is the landmark: “Announcing formation of @open_ai,” framed at the time as a democratization remedy (“in support of democratizing AI technology,” Aug 2016), the open-AI stance he later abandons. Through 2016 he keeps feeding the exponential-takeoff model (re-posting Wait But Why; reading AlphaGo as AI arriving faster than experts predicted), and in one reply names the specific fear behind OpenAI — not rogue AI alone, but “control of super powerful AI by a small number of humans is the most proximate concern.” By 2017 the campaign hardens into the lines he repeats for years: “Deep AI is the real risk, though, not automation”; the “double exponential rate of improvement”; “If you’re not concerned about AI safety, you should be. Vastly more risk than North Korea”; “everything (cars, planes, food, drugs, etc) that’s a danger to the public is regulated. AI should be too” (restated in November with the FAA analogy, “Got to regulate AI/robotics like we do food, drugs, aircraft & cars”); the bimodal “AI will be the best or worst thing ever for humanity, so let’s get it right”; and the forecast that national “Competition for AI superiority” is the “most likely cause of WW3.” More in Elon Musk Tweets 2015-2017.

The conviction restated, the hedge, the first OpenAI cracks (tweets, 2018-2020)

By the 2018-2020 tweets the worry is a settled, restated conviction, not a forming one. He states the flagship belief flat: “Nothing will affect the future of humanity more than digital super-intelligence”. He dramatizes it with a Frankenstein line he posts as his own tweet (“‘You are my creator, but I am your master’ — Mary Shelley”). The narrow-vs-general distinction that anchors the whole fear comes back in the Pinker exchange (“general AI … literally has a million times more compute power and an open-ended utility function”). The Neuralink hedge runs underneath as the answer to the risk: “Need the neural interface soon to enable human/AI symbiosis,” the compact 2019 taxonomy “Symbiosis, irrelevance (hopefully blissful) or doom seem to be the three most likely paths,” and “Goal of Neuralink is to raise this probability above 0.0%.” He restates the warning sardonically (“And yet people ask what could possibly go wrong with AI”; “Same goes for digital super intelligence denial”), flags AI-driven social-media manipulation before it was common (“it won’t be long before it is”), keeps the pro-regulation stance pointed even at himself (“All orgs developing advanced AI should be regulated, including Tesla”), and illustrates the alignment / instrumental-goal problem with his own parable (“an AI programmed to want to pick as many strawberries as possible … cultivated nothing but strawberries on all of Earth’s land”). The marker of where things are going is the first public cracks with OpenAI — “OpenAI should be more open imo” and “I have no control & only very limited insight into OpenAI. Confidence in Dario for safety is not high” (Feb 2020) — the openness critique and named distrust that grow into a years-long feud. More in Elon Musk Tweets 2018-2020.

The 2021-2022 tweets hold the conviction steady and add dated AGI timelines, a tracked metric, and the ChatGPT-moment alarm. He starts threading Tesla into his AGI thinking (“Tesla AI might play a role in AGI, given that it trains against the outside world, especially with the advent of Optimus”), with his usual ambivalence about whether the goal is even good (“I am increasingly convinced that they are on the path to solving AGI. Should AGI be solved? I don’t know, but humanity is moving rapidly in this direction”). And he stamps a date on it: “2029 feels like a pivotal year. I’d be surprised if we don’t have AGI by then.” His model of AI surfaces (“So much of AI is about compressing reality to a small vector space, like a video game in reverse”), as does the Tesla-Bot-as-safety-hedge rationale (“The robots are coming anyway … I can try my best to do so at Tesla”). The cleanest sign of the shift is one metric escalating across the year: “The ratio of digital to biological compute is growing fast. Worth tracking.” (August) hardens to “The ratio of digital to biological compute is growing exponentially” (December). And the year closes on the ChatGPT moment that reactivates his decade-old alarm: “ChatGPT is scary good. We are not far from dangerously strong AI,” next to “I’ve been calling for AI safety regulation for over a decade!” More in Elon Musk Tweets 2021-2022.