Elon Musk: OpenAI Betrayal, His Future at Tesla, and the Next Big Thing — Grokipedia
Summary
Musk’s next X product cycle moves beyond brittle Twitter heuristics with Grok: it will initially read what it judges the top 10 million of roughly 100 million daily posts, then all 100 million, with Musk estimating “on the order of 50K H100s.” The payoff is personalized ranking plus semantic search across text, pictures and video; nearer term, a curated Following tab and bug fixes should restrain the interaction-triggered “food funnel.”
Grokipedia v0.1 is xAI’s bid to make an AI-generated reference product materially broader than Wikipedia, not merely less biased. A “maximally truth-seeking” Grok cycles through Wikipedia’s million most popular articles to “add, modify and delete” after researching public sources; Musk predicts version 1.0 will be 10 times better and the full product “a hundred times better,” with generated explanatory video still to come.
Musk’s three-year X transformation is framed as a free-speech intervention whose evidence layer now combines one-tap Grok analysis with Community Notes. Jason said the Twitter Files exposed 80 FBI agents submitting takedown requests; the retrospective also spans restored accounts and aggressive shadow-banning tools. Musk’s current rule is to follow each country’s law, even when cases such as Brazil leave the platform trapped between a judge’s order and the statute itself.
Tesla’s Optimus program is conditional on Musk retaining enough voting control to resist politically motivated removal. He wants approximately 25%—“a strong influence,” but not enough to prevent dismissal if he “go[es] insane”—because passive funds and proxy advisers ISS and Glass Lewis could otherwise threaten his position. Asked whether he would leave if the vote failed, he stopped short of saying so: “I’m not going to build a robot army if I can be easily kicked out by activist investors.”
Musk expects his OpenAI case to reach a jury in February or March and frames it as enforcement of the company’s original open-source, nonprofit bargain. Jason said the California attorney general’s restructuring analysis excluded Musk’s lawsuit, leaving it outstanding. Musk says the incorporation documents barred founders and officers from benefiting financially and argues that, had he created an ordinary corporation after naming, funding and recruiting for it, he would “probably own half the company.” His proposed name for the reversal is “Closed for Maximum Profit AI.”
AI is a “supersonic tsunami,” but Musk sees radically cheaper inference as the longer-run opportunity. A hypothetical fleet of 100 million Teslas, each contributing roughly 1 kilowatt of onboard inference, would offer 100 gigawatts; meanwhile, the human brain’s roughly 10 watts of higher function shows how much more efficient AI compute could become. Tesla and xAI, he said, continue to see “massive improvements” in inference efficiency.
Tesla’s autonomy milestones remain aggressive but expressly safety-gated: Cybercab production starts in Q2 next year, potentially scaling to millions annually, while Austin cars are expected to lose their safety monitors sometime in December. Musk emphasized that every current Tesla is already capable of unsupervised autonomy, yet said the company is “being paranoid” because one accident could become worldwide news; fleet targets were hedged at roughly 1,000 Bay Area cars and 500 in greater Austin by year-end.
Musk puts climate risk on an approximately 50-year horizon—neither five nor 500—and argues that solar economics and manufacturing scale already point toward the destination. China’s roughly 1.5-terawatt annual solar production capacity, against its use of about 1 terawatt per year and average US electricity use of about 0.5 terawatts, implies in his adjusted arithmetic that 18 months of production could cover US demand. His policy preference is subsidy neutrality, while his physical argument is blunter: “Everything is solar power,” with nuclear as an option and terrestrial fusion a tiny project beside the Sun.
Deep dive
1. Grok is moving X beyond its brittle recommendation machinery
A major bug had stopped in-network posts from appearing, meaning users could follow someone and still miss their posts. Separately, the ranking system underweighted passive dwell time but treated likes, replies and other interactions as overwhelming signals, producing “a torrent of that same thing.”
Musk’s diagnosis was excessive gain: “You had a taste of it. We’re going to give you three helpings.” X has mostly fixed that “food funnel,” but removing legacy Twitter heuristics remains treacherous because “the one bug was covering for the other bug,” like a complicated machine supported by one wooden stick.
The Following tab will gain an optional Grok-curated view, selecting the most interesting posts from accounts a user already follows. That addresses the practical problem that a few prolific accounts can overwhelm reverse chronology without eliminating the existing unfiltered tab.
The larger change is content-level understanding. Grok will initially read what it judges the top 10 million of approximately 100 million daily posts—possibly just the English volume—before scaling to the entire stream; Musk guessed this would require “on the order of 50K H100s.” Semantic search will match natural-language queries against text, pictures and video.
2. Grokipedia turns critical reasoning into a reference product
Musk said Grok was trained to be “maximally truth seeking” and to perform cogent analysis: identify an argument’s axioms, test whether they are likely true and mutually consistent, then determine whether the conclusion naturally follows. “We just trained Grok on a lot of critical thinking,” a task he described as quite hard.
That model was instructed to cycle through Wikipedia’s million most popular articles and “add, modify and delete,” researching whatever was publicly available. The goal extends beyond correcting false statements: propaganda can consist of technically true facts weighted or contextualized so badly that they misrepresent a person, event or subject.
Jason said his Grokipedia biography was five or six times longer and more representative than his Wikipedia page. Musk stressed that this is version 0.1, predicted version 1.0 would be 10 times better and ultimately called the product “a hundred times better,” while Jason noted that images and video had not yet been added; Musk said explanatory videos using Grok Imagine were planned.
Asked whether evidence actually changes minds, Musk said tribal allegiance often wins: some people refuse even to watch disconfirming video and become effectively “a flat-earther.” X’s countermeasure is layered—one tap summons deeper Grok research, while Community Notes requires agreement from people who historically disagree and makes both its code and data open source.
3. Twitter’s empty headquarters exposed operating waste at scale
At acquisition, one of Twitter’s two headquarters buildings was completely empty and the other had roughly 5% occupancy. The cafeteria still operated near its prior scale, turning a nominal $20 lunch into an effective $400 lunch because costs had not adjusted to one-twentieth the attendance.
Whiteboard markers in the unused building had dried out after two years without use, even as cleaners serviced an already clean office and fresh tampons continued arriving. The absurdity was not disorder but processes executing indefinitely after their purpose disappeared.
The financial examples went beyond office theater: Twitter paid for SaaS that had never been installed, retained per-desk management software where nobody worked and spent millions on camera-based pedestrian-traffic analysis for an empty building. Musk rated the last example “11 out of 10 on a Dilbert scale.”
4. X broke with the old censorship regime but still obeys local law
Jason’s retrospective identified three layers: prominent account bans, undisclosed shadow bans and government involvement. Figures including Donald Trump, Jordan Peterson, Jay Bhattacharya and Andrew Tate were restored, while the takeover exposed checkbox tools that could quietly deboost accounts despite years of corporate denials.
Musk called “Trust and Safety” an Orwellian label for the group operating those tools. He gave outside reporters direct, unsupervised access to internal records; the resulting Twitter Files and congressional hearings documented emails, Slack messages and database actions, while Jason said 80 FBI agents had submitted takedown requests.
X’s present policy is simply to follow the applicable law, not independently impose extra restrictions. Brazil illustrated the limit: Musk said a judge ordered X to ban accounts contrary to Brazilian law while imposing a gag order, leaving the company instructed to break the law and unable to disclose it; X was consequently blocked there for a period.
Musk sees a global effort to encode speech suppression as “hate speech,” including in Europe and Australia; Jason called the UK and Germany aggressive in this regard. Musk said his understanding was that roughly 2,000–3,000 people were imprisoned in the UK over social posts. His principle was reciprocal: protected speech must include things people hate—“suppress not lest you be suppressed.”
5. Optimus makes Tesla voting control a product-safety issue
Jason opened with proxy-adviser recommendations he considered incoherent: the advisers recommended against Kathleen Wilson-Thompson, an African-American woman director, while also opposing Ira Ehrenpreis for insufficient diversity. Musk’s structural complaint was that passive index funds control about half the stock market and often outsource voting judgments to ISS and Glass Lewis, neither of which owns the underlying shares.
He argued that retirement savers depend on fund managers to maximize shareholder returns, yet activists gravitate toward proxy advisers because that is where voting power resides. Calling the firms “corporate ISIS,” he worried their politics could effectively decide leadership across public companies without corresponding economic ownership.
Musk’s desired approximately 25% Tesla vote is meant to provide strong influence while leaving him removable if he “go[es] insane.” The specific fear is building an “army of robots,” then losing control for political reasons and being unable to ensure their safety. He did not categorically threaten departure, but categorically ruled out building that army while easily removable.
6. The OpenAI lawsuit tests the original nonprofit bargain
Jason noted that the California attorney general’s analysis of OpenAI’s restructuring excluded Musk’s lawsuit, leaving the suit outstanding. Musk believes it will reach a jury in February or March and says there is “a mountain of evidence.” His account is that OpenAI’s incorporation documents expressly created an open-source nonprofit and prohibited any officer or founding member from benefiting financially; archived versions of its website repeated that positioning.
He said he conceived and named the company, provided its A, B and C rounds of funding, recruited critical personnel and shared everything he knew. Had he chosen a commercial corporation, he estimated he would “probably own half the company,” but said the nonprofit structure was a deliberate gift “for the world.”
The founding rationale was AI safety and competitive balance. Google then possessed essentially all the AI, computing and money, while Larry Page, in Musk’s telling, dismissed his concern that AI might destroy humanity by calling him a “speciesist” for favoring human over machine intelligence. OpenAI was built as Google’s open-source, nonprofit opposite.
Musk now calls the result “closed for maximum profit AI” and “tighter than Fort Knox.” He said China generally supplies today’s best open-source models, praised Grok 2.5 as very good and dismissed OpenAI’s releases as broken “fig leaves” nobody runs. On Sam Altman’s stated $1.4 trillion data-center ambition, Musk’s answer was concise: “I think he means it.”
7. AI’s power curve matters more than today’s job forecast
Asked about job displacement and social mitigation, Musk supplied a metaphor rather than a policy timetable: AI is a “supersonic tsunami,” a giant wall of water moving faster than sound. He said it is arriving whether he wants it or not and reiterated that he had previously tried to slow AI development.
One unconventional inference asset is Tesla’s installed fleet. If it eventually reaches 100 million vehicles, each with approximately 1 kilowatt of inference compute, built-in power, cooling and Wi-Fi, Musk calculates that idle cars could collectively provide 100 gigawatts of distributed inference.
Jason’s power-efficiency question cited architectures that might reduce energy per token by three or four orders of magnitude. Musk answered with biology: the brain uses roughly 20 watts, only about 10 watts for higher cognition, yet that “10-watt meat computer” produced aircraft, lasers, the internet, general relativity and quantum mechanics.
Humans also tolerate a roughly 20-year “boot sequence” and communicate at perhaps 10 bits per second, not terabytes. Even a 100-megawatt or gigawatt AI computer cannot yet do everything a human can, but that comparison is Musk’s evidence that enormous efficiency headroom remains; Tesla and xAI are already seeing “massive improvements.”
8. Every Tesla is a robotaxi candidate, while Cybercab removes the controls
Musk rejected the idea that autonomy belongs only to a specialized vehicle. Model 3, Model Y and other current Teslas already carry the AI computer, software and correctly positioned cameras required for unsupervised autonomy; the sensors look ordinary because Tesla hid them around the mirror, repeaters, B-pillars and license plate.
Cybercab is the dedicated autonomous product, with no steering wheel or pedals because Musk considers both “vestigial in an autonomous world.” Production is scheduled to begin in Q2 next year, followed by a ramp he expects eventually to reach millions of units annually.
Jason pressed for a consumer version with a steering wheel and said his Model Y Juniper running 14.1 in Mad Max mode felt close to autonomy. Musk’s rebuttal was behavioral: passengers in Uber or Lyft virtually never wish they could stop using their phones and take over from the driver—“the reality is that they don’t.”
Musk nonetheless emphasized caution. The cars are already “extremely capable,” but Tesla is “being paranoid” because one accident could produce worldwide headlines, especially for Tesla. The company expects Austin vehicles to operate without a seated safety monitor before year-end, “sometime in December,” but presented that as an expectation, not a guarantee.
9. Robotaxi scaling is now a fleet-software problem
Musk said the Austin rollout had gone smoothly; much of the learning concerns operating a fleet rather than basic driving. Tesla must recreate the ride-hailing and dispatch stack of Uber, except the summoned car has no driver, while coordinating charging, routing, positioning and variable demand.
He estimated Tesla would probably have at least 1,000 cars in the Bay Area and perhaps 500 or more across greater Austin by year-end. Edge cases include preventing every car from choosing the same Supercharger, deciding where idle vehicles wait, reading faded disabled-parking markings and avoiding spaces so tight that passengers cannot enter.
Airports introduce localized integrations rather than generalized autonomy. To drop someone at San Jose airport, the vehicle must call the airport’s servers and trigger an approximately $5 fee. Musk’s comic version of the stack was a robot car making a remote procedure call so the airport can charge its credit card.
10. Climate risk is measured in decades, not emergency slogans
Musk used Bill Gates’s skepticism about a long-range Tesla Semi as an example of conclusion-first reasoning. When Musk asked whether Gates disputed battery watt-hours per kilogram or chassis watt-hours per mile, Musk said Gates knew neither figure despite insisting the truck could not work—even though Pepsi was already using it.
On climate, Musk rejected both “it doesn’t exist” and claims that major cities will be underwater in five years. Atmospheric CO2 can be measured with a roughly $50 monitor and is rising about 2–3 parts per million annually; continually moving billions and eventually trillions of tons of underground carbon into the atmosphere and oceans must change their chemistry.
His order-of-magnitude judgment is “at least 50 years” before climate becomes a serious issue: probably not five, but also not 500. The reasonable response is to lean toward solar and batteries while removing subsidies from every industry, including oil-and-gas tax advantages that have existed in some cases for 80 years and that the industry has come to treat as non-subsidies.
11. Solar’s manufacturing scale overwhelms the alternatives
Musk put China’s annual solar-panel production capacity at roughly 1.5 terawatts, versus its use of about 1 terawatt per year and average US electricity consumption of only 0.5 terawatts. After dividing nameplate solar output by roughly five for night and orientation, he calculated that 18 months of Chinese production could supply steady-state capacity equivalent to all US electricity use.
He supports nuclear fission, citing reactors operated safely inside crowded US Navy submarines and aircraft carriers, but noted regulatory and neighborhood resistance. He also thinks a sufficiently enlarged tokamak could achieve net-positive fusion because surface-to-volume ratios improve with scale; beside the free gravitationally contained reactor overhead, however, terrestrial fusion is “a fun science project.”
The Sun is roughly 99.8% of the solar system’s mass, with Jupiter around 0.1%, and produces well over a billion times more energy than everything on Earth combined. Solar flux reaching Earth is about 1 gigawatt per square kilometer; panels near 25–26% efficiency and 80% packing imply roughly 200 megawatts per square kilometer, or a memorable approximation of 1 gigawatt-hour per square kilometer per day.
Musk sees no raw-material barrier to a solar-battery Earth: panels are primarily silicon, while lithium iron phosphate cells use common iron, phosphorus, carbon and lithium. Earth itself is roughly 32% iron and 30% oxygen by mass—“a rusty ball bearing”—and Tesla’s published arithmetic, he said, shows enough material to power the planet completely.