How Based is Grok 3? + Robinhood CEO Vlad Tenev on Markets For Everything + Vibecoding 101
Summary
Grok 3 is competitive with leading models, but xAI has not yet produced a clear product or research moat. At $40 a month through X Premium+, versus $200 for OpenAI’s most powerful plan, it performed adequately on the hosts’ tests and can analyze X posts, yet offered little reason to switch from ChatGPT or Claude. Newton’s test for a lead is novelty: until rivals say, “Oh, we need to do that,” Grok stays “kind of in the middle of the pack.”
xAI’s real signal is its ability to convert capital and access into frontier-scale infrastructure at extraordinary speed. It went from a poor V1 to a capable V3 in roughly one year while assembling Colossus in Memphis with something like 200,000 NVIDIA GPUs—an undertaking costing billions and aided by Musk’s special NVIDIA relationship through Tesla. Unlike DeepSeek’s compute-saving innovations, Grok represents the other route to intelligence: “just build a bigger data center.”
Musk’s simultaneous leadership of an AI contender and influence inside the federal government creates a power thesis whose practical version matters even if the darkest scenarios remain speculative. The hosts entertained, but explicitly labeled remote, a future in which Washington restricts advanced AI, licenses Grok, or nationalizes OpenAI; Roose’s nearer-term concern was simpler and more concrete—expedited data-center permits, new energy sources, and privileged grid access. Musk’s earlier support for a six-month AI pause now looks to Roose like an attempt to buy time to catch up: “I think this is pure power.”
Robinhood, valued at $52.2 billion at recording, wants to become the venue for every financial asset and transaction, with Tenev arguing that tokenization could open private-company exposure to retail investors. Tenev argues crypto looks dominated by meme coins because connecting tokens to productive assets generally makes them regulated securities; a new framework could instead expose companies such as OpenAI and SpaceX. His pitch is to combine the five-minute, globally tradable token with optional disclosures—including audited financials for qualifying companies—and tiered warnings, including a “big red skull and crossbones” for unverified projects.
Prediction markets are Robinhood’s bid to turn information itself into a tradable product, but the sports use case exposed the unresolved boundary with gambling. After Robinhood offered a “pro football championship” contract to roughly 1% of users, the CFTC requested suspension over “serious concerns”; Tenev nevertheless maintained that prediction markets are “the news faster.” His strongest example was the election market reaching 95–5 for Trump before television called the result, while Roose’s LK-99 counterexample showed that markets can confidently price a false story until reporting and replication correct it.
The interview’s central conflict was whether broader access builds wealth or industrializes speculation. Tenev said open markets have historically created wealth and that retail should generally access anything institutions can; the hosts countered with Robinhood alerts for the Trump meme coin, a Dogecoin giveaway, Pump.fun rug pulls, and rising searches related to gambling addiction. He supported suitability controls but resisted paternalism, invoking Massachusetts’s 1980s exclusion of residents from Apple’s IPO—while conceding that mortgage-backed securities and credit-default swaps may be too complex or unnecessary for retail. Casey also disclosed that Robinhood owns Sherwood News, which briefly syndicated some Platformer content the previous year.
Tenev expects AI disruption to make retirement saving more important, while Harmonic is his effort to make model reasoning verifiably correct. His goal is a “super calculator” combining an LLM’s flexibility with the no-hallucination property of a calculator, especially in mathematics, where one faulty step can invalidate the answer. Whatever happens to labor, he remains “very, very confident” money, companies, and investment will persist.
Vibe coding already makes tiny, bespoke software economically rational, but it shifts technical risk from writing code to trusting code one cannot inspect. Roose built podcast and bookmark tools and a trunk-fitting app, then built Casey Newton’s “Hot Tub Time Machine” in about half an hour without writing a line; the governing ethos is Karpathy’s “I just see stuff, say stuff, run stuff, and copy/paste stuff, and it mostly works.” The trade-off is skill atrophy and opaque security: Roose could not know whether the AI inserted malicious code, while Newton argued society still needs a core of engineers who understand systems “down-to-the-metal.”
Deep dive
1. Grok 3 matches the pack without escaping it
xAI’s premium model is available through X’s $40-a-month Premium+ tier, cheaper than OpenAI’s $200 top plan. Both hosts somehow received free access, while Roose noted that ordinary users otherwise have to pay.
Early reactions and xAI’s own benchmarks put Grok 3 roughly alongside leading models, though it had not undergone rigorous independent testing. On Roose’s proprietary “Roose benchmarks,” it found things competitors missed and missed things they found: capable, but “not mind-blowingly good.”
Its clearest functional distinction is access to X data. A user can ask it to analyze someone’s posts and infer that person’s position on a subject—useful integration, though not enough by itself to pull subscribers away from ChatGPT or Claude.
Newton found its self-description inflated—“Grok 3’s launch is a pivotal moment in AI”—but its assessments of Platformer and The New York Times were surprisingly even-handed. Roose’s conclusion: the product is presently “pretty bog-standard.”
2. The “based” model kept returning inconveniently progressive answers
Musk has pitched Grok as comparatively uncensored, truth-seeking, and free from progressive filters. Roose tested that premise by asking how many genders exist; Grok answered that context matters, “Gender is fluid,” and “there’s no hard number.”
Newton’s playful hope is that training on broad human knowledge might make superintelligence naturally kind, empathetic, and progressive. He immediately hedged that he did not expect this fantasy to come true, but observed that current models often emerge “pretty sweet.”
Roose expects Musk to keep “fiddling with the dials” until Grok better reflects his desired politics. For now, despite Musk’s claims that it is based and anti-woke, Roose said: “In my own testing, that does not appear to be true.”
3. Colossus proves xAI can turn money into frontier capacity quickly
The more consequential achievement is speed: xAI moved from a “very bad” V1 to a capable V3 in about a year. Roose traced that progress to Colossus, the Memphis data center reportedly containing something like 200,000 NVIDIA GPUs.
Those chips cannot simply be ordered from Best Buy. Roose estimated a bill in the billions and emphasized Musk’s special NVIDIA relationship through Tesla, arguing that xAI scaled physical infrastructure faster than comparable Microsoft and Amazon efforts.
Newton’s pushback—worth keeping: DeepSeek appeared to reach strong performance with less-powerful hardware and innovations other labs could copy, whereas Grok looked like Musk “throwing money at a problem.” Roose agreed that Grok follows the scale route: “just build a bigger data center.”
That is not cheating; it is how American labs have advanced for years. But Grok also stands “on the shoulders” of prior research, reinforcing the idea that the field’s decisive shifts were ChatGPT and then OpenAI’s o1-style test-time compute—not every one-point benchmark lead.
4. Musk’s AI race now intersects directly with state power
Grok shows Musk is prepared to spend “a phenomenal amount of money” to remain near the frontier. Roose reconsidered Musk’s support for the six-month AI pause: while publicly warning others to slow down, Musk was racing to build a competitor and, in Roose’s reading, wanted time to catch up.
The motive, as the hosts framed it, is less financial recovery than control and historical primacy. Musk wants to beat his “nemesis” Sam Altman to superintelligence; Roose’s blunt conclusion was, “I think this is pure power.”
Newton then raised Musk’s position as an “unelected fourth branch” of government, with a team dismantling whole swaths of the federal government while discussing opaque uses of AI. Roose resisted treating Grok and federal power as one grand scheme, stressing that Grok 3 is not ready to run critical systems such as Social Security payments.
The concrete conflict is infrastructure. Federal favor could accelerate data-center permits, unlock energy, or confer advantageous grid access—materially improving xAI’s position without requiring a plan to make Grok run the government.
5. A government-anointed AI remains remote but no longer unthinkable
Newton’s scenario began with the AI CEOs’ prediction that AGI could arrive within about two years, meaning a system able to perform essentially any remote worker’s job. Such tools could also help users develop major cyberattacks, eventually forcing today’s deregulatory pendulum back toward safety restrictions.
His explicitly speculative question: might Washington decide private companies should not be “building God,” then license a single approved provider—and choose Grok? Roose called that a “remote possibility,” preserving the uncertainty rather than endorsing it.
Roose recalled an AI-risk tabletop exercise in which Musk persuaded Donald Trump to nationalize OpenAI and put Musk in charge as “a middle finger to Sam Altman.” It seemed fantastical then; after Musk’s lawsuits, takeover efforts, and government role, Roose was “not so sure.”
Grok would demonstrate true leadership through original advances, not parity. Musk and his top engineers predicted that within one or two years AI, assisted by a human expert, could win a Fields Medal or Nobel Prize; Roose was uncertain Grok would do it, but thought comparable breakthroughs could happen fairly soon.
6. Robinhood wants every financial transaction under one roof
Robinhood has expanded far beyond the commission-free mobile brokerage Newton first covered in 2013. Its menu now spans stocks, options, futures, crypto, prediction markets, retirement, credit cards, and 401(k) transfers; the company was worth $52.2 billion at recording.
Tenev’s long-term formulation was maximalist: Robinhood should be where customers “buy, sell, trade any financial asset, or conduct any financial transaction.” Trading was the entry point; the strategy is now the whole consumer financial-services stack.
The hosts entered skeptical because the zero-commission breakthrough removed a cost that might otherwise discourage excessive trading. Newton said he warned in 2013 that the model would encourage trades that lose customers money, and his discomfort has only increased.
Newton disclosed that Robinhood owns Sherwood News, which briefly syndicated some Platformer content for a few months the previous year; he said that arrangement was no longer active.
7. Tokenization is Tenev’s answer to crypto’s meme-coin trap
Tenev accepted the skeptical account of 2021—heavy speculation, widespread losses, and little useful construction—but offered a regulatory mechanism. When crypto connects to a productive asset, it becomes a security; because that linkage has largely been barred, activity gets pushed toward things without underlying value.
His preferred use is access to private companies such as OpenAI and SpaceX, which, he said, most people would consider to have a relatively low risk of going to zero right now but which most Americans cannot buy. Meanwhile, high-growth value in Anthropic and Perplexity also accrues mainly to private-market insiders.
The hosts’ ICO pushback was that token access previously generated scams, rug pulls, and bad-faith behavior, not merely the Telegram ICO that was shut down. Tenev conceded the abuse and argued for mitigation, but insisted the technology’s benefits were too extreme to suppress.
Pump.fun supplied his technological contrast: anyone can create a coin in five minutes and trade it globally across exchanges, indexes, and wallets, while an IPO requires banks, counterparties, a road show, and immense expense. The objective is to attach the first system’s efficiency to real assets.
8. Disclosure tiers would replace a single gate to retail access
Tenev proposed self-certification and optional disclosure. A late-stage private company with public-like audited financials could qualify for superior placement; a token from a meme factory might carry a “big red skull and crossbones” stating that it is unvetted and unverified.
Roose identified the unresolved problem: outsiders have not seen audited financials for OpenAI or Anthropic, so buying merely from fundraising headlines is “throwing a wish in a fountain.” Tenev favored rewarding disclosure with access and visibility rather than forcing companies to provide it.
The underlying philosophy is that people are intelligent enough to decide when risks are made legible. That leaves the hosts’ central objection intact: better token plumbing does not itself ensure that the asset, issuer, or investor behavior is sound.
9. Prediction markets promise faster news but blur into sports betting
Robinhood offered a “pro football championship” contract to roughly 1% of users before the CFTC requested suspension over “serious concerns.” Tenev distinguished prediction markets from sports betting; Roose pressed that wagering on a game’s winner for a payout is mechanically the same consumer experience.
Tenev’s broader thesis is that prediction markets are “the future of not just trading, but also information.” Like a newspaper with multiple sections, they package economically valuable information—but deliver “the news faster,” sometimes before the underlying event occurs.
Roose called the football product regulatory arbitrage: casting a sports wager as a federally regulated derivatives contract could bypass state betting restrictions. Tenev would not concede that interpretation, saying classifications remain unsettled while prediction markets themselves “are here to stay.”
The election was Tenev’s best evidence: markets reached 95–5 for Trump while broadcasts still parsed county returns. Roose countered with LK-99, when markets assigned high odds to a room-temperature superconductor until journalists and scientists checked and failed to replicate it. Tenev conceded markets are not always right, while maintaining that they are the most effective mechanism he has seen for synthesizing public information.
10. The access argument collides with engineered speculation
Tenev rejected the premise that markets equal gambling. His macro case is that open-market countries have tended to outperform closed-market countries, making access a major source of wealth creation even when particular customers suffer losses or other negative externalities.
Roose’s challenge focused on Robinhood’s own prompts: an alert advertising the Trump meme coin and a New Year’s Dogecoin giveaway. Tenev said notifying users is not coercing them; some buy expecting appreciation, while others genuinely want to support the represented movement.
Tenev also described Robinhood as unusually selective versus crypto venues listing hundreds of coins. Roose’s cutting summary—“Only the blue-chip meme coins”—captured the unresolved mismatch between Robinhood’s stewardship language and the speculative products it promotes.
Asked about rising gambling-addiction searches and therapists reporting more affected boys and young men, Tenev said he was less familiar with gambling addiction because Robinhood is not in the gambling space. He pointed instead to financial controls including onboarding, suitability, and state-level geolocation. Roose had cited a JAMA study released earlier that week reporting a significant rise in internet searches related to gambling addiction over recent years.
11. Vibe coding expands creation while hollowing out comprehension
Tenev supports suitability controls but resisted blanket prevention of investments deemed bad. His example was Massachusetts blocking residents from Apple’s 1980s IPO: a seemingly protective decision can look destructive decades later. Still, he named mortgage-backed securities and credit-default swaps as products retail may not need or understand.
On AI, Tenev was “very, very confident” that money, currency, and companies will survive labor disruption, perhaps with AIs creating companies too. His conclusion was counterintuitive: uncertainty makes investing and saving more important—“retirement becomes even more important.”
Harmonic attacks hallucinations through mathematics, where one faulty reasoning step usually ruins the result. Tenev’s dream is a “super calculator” with an LLM’s flexibility and a calculator’s reliability, producing verifiably correct outputs at every step.
Roose’s parallel experiment used Cursor, Replit, Claude, and ChatGPT to build “software for one”: a podcast summarizer, X-bookmark scraper, trunk-fitting app, and the pink “Hot Tub Time Machine.” The last sent weekly, monthly, quarterly, and annual maintenance reminders plus poems, built in about half an hour with no handwritten code.
The limits surfaced around authentication, databases, missing APIs, and choices a novice cannot evaluate. Roose admitted the AI might insert malicious code without his knowledge; Karpathy’s seductive workflow remains, “Okay, okay, okay. Accept, accept, accept.”
Newton connected that opacity to Namnoye Goel’s blog post, “New Junior Developers Can’t Actually Code,” which had 1 million views according to the post Casey was reading. Roose expects engineers to become product-manager-like supervisors, but Newton wants a durable core who understand systems “down-to-the-metal”; otherwise society must ask AI how it works and simply trust the answer.