
14-Year-Old Interviews SBF Behind Bars: $500M Bet on Anthropic Paid Off, but FTX’s Downfall Left Him Regretting It
Curated & Compiled: TechFlow

TechFlow Guide: Accompanied by the automated federal prison voice prompt "Call from a Federal Detention Facility," former FTX founder Sam Bankman-Fried (SBF) engaged in an intense, high-energy deep-dive conversation lasting 82 minutes with 14-year-old tech podcaster Eli Goldfine at the Metropolitan Detention Center (MDC) in Brooklyn. In what is regarded as one of the most information-dense dialogues in recent years, the incarcerated SBF rarely broke his silence, comprehensively recounting his cognitive landscape, investment journey, and extreme philosophy.
From the behind-the-scenes maneuvering that led FTX Ventures to inject $500 million into Anthropic, to radical projections of the future AI landscape worth hundreds or even trillions of dollars; from the cold-eyed scrutiny of traditional software moats and Coinbase's brand value, to facing philosophical questions head-on akin to Tyler Cowen's "51/49 coin flip: destroy the world or double benefits" and personal freedom bets; to finally confessing at the endgame that "the outcome brought by FTX is not worth it, and when the crisis erupted, all reality was distorted by the victors' filter," SBF presented, through the most raw mental deductions, the complete spiritual world of a genius trader after falling from grace.
Guest: Sam Bankman-Fried (Founder and former CEO of FTX, currently serving time in a federal prison; core practitioner of effective altruism)
Host: Eli Goldfine (Creator of the 14-year-old tech and forecasting podcast The Super Cycle, Twitter @realTomBayes)
Podcast Source: The Super Cycle (https://x.com/realTomBayes/status/2123761258306384333)
Original Title: EXCLUSIVE: I'm 14, and I just interviewed Sam Bankman-Fried
Release Date: November 20, 2026 (Beijing Time: November 21, 2026)
Key Takeaways
The greatest torment in prison is the isolation from cutting-edge information: SBF admitted that reflecting on his case behind bars yields no incremental value for the outside world; what truly pains him is the inability to real-time track LLM benchmark performance, code details, market trends, and startup due diligence. The lack of granular, frontier inputs makes meaningful thinking extremely difficult.
The origins of Anthropic's early $500 million wager: FTX Ventures virtually covered Anthropic's entire Series B round at the time. They refrained from injecting more compute capital because excess cash easily tempts companies to overspend (e.g., blind user acquisition), and the team preferred not to overly dilute equity, as reckless expansion could put the cart before the horse.
AI industry to achieve multi-trillion-dollar dominance: Over the next decade, AI will undoubtedly become a pillar, if not the dominant sector, of the global economy; given that only a handful of top-tier labs (Anthropic, OpenAI, DeepMind, xAI) can compete for the crown, it is entirely logical for leading institutions to command trillion-dollar valuations in economic models.
Traditional software companies are not completely sentenced to death by AI: Physical industries (like steelmaking and mining) will feel AI's impact later; for pure software and financial institutions, competitors cannot easily dethrone pioneers even if they replicate identical features at near-zero cost using AI, because the true moat for brands like Coinbase and Robinhood lies in unshakeable consumer trust.
p(doom) assessment falls in a complex gray zone around 25%: SBF believes the probability of humanity being destroyed or facing a survival catastrophe due to AI lies between 10% and 50% (median ~25%); catastrophic scenarios are unlikely to be simple out-of-control nuclear weapons, but rather complex gray-area situations, such as the complete marginalization and obsolescence of non-uploaded populations after mind uploading.
Political prediction markets were once FTX's unfinished ambition: Fundamentally, prediction markets are futures contracts governed by specific settlement rules; FTX initially ran hundreds of millions of dollars worth of U.S. election predictions in offshore markets, but faced restrictions due to lacking a CFTC license. The explosive growth of Kalshi and Polymarket has validated the immense power of this sector.
Highly skeptical of print book information throughput: SBF reiterated his famous stance that most books could be condensed into a six-paragraph blog post; in contrast, audiovisual streams (like Netflix) offer vastly higher information throughput and expressive dimensions per unit of time compared to books bogged down by extensive filler.
Extreme expected value vs. the reality check of the 51/49 coin flip: Under pure linear expected value theory, betting on a 51% chance to double outcomes versus a 49% chance to destroy the world is mathematically sound; but in reality, destruction carries irreversible extinction costs, and real utility suffers from diminishing marginal returns. Faced with risking immediate release via a 51/49 bet, individual risk-utility functions dictate that theoretical gambles cannot be simply mapped onto reality.
A poignant review of the FTX endgame: Asked whether "it was all worth it," SBF firmly stated he "can't dishonestly say it was worth it; if things hadn't fallen apart at the end, it absolutely should have been." He noted that during crises, people easily lose their reason and core judgment, after which all history is reshaped and distorted through the victor's filter.
Survival guidelines for the new generation of 14-year-olds: Stop fixating on Wall Street's "Earning to Give" high-frequency trading careers; the urgent priority is to wholeheartedly understand AI, dive into AI safety, or find practical entry points where AI can genuinely automate infrastructure.
Key Highlights
On the brutal barrier of the prison information gap: "If you want to create value at the absolute forefront, you need to master hundreds or thousands of minute details: exactly where large models excel, where they remain clumsy, what the industry frontier is debating, and how microstructures in markets are shifting. Behind the high walls of a federal prison, acquiring these details is virtually an impossible task."
On Anthropic's valuation and moat: "If you accept that AI will become the core sector dominating tens of trillions of dollars globally over the next decade, and there are only a mere handful of top-tier labs sitting at this table worldwide, then it is entirely logical for each leading company to capture a multi-trillion-dollar market cap. Anthropic's prowess in code generation and software engineering tools remains among the absolute best globally to this day."
On the true moats of brand and technology: "Many developers think technology is everything, but look at Coinbase or Robinhood. What allows them to command hundreds of billions in valuation? Absolutely not because they possess uncopyable proprietary algorithms, but because they hold brand trust that American retail users instinctively name-drop. Even if competitors could replicate an identical software suite at near-zero cost using AI, they still couldn't purchase this brand asset."
On the radical critique of print books: "I still stand by my view that writing a book often means you got it wrong—it should have been a six-paragraph blog post instead. Books have abysmal information throughput, bogged down by lengthy, useless exposition; in the same unit of time, dense visual and audiovisual inputs convey exponentially more valuable information."
On realistic reflections regarding the 51/49 coin flip: "In pure ivory-tower expected value theory, you can deduce betting global utility doubling against a 49% destruction probability; but the real world harbors irreversible systemic zeroing risks, and every positive utility has its limit of diminishing marginal returns. Once you factor in the fatal cost of destruction, you simply wouldn't dare toss most real-world coins."
On the final verdict of FTX's rise and fall: "I want to self-deceptively comfort myself that it was all worth it, but I cannot lie to my conscience and say so. It should have been worth it, had I not completely botched everything at the final hour. The moment an extreme crisis erupts, panic distorts all advice and decisions. Once the smoke clears, all historical narratives are inevitably wholly reconstructed through the victor's lens."
Transcript
Prison Life and Cognitive Barriers: A Black Box World Stripped of Micro-Details
Eli Goldfine: Welcome to The Super Cycle. This is Eli Goldfine. We are deeply honored to connect with Sam Bankman-Fried (SBF) today. My team and I coordinated preparations for months prior to this, making it the most information-dense in-depth interview since our podcast's inception. Hello, Sam, can you hear me?
Sam Bankman-Fried: Yes, loud and clear.
Eli Goldfine: Great connecting with you. I've watched many of your recent interviews, and almost all external attention focuses on the collapse of FTX and your daily routine in prison. You have plenty of solitude and thinking time behind bars. What subjects do you usually spend the most energy pondering?
Sam Bankman-Fried: A very natural question. Obviously, I do spend a considerable amount of time reviewing my own case, reconstructing what exactly happened, and thinking about what I can still contribute. But I am fully aware that for people on the outside, these reflections actually offer little incremental value.
I try to shift my focus toward considering what I can do in the future. But in the prison environment, attempting productive forward-looking thinking is an extremely difficult endeavor.
Eli Goldfine: Why is forward-looking thinking so difficult behind bars? Is it due to a lack of certain key information?
Sam Bankman-Fried: It's not a lack of a single piece of information, but rather a lack of countless microscopic yet crucial inputs. If you want to deduce the future at the absolute frontier of human cognition and business, you must grasp massive dynamic details: How exactly are frontier large language models evolving? On which specific tasks do their benchmark curves break human limits, and in which areas do they remain clumsy? You must personally test these technologies in practice and converse daily with the engineers building them at the front lines.
The same applies financially. You must watch the tape daily, observe capital flows and trading structures, build networks, connect with emerging startup teams, discuss their business models, and conduct deep due diligence. These are daily necessities if I were on the outside, but within the high walls of a federal detention center, accessing each piece of information becomes extraordinarily difficult, bordering on impossible.
Eli Goldfine: If you could somehow get information inside, which dimensions of data would you consider most valuable for modeling the real world?
Sam Bankman-Fried: As I said, it's not singular breaking news, but thousands of microscopic variables. For example, the architectural evolution of frontier models, the practical deployment thresholds for code generation, the cash runways of startups in real commercial environments, and genuine procurement intent across different industries. Without these granular inputs, macro frameworks deduced purely in a vacuum are typically extremely fragile.
Eli Goldfine: You obviously can't just think about your case all day. Besides the legal proceedings, what broader issues occupy your mind?
Sam Bankman-Fried: True, but much of my thinking remains heavily constrained by my current physical environment, as it's the only world where I can directly receive first-hand feedback right now. Not that it matters intensely, but I have spent considerable time contemplating the criminal justice system. Not merely through the lens of my personal ordeal, but the absurd logic exhibited by the entire judicial framework on a systemic level.
Reflections on the Federal Criminal Justice System: Fictional Charges, KPIs, and Perverse Incentives
Eli Goldfine: Setting aside your personal case, what specific universal judicial issues have you observed?
Sam Bankman-Fried: As the saying goes, when you have a hammer, everything looks like a nail. The justice system operates on the exact same principle. If Congress establishes a specialized federal task force to combat a specific type of crime, that task force will desperately hunt for sources regardless of whether the crime actually exists at scale.
Once a task force's headcount and budget exceed the actual volume of that crime in society, it inevitably begins manufacturing fake cases, arresting behaviors that don't even constitute crimes. The federal anti-drug apparatus has exposed this distortion for decades, and in the last ten years, this风气 (trend) has permeated and spread across every domain of federal justice.
Eli Goldfine: Like in which specific domains?
Sam Bankman-Fried: Take the massive federal cybercrime investigation agencies, originally designed to combat extreme offenses, which now extend their reach into private text conversations and normal interactions between lovers. Countless completely innocent people are prosecuted and arrested simply because prosecutors need these cases to pad promotion and bonus KPIs. The case of former Honduran President Juan Orlando Hernandez perfectly illustrates how unhinged prosecution mechanisms can become.
Eli Goldfine: Do you still believe white-collar criminals should serve prison sentences?
Sam Bankman-Fried: Absolutely not. Locking up white-collar offenders does nothing positive for society. It merely strips decades from someone's life and forces taxpayers to bear the enormous financial burden of prison operations. If your sole goal is retribution, it technically achieves that; but I fundamentally reject the notion that this system rehabilitates anyone.
Eli Goldfine: For individuals who actually committed financial fraud, what do you consider the most reasonable course of action?
Sam Bankman-Fried: It depends on the specific factual details of the case. However, in the vast majority of commercial and financial disputes, civil litigation is a far more precise and rational pathway than criminal trial. Civil remedies can resolve matters through massive fines, contract modifications, licensing restrictions, and asset forfeiture.
More importantly, as a modern society, we must learn to curb blind overconfidence. We often assume we can perfectly see the truth and that we are always punishing the guilty, but in reality, many defendants put on trial are merely convenient scapegoats deployed by powerful factions to divert blame.
Eli Goldfine: Should criminal justice reform be a top priority for the effective altruism community?
Sam Bankman-Fried: From an expected utility perspective, this domain is highly complex. While social justice reform undoubtedly holds massive value, it faces incredibly rigid political resistance and institutional inertia. When powerful interest groups build massive industrial ecosystems around incarceration, driving bottom-up change becomes extraordinarily difficult.
Eli Goldfine: In your view, what is the easiest low-hanging fruit to improve the criminal justice system?
Sam Bankman-Fried: The easiest reforms that immediately unlock huge economic dividends involve completely reevaluating pretrial detention and sentencing baselines for non-violent offenses. Our system is saturated with impoverished individuals locked up for years unable to pay nominal bail, and working-class populations sentenced to decades for drug offenses. Providing these individuals with formal vocational training and community-based corrections, rather than throwing them in prison to sever all social ties, offers the highest cost-benefit ratio.
AI Industry Endgame Projections: Trillion-Dollar Valuations and a Few Oligarchs at the Table
Eli Goldfine: Let's discuss frontier tech. Anthropic is recently preparing for an IPO, and external valuation expectations are scaling toward hundreds of billions, even trillions of dollars. As the early lead investor who put $500 million into Anthropic's Series B, how do you view current valuations for these foundation model labs?
Sam Bankman-Fried: I believe they are absolutely worth that, and likely worth significantly more in the future.
Deducing from macroeconomic logic: Throughout human economic history, all core pillars that actually support civilization运转 (operation) have total market capitalizations ranging from several trillion to tens of trillions of dollars. Over the next five to ten years, AI will almost certainly mature into the world's most core, and potentially dominant, economic industry.
Eli Goldfine: If the entire AI sector eventually reaches a multi-trillion-dollar market cap, how will that value be divided?
Sam Bankman-Fried: Assuming two-thirds of that eventual multi-trillion-dollar total market cap stays with innovations and new companies that haven't even been born yet, the remaining board still shares over ten trillion dollars among existing foundational model enterprises.
But count carefully: how many top-tier labs are actually qualified to sit at this core infrastructure table? OpenAI, Anthropic, Google DeepMind, plus xAI. You can count them on one hand. When a multi-trillion-dollar track is ultimately carved up by a single-digit number of oligopolistic leaders, the inevitable mathematical conclusion is that each top-tier frontrunner must be valued in the trillions. Anthropic is unquestionably a top-two global entity.
Eli Goldfine: Why do you rate Anthropic so highly? Is it primarily because Claude series models perform well on benchmarks?
Sam Bankman-Fried: Benchmark scores matter, and the public clearly sees Claude's strong reputation in everyday conversation and consumer-grade applications. But Anthropic's true differentiator is its absolute dominance in software engineering and professional developer tooling. Their development suites and code comprehension capabilities for engineers are globally top-tier, forming the core workhorse machines for the future digital society's transition to AI automation.
Eli Goldfine: If we enter a full-blown superintelligence arms race over the next decade, what will the world look like?
Sam Bankman-Fried: Productivity distribution across society will be completely reshaped. For the past few years, although large models evolved exponentially, their absolute capabilities remained below human averages, keeping the societal order seemingly calm. However, once crossing a specific threshold, machine output speed in programming, law, financial modeling, and drug discovery will be hundreds of times faster than humans. The pace of wealth creation will accelerate massively, but distributing the computational dividends will become the fiercest geopolitical battleground.
Behind the Early $500 Million Heavy Bet on Anthropic: Compute Arms Race and Capital Traps
Eli Goldfine: Many have always wondered why FTX Ventures virtually underwrote Anthropic's entire Series B. Since you were already so confident in AI's future back then, why didn't you write a larger check, like instantly injecting $1 billion or $2 billion to let them unboundedly purchase GPUs and build data centers?
Sam Bankman-Fried: There are two critical considerations here. First, from a founder/team perspective, exceptional early-stage teams never want to overly dilute control or equity; they only raise the exact capital needed for their current development phase.
Second, business history shows many startups died from capital shortages, but just as many were swallowed alive by excessive cash reserves and counterattacked by their own capital structure. Many founding teams develop a blind anxiety to "spend the money" once they secure astronomical funding, leading them to waste billions on meaningless marketing or reckless expansion, ultimately losing reverence for core technical攻坚 (breakthroughs).
In the early days of building foundation models, architectural iteration and algorithm verification require steady, disciplined R&D cycles. You absolutely cannot pull up seedlings to help them grow by simply stacking compute. Anthropic's capital allocation was highly rational then; the funds we injected perfectly satisfied their core needs to launch the next-generation model training.
Eli Goldfine: How did FTX Ventures operate back then? What role did you personally play in specific investment decisions?
Sam Bankman-Fried: In the early days, many deals were sourced organically by team members during daily operations, including numerous crypto-native investments that consistently delivered multi-fold annual returns. Anthropic, like Solana, started as disruptive projects we stumbled upon and deeply connected with.
As we scaled, we onboarded professional partners to build a formal investment structure. For standard deals, I lean heavily on trusting the analysis team handling due diligence; but for strategic heavyweight wagers like Anthropic, I tracked progress weekly or even more frequently, exploring how we could assist them on underlying compute, financials, and resource networks.
Eli Goldfine: When you initially engaged with Dario Amodei's team, what left the deepest impression on you?
Sam Bankman-Fried: Extreme rigor and focus. When Dario and his team split from OpenAI, their objective was razor-sharp: build frontier safety models with high interpretability and alignment guarantees. They refused flashy marketing demos for short-term clicks; every dollar of capital expenditure was tightly locked into critical compute clusters and engineering talent acquisition. This engineering culture is exceptionally rare in today's浮躁 (flashy) Silicon Valley.
The Existential Crisis of Traditional Software and Brand Moats: Why AI Can't Kill Coinbase
Eli Goldfine: With the explosive capabilities of frontier models, many traditional software companies face severe existential anxiety. Do you believe traditional software firms that haven't deeply integrated AI have literally no way out?
Sam Bankman-Fried: It's not that absolute. First, substantial physical industries possess inherent buffering defenses against AI penetration. AI might be brilliant, but it cannot magically extract ore from underground mines overnight.
Second, for purely digital software, winning isn't determined solely by底层 (underlying) code efficiency. Many severely underestimate the absolute moat of "brand." Suppose a competitor uses the latest LLM tools to clone a functionally identical software product at near-zero cost. Does that mean users will abandon the incumbent? Absolutely not.
Eli Goldfine: Just like Coinbase or Robinhood?
Sam Bankman-Fried: Exactly. Why do Coinbase and Robinhood command hundreds of billions in market cap? Is it because they possess some impenetrable, secret matching engine code that other teams can't replicate? Of course not. Their most precious asset is brand recognition—the instinctive trust endorsement spoken by millions of everyday retail investors wanting to buy or sell digital assets.
Even if a new team could rewrite a frontend and trading system at near-zero cost using AI, they simply cannot purchase this沉淀 (accumulated) brand reputation in the short term. Brands often allow traditional enterprises to maintain long-term survival resilience despite homogenized underlying products.
Eli Goldfine: Then which traditional software companies will be eliminated fastest?
Sam Bankman-Fried: Those purely relying on feature stacking, possessing neither brand mindshare nor deep network effects. If your product is merely a basic ticket management or form-filling tool, with zero emotional or brand binding to users, then native AI tools will shatter it within a short timeframe at a fraction of the cost.
Real Alpha in Tech Investing: Minimalist AI-Native Teams vs. Buzzword-Junk Companies
Eli Goldfine: Where, then, do you believe true超额收益 (excess returns/Alpha) currently hides in tech investing?
Sam Bankman-Fried: Real Alpha lies in the ability to pierce through bubbles. Currently, half of the startups in the market slap an AI label onto their pitch decks, even if a T-shirt printing company claims it uses AI to design patterns. These are essentially industrial garbage.
Meanwhile, massive opportunities reside in extremely few native teams consisting of only two or three core members: they leverage the AI toolchain completely, radically restructuring traditional workflows that previously required hundreds or thousands of employees into ultra-lean operations, rapidly scaling into unicorn species valued at tens or hundreds of millions. Distinguishing between genuine breakout momentum and opportunistic hype is the sharpest dividing line for current investing.
Eli Goldfine: If you were managing capital freely outside today, would you return to this kind of investing?
Sam Bankman-Fried: Absolutely. If I were on the outside, I would aggressively seek out and heavily back these two-to-three-person super-species. They lack corporate bureaucracy and internal friction; every member utilizes LLMs to output equivalent to a complete engineering department. The explosive velocity and profit margins driven by this ultra-lean headcount are unimaginable for traditional organizational structures.
Biotech and Longevity Tech: Regulatory Shackles and AI-Rushed Research
Eli Goldfine: Let's talk biotech. FTX also funded numerous anti-aging, longevity, and gene therapy projects early on. What are your thoughts on progress in biomedicine?
Sam Bankman-Fried: Biotech is a track bursting with potential, but its biggest bottleneck isn't scientist intelligence; it's outdated regulatory systems. The FDA clinical approval process is exceedingly long and astronomically expensive. A single new drug often consumes billions of dollars and a decade before hitting the market.
Countless early therapies with genuine life-saving potential die starved in labs because they cannot raise astronomical sums to finish Phase III trials. We aimed to provide early catalytic funding for these high-risk, high-reward frontier attempts.
Eli Goldfine: Will AI change this status quo?
Sam Bankman-Fried: AI already demonstrates disruptive power in protein structure prediction, molecular docking, and compound screening. It compresses preliminary lab screening that used to take years or decades into mere months. Ultimately, however, all drugs must still pass physical human clinical testing. Unless we fundamentally overhaul archaic, rigid regulatory processes, pure technological acceleration will still be severely delayed by offline administrative barriers.
Eli Goldfine: For rare diseases or fatal illnesses, how do you think clinical approval should be optimized?
Sam Bankman-Fried: Introduce more flexible informed consent and parallel pathways. For patients with terminal illnesses whom modern medicine has already abandoned, if they are fully informed and voluntarily opt into experimental therapies in Phase I or II, regulators shouldn't crudely strip them of survival rights. Dynamic tracking using Real-World Data (RWD) holds far more ethical warmth than mechanically waiting ten years for dual-blind Phase III trials.
Politics, Regulation, and Reflections on the Biden Administration: Crypto Hostility and Democratic迷失 (Loss)
Eli Goldfine: Let's discuss politics and macro-regulation. The Biden administration adopted an extremely aggressive posture on tech and financial regulation over the past four years, particularly the SEC's comprehensive enforcement actions against the crypto industry. How do you evaluate their regulatory performance?
Sam Bankman-Fried: I found the Biden administration's performance in this area profoundly disappointing.
Not just regarding crypto, but across the entire innovation ecosystem, regulators displayed a deep bureaucratic hostility. Instead of establishing clear, executable, innovation-promoting legal frameworks, they relentlessly executed surprise raids to打击 (crack down on) compliant domestic enterprises, directly driving massive pools of top tech talent and innovation capital toward offshore markets.
Eli Goldfine: You made massive donations during political campaigns. Why do you think policymakers' understanding of the tech industry has become so disconnected?
Sam Bankman-Fried: Many politicians and regulators fundamentally misunderstand the底层 (foundational) logic of technology. They only care about political narratives and power matrices. Policy decisions are rarely based on which方案 (approach) benefits socioeconomic welfare most, but rather which brings the most votes and political筹码 (leverage). I believe Democrats have seriously lost their way on numerous economic and tech issues over the past decade.
Eli Goldfine: Facing international tech competition, should tech companies slow down to guarantee security?
Sam Bankman-Fried: That is an extremely dangerous game-theory trap. If one security-conscious company unilaterally slows down while global competitors, or even hostile nations, sprint full-speed ahead, unilateral self-restriction won't make the world safer; it simply hands control of frontier technology entirely to rivals who ignore safety and ethics altogether. The ideal state requires major participants to establish substantive safety compacts, but in real-world geopolitics, that is extraordinarily difficult.
The Nature and Explosion of Political Prediction Markets: From FTX 2020 Election Contracts to Kalshi and Polymarket
Eli Goldfine: Let's discuss financial innovation. Kalshi's recent valuation surge and Polymarket's staggering trading volumes in global political elections are notable. Rumor has it that building a global political prediction market was once one of your original visions for FTX?
Sam Bankman-Fried: From a financial engineering perspective, so-called prediction markets are essentially futures contracts that settle cash based on specific real-world event outcomes. They are no different fundamentally from Bitcoin or crude oil futures.
During the 2020 U.S. presidential election, FTX International hosted the network's most active Trump vs. Biden win-probability contracts, reaching hundreds of millions in trading volume. From day one, we craved to scale this market, but the biggest constraint remained regulatory licenses. Lacking a本土 (domestic) CFTC permit, we strictly banned U.S. resident participation and enforced mandatory KYC verification across our centralized exchange.
Eli Goldfine: Why have Kalshi and Polymarket completely exploded in this cycle?
Sam Bankman-Fried: They paved two distinct paths. Kalshi spent years fighting lawsuits and formal procedures to secure a compliant domestic license; Polymarket leveraged offshore and decentralized architecture to aggregate massive liquidity globally. This thoroughly validates our original judgment: human demand for pricing historic events is universal and genuine.
Beyond prediction markets, tokenized equities still hold enormous growth potential. Traditional stock trading markets rely on archaic, inefficient underlying settlement infrastructures. Blockchain-based 24/7 seamless clearing systems possess complete dimensional-strike advantages in user experience.
AGI Threshold and p(doom) Assessment: 25% Extinction Probability and the Gray Dilemma of Mind Uploading
Eli Goldfine: Everyone discusses when Artificial General Intelligence (AGI) or Artificial Superintelligence (ASI) will arrive. What is your personal timeline expectation for AGI?
Sam Bankman-Fried: I don't view AGI arrival as a sharp, black-and-white discrete event. It will more likely manifest as a continuous, progressive substitution process.
For a long period, although AI computation doubled constantly, its intelligence remained far below human baselines, causing virtually no substantive real-world impact. But as its intellect gradually approaches human levels and breaks through critical thresholds in specific verticals, qualitative shifts will spread rapidly. We already see this trend in software development, advanced mathematical derivation, and game theory strategies. Within the next three to five years, critical societal sectors will be deeply taken over by AI piece by piece.
Eli Goldfine: Within the effective altruism community, AI safety and human extinction probability have always been the central debate. You've never publicly given a specific numerical p(doom) for yourself. Can you give a concrete assessment today?
Sam Bankman-Fried: If forced to give a number, I place it below 50%, above 10%, roughly hovering around the 25% bracket.
But one point must be clarified: in my assessment model, catastrophic scenarios rarely resemble Hollywood movies featuring maliciously rogue AI suddenly grabbing nuclear launch codes and declaring war on all humanity. The higher probability involves extremely complex gray ethical dilemmas.
Eli Goldfine: What kind of gray dilemma?
Sam Bankman-Fried: Take mind uploading. If technology advances to a stage where human consciousness can be digitally uploaded to cloud servers, those achieving digital immortality will surpass biological humans geometrically in compute and evolutionary speed. In that context, will native biological humans remaining in the physical world be completely marginalized and obsolete?
Does this gradual de facto extinction of biological intellectual species count as a human civilization catastrophe? Different philosophers will give diametrically opposite answers. Once you fold these ethically ambiguous evolutionary pathways into consideration, survival risk probability assessments violently oscillate between 20% and 60%.
AI Agents and Future Workflows: Tenfold Human Super Research Assistants
Eli Goldfine: If a super-intelligent assistant emerges that comprehensively surpasses you in logical reasoning and decision-making expected value, would you willingly surrender all personal decision-making authority to it?
Sam Bankman-Fried: That depends entirely on its optimization objective function. Is it maximizing its own survival probability, maximizing wellbeing for all sentient life, or faithfully executing the delegated goals I assigned?
If its core aligned objective functions as a loyal, meticulous, and rationally prudent research assistant—capable of completing in one minute what previously demanded a full day of grueling desk due diligence and deduction at human speeds—I would absolutely entrust vast execution and analysis tasks to it. But ultimate value orientation and ethical baselines must always anchor to core human demands.
Eli Goldfine: How would this super assistant disrupt finance and investing?
Sam Bankman-Fried: The pyramidal structures built on hundreds of junior analysts in traditional institutions would collapse instantly. One partner paired with a few autonomous agents specifically tuned for financial statement cross-verification, macro data cleaning, and sentiment tracking could replicate the output of an entire investment bank building. Inter-institutional胜负 (winning or losing) will no longer depend on hiring Ivy League graduates, but on who deeply understands agent fine-tuning strategies and underlying architectures.
Why Profoundly Skeptical of Print Books? Information Throughput and Dimensional Strikes from Audiovisual Media
Eli Goldfine: Let's touch on a fascinating lifestyle topic. Many clever people pride themselves on reading extensively, but I recall you once made a highly counterintuitive claim: "I deeply doubt books' value; the vast majority belong in a six-paragraph blog post." Now behind bars, how many books have you read? Has your stance changed?
Sam Bankman-Fried: Prison limited my options, forcing me to read quite a bit, which actually reinforced my conviction that my original judgment was correct.
If you write a full-length book, it often signals you failed to crystallize your core argument. In a multi-hundred-page tome, genuinely valuable insights could be fully conveyed in a few paragraphs; the remaining ninety percent is usually padded with lengthy examples, buildup, and trivial narratives just to meet page-count expectations.
Eli Goldfine: What genres did you primarily read in prison? Any single book that proved worthwhile?
Sam Bankman-Fried: I read heavily in science fiction. Honestly, most sci-fi is merely average. Given the choice, I'd far rather watch hardcore sci-fi series on Netflix. Rarely excellent sci-fi works barely勉强 (barely) match mediocre television standards, and still cannot compete with streaming media on information transmission dimensions.
Eli Goldfine: Compared to books, which medium do you respect more for information transmission efficiency?
Sam Bankman-Fried: Dense audiovisual media drastically outperforms print text in information throughput per unit of time. In an outstanding film or deep series, visuals, micro-expressions, sound design, editing pacing, and dialogue transmit dozens of simultaneous cognitive dimensions to your brain in a single second. This density of emotional and intellectual resonance is utterly incomparable to the sluggish efficiency of one-way printed text.
Effective Altruism (EA) and Quantitative Metric Boundaries: QALYs and Reality Distortion
Eli Goldfine: Within the EA community, people habitually use strict mathematical formulas and Quality-Adjusted Life Years (QALYs) to quantify expected utility per charitable dollar. Now incarcerated, how do you view this purely quantified charity philosophy?
Sam Bankman-Fried: Pure theoretical expected value is elegant in mathematics, but reality's complexity lies in the fact that metrics you choose to quantify are often not humanity's ultimate core values.
When you simplify complex realities into a single-dimensional number solely for mathematical measurability, you easily fall into a blind-men-touching-an-elephant trap. Donating mosquito nets against malaria has easily measurable costs and short-term returns; funding institutional building, foundational science breakthroughs, or preventing civilizational catastrophes involves long-term impacts filled with unquantifiable parameters. Abandoning profoundly civilization-shaping attempts merely because they resist precise quantification is itself the greatest betrayal of rationality.
Eli Goldfine: Does this mean over-quantification ironically becomes a dogmatic trap for effective altruism?
Sam Bankman-Fried: It absolutely does. When people become obsessed with mathematical verifiability, they instinctively reject truly disruptive, highly uncertain grand propositions. A healthy evaluation framework must acknowledge human cognitive boundaries, dynamically balancing quantitative precision with intuitive big-picture格局 (vision).
Philosophical Interrogation of the 51/49 Coin Flip: Pure Expected Value vs. Irreversible Ruin and Diminishing Marginal Returns
Eli Goldfine: I want to explore a widely known thought experiment. Economist Tyler Cowen once posed this to you: If a coin flip gives a 51% chance of doubling global utility, and a 49% chance of destroying the world, would you toss it? Your original answer was yes. Now imprisoned, do you still adhere to this pure expected-value logic?
Sam Bankman-Fried: This is a classically compelling philosophical and probabilistic puzzle. If strictly bounded within pure mathematical assumptions and linear expected utility functions, the mathematical answer is obvious: 51% multiplied by 2 exceeds 1. Expected value is positive; any rigid expected-value maximizer must logically accept this gamble.
But in the real world, why does this thought experiment trigger visceral rejection from most normal people? Because it ignores fundamental constraints in reality that cannot be simply abstracted away.
Eli Goldfine: What are those ignored constraints?
Sam Bankman-Fried: First, ruin is absolutely irreversible. In probability games, if you harbor even a slim chance of going permanently to zero and dropping out of the game, you forfeit all opportunity to enjoy subsequent positive-expectancy gains.
Second, utility is never strictly linear; it exhibits profound diminishing marginal returns. Doubling global utility delivers incremental happiness to civilization far outweighed by the infinite agony inflicted by a 49% chance of total human annihilation. Blindly applying an extreme model stripped of diminishing returns and bankruptcy protection to reality invariably breeds dangerously skewed judgments.
Eli Goldfine: Would you support slowing technological progress if it traded lower destruction probabilities?
Sam Bankman-Fried: In theoretical models, delaying human entry into superintelligence by a century—even if it lowers civilizational destruction probability by merely 1%—is mathematically unquestionably worthwhile when viewed against cosmic timescales spanning billions of years. The universe's future stretches infinitely, while civilization once zeroed ends forever.
High-Expected-Value Gambles in Reality: Would You Toss 51/49 for Personal Freedom?
Eli Goldfine: What if this gamble involves chips intimately tied to you? Suppose a coin flip offers a 51% chance of your无罪 (acquittal) release tomorrow, and a 49% chance of doubling your sentence or imposing the death penalty. Would you toss it?
Sam Bankman-Fried: This vividly illustrates why real-world utility functions are non-linear. Faced with core assets like personal liberty and life, human aversion to catastrophic downside risk is astronomical.
If released tomorrow, my life naturally improves dramatically; but if the price carries nearly a fifty-percent probability of enduring prolonged or complete deprivation of life, that risk premium is unconscionable. This is precisely why complex real-world decisions absolutely cannot be rashly gambled on based solely on a rough expected-value number.
Eli Goldfine: Many argue your past investments and business decisions were direct products of this extreme expected-value philosophy. Do you agree with that assessment?
Sam Bankman-Fried: I believe outsiders oversimplified and caricatured my thought processes. In routine trading, we meticulously hedged and managed risk exposure; but the moment an extreme crisis erupts, theoretical model safety buffers vaporize instantly. Reality isn't an ivory-tower Monte Carlo simulation; black swan destructive forces vastly exceed mathematical equation expectations.
The Endgame Interrogation: Was It Really Worth It? Decision Distortion in Crises and the Victor's Narrative
Eli Goldfine: We're down to our final minutes. I want to ask the most soul-striking question: Considering FTX's ultimately shocking collapse, but simultaneously, the FTX Future Fund early funded massive top-tier tech and safety projects including Anthropic. Looking back at the entire rise and fall, was it truly worth it macroscopically?
Sam Bankman-Fried: No. I desperately want to self-deceptively give an affirmative answer, but I cannot lie to my conscience and say it was worth it.
I believe it originally should have been. Had I not completely botched the situation at the final hour, had those risk firewalls not shattered, it should have propelled the world forward as a massive catalyst. If posterity extracts correct lessons from this disaster, perhaps it retains a sliver of warning value. But realistically, most people failed to learn anything.
Eli Goldfine: What exactly should people have learned?
Sam Bankman-Fried: The core lesson boils down to one thing: the profoundly distorted decisions humans make in extreme panic when crises actually land. Under life-and-death pressure, people rapidly lose judgment on what truly matters; all rules instantly disintegrate.
And once the dust settles, history is forever written by victors. Every causal detail gets reshaped, distorted, and simplified through the victor's filter. For those standing outside the storm, piercing through that narrative fog authored by winners to restore underlying mechanics is virtually impossible.
Survival Guidelines for the New Generation of 14-Year-Olds: Don't Go to Wall Street for Quant Trading, Dive into AI
Eli Goldfine: As a 14-year-old facing the incoming superintelligence era, if you want to create maximum positive value in this world, what is your core advice?
Sam Bankman-Fried: Ten years ago, the standard EA community answer handed to you would be "Earning to Give": go to Wall Street, execute high-frequency trading, rake in billions, and donate the wealth to charity.
But the era has completely shifted. At this precipice approaching singularity, Wall Street high-frequency trading is no longer the premier battlefield. The window between now and AI's full ascension is narrowing sharply; the value of deeply understanding and harnessing machine intelligence rises exponentially.
Whether diving into frontier AI alignment and safety research, or locating a single node of traditional infrastructure and ruthlessly automating it with AI, plunging straight into the core of machine intelligence is the only truly worthwhile battlefield demanding your absolute commitment this era.














