Compiled by: Shenchao TechFlow
Original link: https://www.techflowpost.com/article/35581
Accompanied by the automatic voice prompt from the federal prison that "the call is coming from a federal detention facility," former FTX founder Sam Bankman-Fried (SBF) received an 82-minute high-energy deep connection with 14-year-old technology blogger Eli Goldfine at the Metropolitan Detention Center in Brooklyn (MDC Brooklyn). In this conversation, which is regarded as the most information-dense conversation in recent years, SBF, who is behind bars, rarely broke his silence and comprehensively reviewed his cognitive map, investment history and extreme philosophy.
From the behind-the-scenes game that led FTX Ventures to inject $500 million into Anthropic, to the radical deduction of hundreds or even trillions of dollars in the future AI landscape; from the cold examination of traditional software moats and the Coinbase brand value, to the philosophical questioning of Tyler Cowen's "51/49 coin toss to destroy the world or multiply the benefits" and personal freedom; to the final moment when he confessed "FTX The outcome was not worth it. When the crisis broke out, all reality was distorted by the winner's filter." SBF used the most naked thinking deduction to present the complete spiritual world of a genius trader after he fell from the altar.
Guest: Sam Bankman-Fried (Founder and former CEO of FTX, currently serving time in federal prison; core practitioner of effective altruism)
Host: Eli Goldfine (14-year-old host of The Super Cycle, a tech and predictions podcast, 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
Broadcast date: November 20, 2026 (Beijing time: 2026-11-21)
Summary of key points
The biggest torture in prison is the isolation of cutting-edge information: SBF admitted that thinking about the case in prison has no impact on the outside world. What really pains him is the inability to follow up in real time the measured performance of large language models, code details, market conditions, and business due diligence of start-ups. The lack of cutting-edge micro input makes effective thinking extremely difficult.
The story of Anthropic’s early $500 million bet: FTX Ventures almost took over Anthropic’s entire Series B financing at the time. The reason why more computing power funds were not injected is because too much cash can easily induce corporate squandering (such as blind buying), and the team is unwilling to excessively dilute the equity. Blind expansion will put the cart before the horse.
The AI industry will reach a dominance of tens of trillions of dollars: AI will surely become the backbone or even the dominant sector of the global economy in the next ten years; since there are only a few top laboratories that can compete for the throne (Anthropic, OpenAI, DeepMind, xAI), it is completely logical for the leading institutions to enjoy trillion-dollar valuations in terms of economic deduction.
Traditional software companies have not been completely sentenced to death by AI: physical industries (such as steel and mining) in the physical world are lagging behind in the impact of AI; for pure software and financial institutions, even if competitors use AI to copy the same functions at very low cost, they cannot easily overturn the first mover, because the real moat of Coinbase and Robinhood is unbreakable brand trust.
p(doom) is evaluated in a complex gray range of about 25%: SBF believes that the probability of human beings being destroyed or encountering an existential disaster due to AI is between 10% and 50% (median is about 25%); the form of disaster is not a simple out-of-control nuclear weapon, but more likely to be a complex gray scenario, such as the complete marginalization and elimination of non-uploaded groups after the uploading of human consciousness.
The political prediction market was an unfinished ambition of FTX: the prediction market is, in essence, a futures contract with specific settlement rules; FTX had carried out hundreds of millions of dollars of U.S. election predictions in the offshore market in its early days, but was restricted due to its failure to obtain a license from the U.S. Commodity Futures Trading Commission (CFTC). The outbreak of Kalshi and Polymarket verified the explosive power of this track.
Extremely skeptical of the information throughput efficiency of paper books: SBF reiterates its famous view that the vast majority of books should be compressed into six natural paragraphs of blogs; in contrast, the information throughput and expression dimension of audio-visual streams (such as Netflix) per unit time are much higher than books that are dragged down by a lot of empty words.
Extreme expectations and the reality of the 51/49 coin toss are shattered: Under the theoretical pure linear expectation, the bet of 51% doubling and 49% destruction of the world is mathematically valid; however, in the real world, destruction has an irreversible cost of extinction, and the real utility is accompanied by diminishing margins; faced with whether you can bet yourself out of prison immediately with 51/49, the personal risk utility function determines that the theoretical gamble cannot be simply copied into reality.
A painful review of the ending of FTX: When asked "Is it all worth it?", SBF made it clear that "You can't say it was worth it without conscience. It should have been worth it if it hadn't been messed up in the end." When a crisis comes, people can easily lose their rationality and core judgment, and then all history will be reshaped and distorted by the filter of the winner.
Survival rules for the new generation of 14-year-olds: Don’t be obsessed with going to Wall Street to do high-frequency trading of “Earning to Give”; the top priority is to go all out to understand AI, devote yourself to AI security, or find a way for AI to truly automate infrastructure.
Wonderful views
On the brutal barriers to information disparity behind bars: “If you want to create value at the cutting edge, you need to know hundreds of tiny details: what tasks the big model excels at and where it still sucks, what’s being discussed at the forefront of the industry, how the microstructure of the market has changed. And within the walls of federal prisons, accessing these details is an almost impossible task.”
Regarding Anthropic's valuation and moat: "If you accept that AI will become a core sector that dominates the world's tens of trillions of dollars in the next ten years, and there are only a handful of top core laboratories in the world that can sit at this table, then it is completely in line with the market value of trillions of dollars for each leading company. Anthropic's strength in code generation and software engineering tools is still among the best in the world."
Regarding the real barriers between brand and technology: "Many programmers think that technology is everything, but if you look at Coinbase or Robinhood, why are they worth tens of billions of dollars? It is definitely not because they have exclusive algorithms that cannot be copied, but because they have the brand trust that retail users across the United States have. Even if competitors can use AI to copy an identical set of software at almost zero cost, they cannot buy this kind of brand equity."
On the radical criticism of paper books: "I still believe that if you write a book, it often means you did it wrong. It should have been a long blog post with only six natural paragraphs. The information throughput of books is too low, full of long useless narratives; in the same unit of time, high-density visual and audio-visual input can convey much more effective information."
Realistic reflections on the 51/49 coin toss: "In a pure ivory tower expected value theory, you can deduce that doubling the global utility will give you a 49% probability of destruction; but in the real world, there is an irreversible systematic risk of zeroing, and any positive utility has its limit of diminishing marginality. Once you incorporate the fatal cost of destruction into the model, you simply don't dare to throw down most coins in reality."
Regarding the final conclusion of the rise and fall of FTX: "I want to delude myself and comfort myself that it was worth it, but I can't say this without conscience. It should have been worth it, if I didn't completely screw up everything at the last moment. At the moment of extreme crisis, panic will distort all suggestions and decisions, and after the smoke clears, all historical narratives will inevitably be completely reconstructed by the winner's position."
text
Prison life and cognitive barriers: a black box world that loses microscopic details
Eli Goldfine: Welcome to The Super Cycle, I'm Eli Goldfine. Today we are honored to be connected with Sam Bankman-Fried (SBF). Before this, my team and I coordinated and prepared for several months. This is also the most information-dense in-depth interview since the founding of our podcast. Hello Sam, can you hear me?
Sam Bankman-Fried: Okay, hear me clearly.
Eli Goldfine: Great to be with you. I've seen a lot of your recent interviews, and the overwhelming focus has been on the collapse of FTX and your current day-to-day life in prison. In prison, you have plenty of time to be alone and think. What issues do you usually spend the most energy thinking about?
Sam Bankman-Fried: It's a natural question. Obviously, I do spend a fair amount of time reviewing my own cases, going over what happened, and thinking about what else I could do about it. But I know very well that these reflections do not actually have much incremental value to people in the outside world.
I try to turn my energy toward thinking about what else I can do in the future. But productive forward thinking is extremely difficult in a prison setting.
Eli Goldfine: Why is forward thinking so difficult in prison? Is it because there is a lack of key information?
Sam Bankman-Fried: It's not a lack of a single piece of information, but a lack of countless small but crucial inputs. If you want to stand at the forefront of human cognition and business to deduce the future, you must master massive dynamic details: How do cutting-edge large language models evolve? In which specific tasks do their evaluation curves break through human limits, and in which areas are they still poor? You get to see these technologies in action and talk to the people who are out there building these models every day.
The same is true from a financial perspective. You must keep an eye on the market every day, observe the flow of funds and transaction structures, build a network, communicate with emerging start-up teams, discuss their business models and conduct in-depth due diligence. These are the tasks I would do every day if I were outside, but within the high walls of a federal detention facility, obtaining every piece of information becomes extremely difficult or almost impossible.
Eli Goldfine: If information can be obtained in prison, which dimensions of information do you think are most valuable for your deduction of the real world?
Sam Bankman-Fried: Again, it’s not a single breaking news, but thousands of micro variables. For example, the architectural evolution of cutting-edge models, the actual implementation threshold of code generation, the cash flow runway of start-ups in real business environments, the real purchasing intentions of customers in different industries, etc. In the absence of these micro-detailed inputs, macro-frameworks derived purely behind closed doors are often extremely fragile.
Eli Goldfine: You obviously can't just think about the case all day long. Beyond the case, what broader issues are you thinking about?
Sam Bankman-Fried: Yeah, but a lot of that thinking is still highly constrained by the physical environment that I'm currently in, because that's the only world where I can get direct first-hand feedback right now. Not that it matters that much, but I do spend a lot of time thinking about the criminal justice system. Not just for my personal experience, but for the absurd logic displayed by the entire judicial system on a general level.
Reflection on the federal criminal justice system: fictional charges, indicator assessments and reverse incentives
Eli Goldfine: Outside of the context of your individual case, what are the specific common judicial issues that you've observed?
Sam Bankman-Fried: As the saying goes, with a hammer in your hand, everything looks like a nail. The exact same pattern applies to the judicial system. If you set up a federal task force to fight a particular crime, that task force will go crazy looking for the source of the crime, whether or not the crime actually exists.
Once the size of the working group's staff and budget exceeds the actual volume of this type of crime in real society, it will inevitably start to create false cases to catch behavior that does not constitute a crime at all. Over the past few decades, this distortion has been widely exposed in the federal anti-drug system, and over the past decade, this ethos has permeated and spread to every area of federal justice.
Eli Goldfine: Like what specific areas?
Sam Bankman-Fried: For example, the huge federal cybercrime investigation agency, which was originally established to combat extreme violations, has now begun to reach out to private text message chats and normal interactions between lovers. Countless completely innocent people have been prosecuted and arrested simply because prosecutors need to use these cases as performance evaluations for promotions and salary increases. The case of former Honduran President Juan Orlando Hernandez is a prime example of the outrageous results that can occur when the prosecution machinery is completely out of control.
Eli Goldfine: Do you still believe that white-collar crimes should be punishable by prison time?
Sam Bankman-Fried: Not convinced at all. Putting white-collar criminals behind bars does nothing positive for society. The only thing it can do is take decades away from someone's life and impose a high financial burden on taxpayers to run prisons. If you're doing it purely for revenge, it does serve its purpose; but I don't think this system can reform anyone in any way.
Eli Goldfine: What do you think is the most reasonable way to deal with someone who actually committed financial fraud?
Sam Bankman-Fried: It depends on the specific factual details of the case. But in the vast majority of commercial and financial cases, civil litigation procedures are far more sophisticated and rational than criminal trials. Civil remedies can solve problems by imposing large fines, modifying business contracts, restricting business qualifications, and depriving assets.
More importantly, as a modern society, we must learn to reduce blind overconfidence. We often think that we can accurately see the truth, and we think that those who are punished must be bad people, but in reality, many people who are put on trial are simply because they have become the most convenient scapegoats used by a certain force to deflect conflicts.
Eli Goldfine: Is improving the justice system a priority for the Effective Altruism community?
Sam Bankman-Fried: From a utility expectation perspective, this area is actually very complex. The social value of judicial reform is undoubtedly huge, but it faces extremely solid political resistance and institutional inertia. When an interest group forms a huge industrial closed loop by imprisoning people, it is extremely difficult for external forces to promote changes at the bottom.
Eli Goldfine: In your opinion, what is the low-hanging fruit in improving the criminal justice system?
Sam Bankman-Fried: The easiest thing to push, and the thing that would deliver huge immediate economic dividends, is a complete reexamination of pretrial detention and sentencing benchmarks for nonviolent crimes. The current system is filled with impoverished people who are detained for long periods of time because they cannot afford meager bail, and working-class people who are sentenced to decades-long sentences for drug-related crimes. Providing them with formal market-oriented employment training and community corrections, rather than directly imprisoning them and completely rupturing their social relations, is the most cost-effective reform approach.
The endgame of the AI industry: Trillion-dollar valuations and the poker table of a few oligarchs
Eli Goldfine: Let’s talk about cutting-edge technology. Anthropic is preparing for a public offering recently, and expectations for its valuation are sprinting toward hundreds of billions or even trillions of dollars. As an early lead investor in Anthropic’s $500 million Series B investment that year, what do you think of the current valuations given to these large model labs?
Sam Bankman-Fried: I think they're definitely worth the price and probably even more in the future.
Deduced from macro business logic: Throughout the history of human economics, the total market value of all core pillar sectors that truly support the operation of human civilization ranges from trillions to tens of trillions of dollars. In the next five to ten years, AI will most likely grow into the world's most core and even the most dominant economic industry.
Eli Goldfine: If the entire AI sector can reach a market value of tens of trillions of dollars in the future, how will this value be divided up by the market?
Sam Bankman-Fried: Assuming that two-thirds of the total market value of tens of trillions of dollars in the future will be reserved for innovative applications and emerging companies that have not yet been born, then the remaining sectors will still have a scale of at least more than ten trillion dollars shared by existing top and bottom model companies.
However, if you count carefully, how many top laboratories in the world are truly qualified to sit on this core base card table? OpenAI, Anthropic, Google DeepMind, plus xAI, you can count them on the fingers of one hand. When a multi-trillion-dollar industry is eventually divided up by a few top oligarchs in single digits, the inevitable result is that the intrinsic value of each leader at the top must be measured in trillions of dollars. Anthropic is undoubtedly among the top two in the world.
Eli Goldfine: Why do you think so highly of Anthropic? Is it because of how the Claude series models perform on benchmarks?
Sam Bankman-Fried: Although running scores are important, the outside world can also see Claude’s strong reputation in daily conversations and mass consumer applications. But Anthropic's core asset that really sets the difference is its absolute dominance in software engineering and professional development tools. The development kits and code understanding capabilities they have built for engineers are among the best in the world. They are the core working machine for the entire digital society to shift to AI automation in the future.
Eli Goldfine: What do you think the world will look like if we enter into a complete superintelligence arms race in the next ten years?
Sam Bankman-Fried: The distribution of productivity across society will be completely reshaped. In the past few years, although large models have evolved exponentially, because their absolute capabilities are still below the human average, the social order seems to be stable. But once a certain critical point is crossed, machines can produce hundreds of times faster than humans in areas such as programming, law, financial modeling, and drug discovery. The speed of wealth creation will be greatly accelerated, but how to distribute the dividends brought by these computing power will become the focus of the most intense geopolitical game in the world.
The whole story of Anthropic’s early $500 million investment: computing power arms race and capital trap
Eli Goldfine: Many people have been curious. FTX Ventures almost took over Anthropic’s entire Series B financing back then. Since you were so sure of the future of AI at that time, why didn't you write a bigger check, such as directly injecting $1 billion or even $2 billion to allow them to purchase unlimited graphics processing units (GPUs) and build data centers?
Sam Bankman-Fried: There are two key considerations here. First of all, from the perspective of the founder team, any excellent team does not want to overly dilute the control rights and equity ratio of their company in the early stage. They are only willing to raise the amount of funds necessary for the current stage of development.
Secondly, in the history of business, many companies have died due to lack of funds, but it is also not uncommon for companies to be suffocated by excessive cash flow and then suffer backlash from their own capital. After raising astronomical amounts of funds, many start-up teams often develop a blind anxiety of "I have to spend this money." They then squander huge sums of money on meaningless marketing or blind expansion, and ultimately lose their respect for the core technology.
In the early stages of building a large model, iteration of the model architecture and algorithm verification require a steady R&D cycle, which cannot be achieved by simply stacking computing power. Anthropic planned the use of funds very rationally at the time, and the funds we invested perfectly met their core requirements for launching the next generation of large model training.
Eli Goldfine: How was FTX Ventures functioning at that time? What role do you personally play in specific investment decisions?
Sam Bankman-Fried: Many of the early cases were discovered by our team members in their daily business, including a large number of early crypto-native investments, many of which achieved several times of asset appreciation every year. Anthropic is similar to Solana. They were both disruptive projects that we came into contact with by chance.
As our scale expanded, we attracted professional partners to form a formal investment structure. For ordinary cases, I tend to fully trust the analysis team responsible for due diligence; but for strategic heavy bets like Anthropic, I will follow up in depth on a weekly or even more frequent basis to explore how we can assist them in the underlying computing power, finance and resource network.
Eli Goldfine: When you were in contact with the Dario Amodei team, what was the biggest impression they made on you?
Sam Bankman-Fried: Extreme rigor and focus. When Dario and his team spun out from OpenAI, they had a clear goal of building a cutting-edge security model with high interpretability and alignment guarantees. They will not make fancy marketing presentations to win short-term attention. Every capital expenditure is extremely restrained and targeted at key computing clusters and engineering talent recruitment. This kind of engineering culture is extremely rare in today's impetuous Silicon Valley.
The existential crisis and brand moat of traditional software companies: Why can’t AI kill Coinbase?
Eli Goldfine: With the explosion of cutting-edge model capabilities, many traditional software companies are facing huge existential anxiety. Do you think that traditional software companies that have not deeply integrated AI have no way out?
Sam Bankman-Fried: It's not so absolute. First of all, there are quite a few industries in the real economy that have natural physical defense buffers against the penetration of AI, such as steel smelting and underground mining. No matter how smart the AI is, it cannot dig out ore overnight.
Secondly, for purely digital software, it is not just the efficiency of the underlying code that determines the outcome. Many people seriously underestimate the absolute moat of the "brand". Suppose a competitor uses the latest large-model tools to clone a software product with almost identical functions at a very low cost. Does this mean that users will abandon the original product? Absolutely not.
Eli Goldfine: Like Coinbase or Robinhood?
Sam Bankman-Fried: Exactly. Why do Coinbase and Robinhood enjoy market caps in the tens of billions of dollars? Is it because they have top-secret matching engine code that other teams can never copy? Of course not. Their most valuable asset is brand recognition, which is the trust and endorsement that thousands of ordinary retail investors across the United States speak out when they want to buy or sell digital assets.
Even if the new team can use AI to copy a front-end and transaction system at almost zero cost, it will not be able to buy the brand reputation that has been accumulated for many years in a short period of time. Brands often allow traditional companies to maintain long-term survival resilience under the impact of homogenization of underlying products.
Eli Goldfine: So what kind of traditional software companies will be eliminated the fastest?
Sam Bankman-Fried: Those middleware software that relies solely on stacking functions, has neither brand mind occupation nor deep data network effects. If your product is just a simple work order management or form filling tool, and there is no emotional binding between users and your brand, then AI-native tools will completely crush it in a very short time and at one-tenth the cost.
The Real Alpha of Technology Investment: Minimalist AI Native Team vs. Rubbish Companies with Concepts
Eli Goldfine: So where do you think the real excess returns (Alpha) in the current technology investment field are hidden?
Sam Bankman-Fried: The real alpha is the ability to discern through bubbles. Half of the startups on the market currently use AI labels on the covers of their business plans. Even a company that sells cultural shirts claims that it is using AI to design patterns. These are essentially industrial garbage.
At the same time, truly huge opportunities lie in a very small number of native teams with only two or three core members: they rely entirely on AI tool chains to reconstruct enterprise workflows that traditionally require hundreds or thousands of people to maintain with incredibly streamlined human efficiency, and have rapidly grown into a super species with a valuation of tens to hundreds of millions of dollars. Distinguishing which is a real outbreak and which is taking advantage of the situation is the core watershed in current investment.
Eli Goldfine: If you were out there running free money today, would you go back into this kind of investing?
Sam Bankman-Fried: Absolutely. If I were outside, I would be very active in finding and heavily supporting these two or three super species. They do not have the bureaucratic internal friction of traditional companies, and each member can call the output of an entire engineering department through large models. The explosive power and profit margins brought about by this streamlined human efficiency are completely unimaginable in traditional organizational forms.
Biomedicine and longevity technology: regulatory constraints and acceleration of AI R&D
Eli Goldfine: Let’s talk about biotech. FTX also funded many anti-aging, longevity and gene therapy projects in its early days. How do you see progress in the biomedical field?
Sam Bankman-Fried: Biotechnology is a field full of huge potential, but the biggest bottleneck it faces is not the intelligence of scientists, but the outdated regulatory system. The clinical approval process of the U.S. Food and Drug Administration (FDA) is extremely long and costly, with a new drug often costing billions of dollars and taking up to ten years to be launched on the market.
Many early-stage therapies that were originally promising to save lives were stuck in the laboratory because they could not raise astronomical funds to complete the third phase of clinical trials. We hoped to provide early catalytic funding for these high-risk, high-return cutting-edge attempts.
Eli Goldfine: Will AI change that?
Sam Bankman-Fried: AI has demonstrated disruptive power in protein structure prediction, molecular docking and compound screening. AI can compress preliminary laboratory screening that would have taken years or even decades into just a few months. But ultimately, all drugs must still pass human clinical testing in the physical world. If outdated and rigid regulatory processes are not fundamentally reformed, pure technological acceleration will still be greatly delayed by offline administrative barriers.
Eli Goldfine: How do you think the clinical approval process should be optimized for rare or fatal diseases?
Sam Bankman-Fried: More flexible informed consent and parallel channels should be introduced. For those patients who are terminally ill and for whom modern medicine has no solution, if they are fully informed and voluntarily try innovative treatments in Phase I or Phase II clinical trials, regulatory agencies should not brutally deprive them of their right to live. Dynamic tracking through real-world data (RWD) is far more ethical than mechanically waiting for ten years for phase III double-blind trials.
Politics, Regulation, and Reflections on the Biden Administration: Crypto Hostility and the Democratic Party’s Disorientation
Eli Goldfine: Let’s talk about politics and macro regulation. The Biden administration has taken a very aggressive stance on technology and financial regulation over the past four years, especially the Securities and Exchange Commission (SEC), which has launched comprehensive enforcement actions against the cryptocurrency industry. How would you evaluate the Biden administration's regulatory performance?
Sam Bankman-Fried: I think the Biden administration's performance on this is extremely disappointing.
Regulators have demonstrated a deep bureaucratic hostility not only to the crypto industry, but to the entire field of technological innovation. Instead of committing to establishing a clear, enforceable legal framework that promotes innovation, they use endless law enforcement raids to crack down on local companies that operate in compliance, which directly leads to the flow of a large number of top scientific and technological talents and innovative capital to offshore markets.
Eli Goldfine: You've made huge donations to political campaigns. Why do you think there is a disconnect in the political world’s understanding of the tech industry?
Sam Bankman-Fried: Many politicians and regulators simply don’t understand the underlying logic of technology. They only care about political narratives and power maps. In many policy formulations, the basis for decision-making is not which option is most beneficial to the social economy, but which option can bring the most votes and political chips. I think the Democratic Party has seriously lost its way on a lot of economic and technology issues over the past decade.
Eli Goldfine: Should technology companies slow down to stay safe in the face of international technology competition?
Sam Bankman-Fried: This is an extremely dangerous gambling trap. If a company that adheres to the safety concept unilaterally slows down, while other global competitors or even non-friendly countries are running at full speed, then unilateral self-restriction will not make the world safer, but will directly hand over control of cutting-edge technology to competitors who don't care about safety and ethics at all. The ideal situation is for all major parties to reach a substantive security agreement, but in real geopolitics, this is extremely difficult.
The Nature and Explosion of Political Prediction Markets: From FTX 2020 Election Contracts to Kalshi and Polymarket
Eli Goldfine: Let’s talk about financial innovation. Kalshi’s valuation has skyrocketed recently, with Polymarket seeing staggering trading volumes during global political elections. I heard that in the early days of FTX, one of your original ideas was to create a global political prediction market?
Sam Bankman-Fried: From the perspective of the underlying logic of financial engineering, the so-called prediction market is essentially a futures contract that liquidates cash based on the results of specific real-life events. They are not fundamentally different from Bitcoin or crude oil futures.
During the 2020 U.S. presidential election, FTX International ran the most active Trump-Biden victory prediction contract on the entire network, and the transaction size at that time reached hundreds of millions of dollars. We have been eager to expand this market from day one, but the biggest obstacle has always been regulatory licenses. Due to the lack of a local license from the U.S. Commodity Futures Trading Commission (CFTC), we strictly prohibited U.S. users from participating, and centralized exchanges mandated comprehensive real-name authentication (KYC).
Eli Goldfine: Why did Kalshi and Polymarket explode this cycle?
Sam Bankman-Fried: They took two different paths. Kalshi spent many years obtaining a local compliance license through legal proceedings and formal procedures; while Polymarket relied on offshore and decentralized structures to gather huge liquidity around the world. This fully proves our judgment back then: Humanity’s demand for pricing of major historical events is universal and real.
In addition to prediction markets, tokenized equities still have huge growth potential. The underlying settlement system of the traditional stock trading market is extremely outdated and inefficient. The 7×24-hour seamless settlement system based on the blockchain fully possesses the advantage of dimensionality reduction in terms of infrastructure experience.
AGI critical point and p(doom) evaluation: 25% extinction probability and the gray dilemma of consciousness uploading
Eli Goldfine: Many people are discussing when artificial general intelligence (AGI) or even superintelligence (ASI) will arrive. What is your personal timeline expectation for AGI?
Sam Bankman-Fried: I don’t think the arrival of AGI is going to be a clear, black and white, single point discrete event. It is more likely to be a gradual replacement process that continues to evolve.
For a long time in the past, although AI's computing power has continued to double, its intelligence level has been far below the human baseline and has had almost no substantial impact on the real world. But when its intelligence gradually approaches that of humans and breaks the critical point in specific vertical fields, qualitative changes will spread at an extremely fast speed. We’re already seeing this trend in software development, advanced mathematical derivations, and game theory strategies. In the next three to five years, key sectors of human society will be deeply taken over by AI one after another.
Eli Goldfine: In the effective altruism community, AI safety and the probability of human extinction have always been the core debate issues. You have never publicly given your personal p(doom) (the probability that AI will cause human extinction or complete destruction) specific value. Can you give an exact judgment today?
Sam Bankman-Fried: 如果必须给出一个数值,我认为它低于 50%,但高于 10%,大致落在 25% 左右的区间。
但有一点必须厘清:在我的评估模型里,灾难性情景往往不是好莱坞电影里那种具有清晰敌意的失控人工智能突然掌握了核按钮并向全人类宣战。现实中更大概率发生的是极其复杂的灰色伦理困局。
Eli Goldfine: 比如怎样的灰色困局?
Sam Bankman-Fried: 比如意识上传(Mind Uploading)。如果未来技术发展到了人类可以把大脑意识全盘数字化上传至云端的阶段,那些完成数字化永生的智能体在算力和进化速度上会以几何级数超越肉身人类。在那种情境下,留在物理世界的原生人类是否会被彻底淘汰并边缘化?
这种导致肉身智人物种在事实上走向消亡的过程,到底算不算一场人类文明的灾难?不同的人会给出截然相反的哲学答案。一旦你把这些充满伦理模糊地带的演化路径纳入考量,对生存风险的概率判断就会在 20% 到 60% 之间剧烈摇摆。
AI 智能体与未来工作流:十倍于人类的超级研究助理
Eli Goldfine: 如果未来出现了一个在逻辑推理和决策期望值上全面超越你的超级智能助理,你会愿意把个人的决策权全盘交由它代理吗?
Sam Bankman-Fried: 这要看它的优化目标函数到底是什么。它是为了最大化它自己的存活概率,还是为了最大化所有感知生命的福祉,抑或是忠实履行我赋予它的委托目标?
如果它的核心对齐目标是作为一个忠诚、细致且理性审慎的研究助手,能够以人类百倍的速度在一分钟内完成原本需要一整天的高强度案头尽调与推演,那我绝对愿意把大量具体的执行和分析任务托付给它。但终极的价值取向和伦理底线,必须始终锚定在人类自身的核心诉求上。
Eli Goldfine: 这种超级助理会如何颠覆金融与投资行业?
Sam Bankman-Fried: 传统投资机构中由成百上千名初级分析师构成的金字塔结构会瞬间崩塌。一个合伙人配合几套专门针对财报交叉核验、宏观数据清洗与舆情追踪的自主智能体,就能够完成过去整座投行大楼的产出。机构之间的胜负将不再取决于雇佣了多少名校毕业生,而取决于谁对智能体的调优策略和底层逻辑理解得更深刻。
为什么极度怀疑纸质书?信息吞吐量与视听媒介的降维打击
Eli Goldfine: 聊一个很有趣的生活议题。许多聪明人都以博览群书为荣,但我记得你曾经说过一句非常著名的反常识观点:「我极度怀疑书的价值,绝大多数书本该写成一篇只有 6 个段落的博客」。今天身在狱中,你读了多少书?你的观点有改变吗?
Sam Bankman-Fried: 在狱中我确实没有太多选择,只能读了不少书,但这反而让我更加坚信我当初的判断是对的。
如果你写了一整本书,往往意味着你根本没有想清楚核心论点。一本数百页的著作里,真正有价值的核心洞见往往用几个自然段就能完全讲透,其余百分之九十的篇幅全是在为了凑齐书本厚度而堆砌冗长的案例、铺垫和无关紧要的叙事。
Eli Goldfine: 你在狱中主要读了哪些类型的书?有任何一本让你觉得值得一读吗?
Sam Bankman-Fried: 我读了大量的科幻小说。但说实话,绝大多数科幻小说也就是差强人意。如果我有条件,我宁愿去看 Netflix 上的硬核科幻剧集。极少数写得极好的科幻作品,或许勉强能达到一部平庸电视剧的水准,但依然无法在信息传输维度上与流媒体竞争。
Eli Goldfine: 相比书籍,你更认可哪种媒介的信息传输效率?
Sam Bankman-Fried: 高密度的视听媒介在单位时间内的信息吞吐量远胜纸质文字。在一部优秀的电影或深度剧集里,视觉画面、人物微表情、声效、剪辑节奏和台词在同一秒内向你的大脑并发传输数十个维度的信息。这种高密度的情感与认知共鸣,是效率极其低下的单向印刷文字根本无法比拟的。
有效利他主义(EA)与量化指标的边界:QALYs 与现实评估失真
Eli Goldfine: 在有效利他主义社区中,人们习惯用严格的数学公式和质调整生命年(QALYs)来量化每一笔慈善支出的期望效用。身陷囹圄之后,你如何看待这种纯量化的慈善哲学?
Sam Bankman-Fried: 纯理论的期望值在纯数学世界里是优雅的,但现实世界的复杂性在于:你所选择量化的指标,往往并不是人类社会终极关切的核心价值。
当你为了追求数学上的可衡量性,把复杂的现实问题简化为一个单一维度的数字时,你很容易陷入盲人摸象的陷阱。比如资助蚊帐防疟疾,它的成本和短期收益极其容易量化;而资助制度建设、基础科学突破或防止文明灾难,其长期影响在数学模型里充满了无法量化的参数。如果因为无法精准量化就放弃那些真正决定文明命运的深远尝试,这本身就是对理性最大的背叛。
Eli Goldfine: 这是否意味着过度量化反而成了有效利他主义的一种教条主义陷阱?
Sam Bankman-Fried: 确实如此。当人们过度痴迷于数学模型的可验证性时,他们会本能地排斥那些真正具有颠覆性、但充满不确定性的宏大命题。一个健康的评估框架必须承认人类认知的边界,在量化精密与直觉格局之间保持动态平衡。
51/49 抛硬币的极端哲学拷问:纯期望值 vs 不可逆毁灭与边际递减
Eli Goldfine: 我想跟你探讨一个广为人知的思想实验。经济学家 Tyler Cowen 曾向你提出过一个假设:如果有一枚硬币,正面抛出有 51% 的概率让地球的整体效用翻倍,反面有 49% 的概率导致地球彻底毁灭,你会不会选择抛下这枚硬币?当时你的回答是会抛。今天身在监狱中的你,依然坚守这个纯粹的期望值逻辑吗?
Sam Bankman-Fried: 这是一个在哲学和概率论上极具魅力的经典辩题。如果你把问题严格限定在纯粹的数学假设与线性期望效用函数(Linear Expected Utility)之内,那么数学上的答案显而易见:51% 乘以 2 大于 1,期望值为正,任何坚定的期望值最大化者在逻辑上都必须接受这个赌局。
但在现实世界中,这个思想实验之所以会遭到绝大多数正常人的本能排斥,是因为它忽略了现实世界中无法被简单抽象的根本约束。
Eli Goldfine: 这些被忽略的现实约束具体是什么?
Sam Bankman-Fried: 第一,毁灭是绝对不可逆的(Catastrophic Ruin)。在概率游戏中,如果你有一定概率彻底归零并永久退出游戏,你就再也没有机会去享受后续正期望值带来的收益。
第二,效用从来不是严格线性的,它存在强烈的边际效用递减。在现实中,地球效用翻倍带给人类文明的增量幸福感,远远无法抵消 49% 导致人类文明彻底湮灭所承受的无限痛苦。把一个剥离了边际效用递减和破产保护的极端模型直接套用到现实世界,往往会导致极其危险的判断偏差。
Eli Goldfine: 如果延迟科技发展能换取更低的毁灭概率,你会支持放慢脚步吗?
Sam Bankman-Fried: 如果在理论模型里,将人类跨入超级智能的时间推迟一百年,能够把文明毁灭的概率哪怕只降低 1%,那么站在宇宙未来数十亿年的宏观长周期来看,这一百年的停滞在数学期望上是绝对合算的。因为宇宙的未来无限漫长,而文明一旦归零就彻底终结了。
现实中的高期望值博弈:如果用 51/49 赌人身自由,你会抛吗?
Eli Goldfine: 如果把这个赌局换成与你切身相关的现实筹码呢?假设现在有一枚硬币,51% 的概率能让你明天立即无罪出狱,49% 的概率会导致你的刑期加倍甚至判处死刑,你会抛这枚硬币吗?
Sam Bankman-Fried: 这非常生动地说明了为什么现实中的效用函数不是线性的。面对人身自由与生命这种核心资产,人们对下行极端风险的厌恶程度是极其巨大的。
如果我明天出狱,我的生活当然会发生翻天覆地的改善;但如果代价是有近一半的概率承受更漫长甚至彻底剥夺生命的深渊,这种风险溢价是不可承受的。这也是为什么现实中的复杂决策绝不能仅凭一个粗糙的期望值数字就草率下注。
Eli Goldfine: 很多人认为你过去的投资和商业决策就是这种极端期望值哲学的产物,你认同这种评价吗?
Sam Bankman-Fried: 我认为外界把我的思考方式过度脸谱化了。在做常规交易时,我们极其注重对冲和风险敞口管理;但在极端危机爆发时,很多理论模型的安全垫会在瞬间被击穿。现实世界不是象牙塔里的蒙特卡洛模拟,黑天鹅的破坏力远远超出数学方程的预期。
终局拷问:这一切真的值得吗?危机中的决策变形与胜者叙事
Eli Goldfine: 我们只剩最后几分钟了。我想问一个最直击灵魂的问题:考虑到 FTX 最终令人震惊的崩塌,但与此同时,FTX Future Fund 早期资助了包括 Anthropic 在内的海量顶尖科技与安全项目。回顾整个起落,这一切在宏观上真的值得吗?
Sam Bankman-Fried: 不。我很想自欺欺人地给出肯定的回答,但我不能昧着良心说它值得。
我认为它原本是应该值得的。如果我在最后关头没有把局面搞砸,如果那些风险防线没有被击溃,它本该成为推动世界前行的一股巨大力量。如果后人能够从这场灾难中汲取正确的教训,或许它还能保留一丝警示价值,但现实中大多数人并没有在汲取教训。
Eli Goldfine: 人们真正应该吸取的教训是什么?
Sam Bankman-Fried: 核心的教训只有一条:当危机真正降临、局势彻底失控时,人类在极度恐慌中会做出多么变形的决策。在生死存亡的压力下,人们会以极快的速度丧失对真正重要事物的判断力,所有规则都会在瞬间分崩离析。
而在尘埃落定之后,历史永远是由胜利者书写的。一切事件的因果细节都会被胜利者的滤镜所重塑、扭曲和简化。对于身处风暴之外的人来说,想要穿透那层胜者书写的叙事迷雾去还原真正的机理,几乎是不可能的。
给 14 岁新一代年轻人的生存准则:别再去华尔街做量化,投身 AI
Eli Goldfine: 作为一个 14 岁的年轻人,面对即将到来的超级智能时代,如果想在这个世界创造最大的正向价值,你给我的核心建议是什么?
Sam Bankman-Fried: 放在十年前,有效利他主义社区会给你的标准答案是「赚取后捐赠(Earning to Give)」:去华尔街从事高频交易,赚取数十亿美元,然后将财富捐赠给慈善事业。
但时代已经彻底变了。在这个奇点临近的节点上,华尔街的高频交易已经不再是最重要的战场。从现在到 AI 全面腾飞之间的窗口期正在急剧收窄,深入理解并驾驭机器智能的价值正以指数级速度攀升。
无论是投身于前沿的 AI 对齐与安全研究,还是寻找某一处传统基础设施并利用 AI 将其彻底自动化,深入机器智能的核心,是这个时代唯一值得全力以赴的真正战场。
风险提示
根据中国人民银行等八部门最新发布的银发〔2026〕42号文件,本平台内容仅用于区块链行业技术与信息分享,不构成任何投资建议,不推广、不背书任何经营与投资行为。请读者严格遵守所在地区法律法规,谨防投资风险与诈骗。目前存在大量诈骗类仿冒账号,敬请警惕甄别。本平台所有内容未经许可,禁止转载、复制及商用。
北京蓝色海马文化传播有限公司 版权所有 京ICP备2023026470号




