Martin Shkreli：大药企的盲点、做空量子计算，以及监狱教会他的事
Shkreli 对量子计算的看空，既是技术判断，也是商业模式判断。 他认为，量子计算的真正优势，是把整数分解和椭圆曲线离散对数从指数时间降到多项式时间——不是排序，也不是通用计算。现有机器的运行频率只有几百或几千赫兹，在普通计算上却输得很惨；它们没有量子 RAM，而且按他的测试，“连300乘200都算不出来”。他说，企业在这个领域投入数亿美元，无非是在押注“总有一天可能会有点东西出来”。
主持人最有力的反驳——量子计算也许只是处在晶体管时代的起点——并未改变 Shkreli 的投资结论。 他承认这项技术最终可能奏效，但随即反问：投资者当年会不会因为打孔卡、真空管或铁氧体磁芯存储是必要阶段，就想持有它们？这些重要阶段的拥有者大多已经消失。他做空量子计算公司已经亏了数百万美元，但开玩笑说自己“赚到了数十亿美元的洞见”，同时承认，市场择时才是难点。
他认为自己最看好的算力颠覆方向是模拟光学计算，因为 AI 工作负载有“99.9%”的时间都在做矩阵乘法。 Shkreli 说，在 Nvidia 机器上做一次1,000×1,000的矩阵乘法，大约需要10亿次运算；而光学中的乘法可以瞬间完成。他正在研究光子核心，并预测20年内 H100、B100 和 B200 最终都会“被扔进垃圾桶”。
这种技术层面的看空，并不等于对 Nvidia 的明确做空：Shkreli 持有半导体股票，仍在押注 AI 上行。 Jevons 悖论可能会吞噬效率提升带来的收益，但他坚持认为，一个效率高出1,000倍的 transformer 对 Nvidia 依然“不是好事”；之后需求是否会扩大100万倍，仍未可知。他的警告来自历史：如果身处2000年，他大概也会买入 Yahoo 和 Cisco。
在 Shkreli 看来，大药企的盲点不是缺乏科学创意，而是错定了被弃置分子的价值。 他花200万美元买下一款 Bristol-Myers 的药，并称同一家公司如今已经上市、价值60亿美元；sparsentan 则从高血压适应症转向 FSGS，最终让约3万名此前从未获得药物治疗的人拿到了获批疗法。“我也可以做个好人”，他说，同时又称这“是赚到10亿美元的好方法”。
Shkreli 否定近期 AI 药物研发的炒作，但对机器智能市场依然极度看好。 他称 AI 在5年内开发出药物是“完全胡说”，并表示未来10年内不会出现一款 AI 相关药物；但他认为，对智能和劳动力的需求几乎没有上限。他设想的终点不只是聊天机器人，而是由平台搭建完整企业——OpenAI、Anthropic 或它们的后继者内部，可能容纳“一个完整的 S&P 500”。
监狱让 Shkreli 对自身失败的解释，从过度绑定成就转向更脚踏实地地看待自己。 他说，自己不肯承认并非世界上最好的对冲基金经理——尽管是很强的企业家——“把我送进了监狱”。他更深层的教训是，天赋异禀的人可能把实现某种宏大成就视为自己的义务，最终走向自我毁灭；而监狱让他看见，地位之下人人共有的普通需求，是想念朋友和所爱之人，以及希望成为一个好父母。
1. 量子计算的科学魔法，不等于一门生意
Shkreli 的核心区分，是更快的计算与一种特定问题上的加速并非一回事。Peter Shore 的思想实验可以把整数分解和椭圆曲线离散对数变成多项式时间问题，却不会加速归并排序这类普通算法。“除非你在分解数字，否则没什么值得兴奋的。”
他认为，今天的量子机器运行频率只有几百或几千赫兹，而不是 GHz 级；在传统任务上，它们更像单核、单线程计算机，也没有量子 RAM。他的结论没有留下余地：“它们永远无法跑赢一台 Nvidia 机器”，更不用说100万台 Nvidia 机器。
他还把问题定义为一场商业押注。他说，Google、IBM 和 Microsoft 每年投入数亿美元，是因为“总有一天可能会有点东西出来”，但他不认为这是一门很好的生意。
主持人的反驳值得保留：这个行业也许正处在经典计算的晶体管阶段。Shkreli 同意进步是可能的，但反问：仅仅因为打孔卡、真空管或铁氧体磁芯存储是必要的先导阶段，投资者当年就应该买入它们吗？
他的实证测试刻意选择了最普通的任务：在他和妻子 Chelsea 测试了全球每一台量子计算机后，“我们在任何一台上都没法算出300乘200”。他做空量子计算公司已经亏了数百万美元，也承认“猜市场时点”并不容易，同时开玩笑说自己因此“赚到了数十亿美元的洞见”。
2. 光子技术可能颠覆 Nvidia 的矩阵乘法范式
Shkreli 把 AI 归结为“matmul”：GPU 有“99.9%”的时间都在做矩阵乘法，而传统乘法的计算量按 N 的三次方增长。两个边长为1,000的方阵，因此大约需要10亿次运算。
他的替代方案来自一篇 Microsoft 关于 AOC，即模拟光学计算的文章：光学中的矩阵乘法是“免费的”，可以瞬间完成。他提到 Neurophos、OptiCore 和 Lumai 都在研发光子核心，同时承认非线性和存储问题仍有大量工程工作要解决。
他的长期判断非常激进：20年内，H100、B100 和 B200 可能都会变成垃圾。但投资组合上的态度并不一致：他持有半导体公司，仍在“押注上行”，即便同时警告，一个效率高出1,000倍的 transformer 显然会伤害 Nvidia。
3. 被忽视的分子资产，是 Shkreli 最高 Alpha 的市场
Shkreli 16岁时加入 Jim Cramer 的对冲基金；他认为 Cramer 扣除佣金后仍能实现23%的年化回报，但后来发现，自己“更擅长做企业家和私募股权投资，而不是交易员”。相比公开市场交易，买入被忽视的医药资产能带来更多 Alpha。
他最具代表性的案例，是以200万美元买下一款 Bristol-Myers 的药；他说，同一家公司如今已经上市、价值60亿美元。他还表示，Pfizer 等药企有时会以约100万美元出售资产权益，即便这些资产可能支持价值10亿美元级别的产品，因为对大药企而言，它们并不重要。
Sparsentan 把这套模式从金融套利推进到临床适应症重定位：它最初用于高血压，后来转向 FSGS——一种此前没有获批药物的疾病。FDA 批准后，Shkreli 说，大约3万名此前从未用上药物的人终于有了药可用；他称这一结果既有用，“也是赚到10亿美元的好方法”。
真正的瓶颈在于信息处理。他的团队需要有人阅读横跨大学和药企的论文与专利；2015年，他在发现 Watson “不太行”之后，创办了一家 NLP 公司，试图处理由4,000万篇医学文章组成的语料库。
4. AI 药物研发炒作为时过早，但对智能的需求可能没有上限
Shkreli 对 Dario 的预测持明确反对态度：AI 在5年内开发出药物是“完全胡说”，未来10年内他也不认为会出现一款 AI 相关药物。他说，电脑和 GPU 可以帮助分析，但药物研发和 FDA 审批依然困难，AI 也无法真正完成药物开发。
当主持人提出，约4000亿美元的风险资本可能正把投资者资金导向购买代币时，Shkreli 否定了由此引申出的估值陷阱论。他认为，对智能和劳动力的需求都是“无限的”，而且远不止知识工作者。
他提到一位离开 OpenAI、创办 Core Automation 的朋友；据他说，此人发明了 OpenAI 的推理模型。随后他设想，AI 平台可以建立 OpenAI Pharma、OpenAI Motors、OpenAI Aviation 等垂直业务——OpenAI、Anthropic 或同类平台内部，可能容纳“一个完整的 S&P 500”。
麦当劳的例子提供了经济学上的检验：只要由此产生的智能拥有正的净现值，花费1000万美元购买“1000万个 GPU 大脑”来挑选下一款三明治就是合理的。只有当继续增加 token 最终不再改善答案时，这一论点才会失效。
5. 分发能力捕获价值，能否活下来取决于自我认知
Shkreli 对市场的类比是互联网：消费者和企业获得了近似“免税”的基础设施，而供应商可能只能赚取短暂的超额回报。他看好 Meta 通过 Messenger 获得分发能力以及产品执行力，称 OpenAI 的硬件策略“非常聪明”，并通过多条路径持有 OpenAI 的投资敞口。他还把 io 和 Jony Ive 的创意产品工作视为一条新路径的证据。
在 AI 安全问题上，他站在有效加速主义一边——“快速行动，打破刻板印象”（move fast and break stereotypes）——并驳斥当前关于生存危机的论调：系统归根结底仍是软件，“你可以关掉电脑”，“它不会杀死我们所有人”。主持人提出的同行评审折中方案，也没有改变他的看法。
被问及他自称曾提供过建议的 Sam Bankman-Fried（SBF）时，Shkreli 提到 SBF 很早就进入 Anthropic，而他在2021-22年的投资组合事后看起来像天才之举。他把非典型的天赋与危险的自我确信联系起来：天赋异禀的人可能觉得自己对某种宏大使命负有责任，而推进过快、拒绝承认失败，最终会把这份负担变成不当行为。
他对自己的检讨，是本集最后的转折：“我很难承认自己不是世界上最好的对冲基金经理，但我是一个很棒的企业家。”监狱剥离了身份和地位，揭示出人们始终会想念朋友和所爱之人，也都希望成为好父母。
完整逐字稿
He has been called the most hated man in America because he raised the cost of life-saving drugs, raising the price of vital medicines by 5,000%. He became a symbol of capitalist greed. Martin Shkreli is serving 7 years in prison.
“We definitely plan to raise the price, that’s for sure.” “You can be held liable for not maximizing profits, and you should do everything in your power to make as much money as possible.” “If I’m your enemy, that’s fine. I’m happy, man. I am fulfilling a certain need in your life.” “I bought the only copy of the legendary Wu-Tang Clan album without any immediate plans to listen to it. This is a good, good record.” “The ego is the source of all progress.”
That’s good. Welcome, Martin. You’re like, “Thank you for coming.” Nice to meet you. You wore Brad’s outfit today.
That’s right.
1. Why quantum computing is a one-trick business and why he’s betting against it
Congratulations, Martin. We have so many topics to discuss with you, but I want to start with one. You essentially criticized quantum computing. You said you thought it was a scam—a “landing slip,” or whatever.
A two-part question: First, perhaps you could lay out your thesis as to why this is a scam. And secondly, your best advice for those CEOs when they do end up in jail. You have some experience.
Best advice, considering you don’t seem like someone who would thrive in prison. But from what I understand, Martin, you were basically a leader in prison. You turned into Lex Luthor and ran the prison yard, I understand. So tell us about the quantum scam and best practices when you get locked up.
Yes, I have to congratulate my wife, Chelsea. Very few people know her, but she’s, I think, an important figure in technology. She was one of the first people at OpenAI. And her father, believe it or not, was one of the first people in the video. I was one of the first in my family to go to prison.
I thought it would be worth changing the gene pool, shifting it a little bit. Chelsea was at MIT and taught quantum computing with Scott Aaronson, the most prominent expert in the field. She had to teach me for many hours to pound it into my thick head and make me understand what quantum computing was.
I think a lot of people equate this with faster computing, which sounds great. But that’s not what it is at all.
No.
I think the main point of quantum computing is one algorithm. In fact, this entire industry was born out of a thought experiment by Peter Shore. Peter was considering the idea that there are types of algorithms that can run faster in a different environment. This was a revolutionary idea in computing.
But it’s not exactly a revolutionary idea for business, because—and this is, of course, my thesis—there is only one algorithm that runs faster on a quantum computer. If you ask someone, “What kind of computer can sort a deck of cards?” and you know the famous CLRS algorithms book, the first 5 pages are devoted to sorting algorithms such as bubble sort, merge sort, and so on.
I asked the quantum enthusiasts, “Well, which one will sort faster?” They said, “Well, of course, a quantum computer.” And I replied, “Of course not.” This is not true at all.
Modern quantum computers operate at clock speeds of hundreds or thousands of hertz, not the gigahertz we’re used to. They’re single-core and single-threaded when it comes to performing a task like merge sort or something like that.
The only thing they’re really good at—and that’s definitely the focus—is going from an exponential-time algorithm, like integer factorization or the discrete logarithm problem on elliptic curves, which I’ll mention later—hello to the ECDLP problem. Quantum computers can do this in polynomial time, which is impressive.
For a nerd, this is the best and coolest magic trick in the world. But for a businessman, unless you’re factoring numbers, there’s nothing to get excited about. Other algorithms that a quantum computer can perform are not much faster. Therefore, they will never outperform an Nvidia machine. They will never surpass a million Nvidia machines.
Quantum computers have no data storage. There is no such thing as quantum RAM. So you’ll have to design a task that has a very, very small amount of input data. You can’t fit all the computers in the world into it.
Whenever someone proposes the idea of a quantum computer to OpenAI, I’ve heard them say, “No, we’re not interested in that.”
My views on quantum technology are closer to yours, but let me take the other side. Let’s just consider another point of view.
The other side would say, “Listen, we’re in this very early, embryonic stage of actually creating the transistor. With a transistor, we’re going to connect a bunch of these things together. We’ll deal with logic gates. We’ll figure out the logic. Then we can build some kind of acceptable operating system on this.
“We’re going to have some clever mechanisms for writing software for these things. None of this actually exists. So you believe that we’re only at an early stage, and that is why we’re disappointed by the computational limitations, other than very narrow use cases?”
Do you believe it’s possible?
Yes, of course. I guess the problem is, would you invest in punch cards and vacuum tubes? Mercury delay lines, ferrite-core memory—these are all things that disappeared and went bankrupt. These were technologies of their time.
CTR—Computing-Tabulating-Recording Company—became IBM Corporation.
So your criticism is about the business model and timing, or is it technological?
I think it’s both. If you talk to Google, IBM, or Microsoft and ask them why they spend hundreds of millions of dollars a year on this, they’ll say, “Well, something might come of this someday.” For me, this is not a very good business.
Well, this is very shocking. You can probably tell this story, but how many qubits are there actually? How many machines do IBM and Google actually have?
Very few. Google has several prototypes. IBM has a small cloud. Chelsea and I have tested every quantum computer in the world. We can’t multiply 300 by 200 on any of them.
And you’re betting on their shares going down?
Yes, I’m playing for a downgrade. Unfortunately, we lost millions of dollars shorting quantum-computing companies, but I told my wife that we gained billions in insights.
This is definitely not easy—guessing the timing in the markets. So where does the natural computational advantage lie?
Some people would say it’s all about encryption and overcoming it, because now we can actually use what it’s really good at to essentially brute-force solutions to things that we can’t solve today.
I think it’s crazy to even think about it, because we have a computational problem that we’re trying to solve. Every time someone uses the word “AI,” I actually think of the word “matmul.” These are matrix-multiplication machines.
He means matrix multiplication.
True. Multiplying two lists of numbers is what the GPU spends 99.9% of its time on. But it does it in time N cubed.
Interestingly, I think complexity analysis will become commonplace over the next 5–10 years as we move toward an AI society. So N cubed means that if you want to multiply a square matrix of 1,000 elements by another square matrix of the same type on an Nvidia machine, it requires 1 billion operations. That’s a lot of operations.
When we were learning these computer concepts through my dear wife, I was flipping through Nature, which we do as a hobby, and I saw an article from Microsoft called AOC, or Analog Optical Computing. In New York, it means something different, but AOC in this case is analog optical computing.
I saw, out of the corner of my eye, that matrix multiplication in light is free. It happens instantly, which is a truly remarkable thing. So we’re exploring maybe starting a company or something.
There are a few startups, like Neurophos, OptiCore, and Lumai, that want to create these photonic cores for matrix multiplication that will leave behind everything Nvidia is doing. Of course, this is not the whole story. There’s a lot of detail involving nonlinearity and storage.
2. The drugs big pharma sells too cheap and his journey from hedge fund trader to pharma
Besides hating quantum technology and your crazy short position, which may or may not be winning, you were a hedge fund manager, right?
The famous episode, just to bring everyone up to speed, was this: You bought the drugs and then inflated the price of them. Is that what happened? Can you just give us a short answer—yes or no—what happened and what led you to it?
Yes. I started at 16 at Jim Cramer’s hedge fund. Cramer was actually an extremely effective and skilled trader. But Cramer’s idea of investing was buying and selling all day, every day, all at once. He somehow squeezed profits out of it.
That was the way it was done in the ’80s and ’90s. He actually earned 23% per annum after deducting commissions.
So you worked for him at 16. How did you get that job?
Well, just being in Manhattan, having a boneless tongue—it’s a great city. I was excited about stocks and investing. Bill Gates is my hero. I found a way to get into that environment.
One day, Cramer pointed at me and said, “Genomics. You’re going to do genomics.” And I said, “Okay.” It was very interesting, and I moved on to a Tiger Cub.
I opened my own fund. I realized that I’m a much better entrepreneur and private-equity specialist than a trader. We found that the small-molecule market gives huge alpha, whereas in the public-company market, we could buy a drug.
I bought a drug from Bristol-Myers Squibb for $2 million, and that same company is public today. It’s worth $6 billion. So this was my first startup.
And it was a truly inefficient market. Pfizer could sell you a drug for $1 million that cost $1 billion. They didn't care. They sold you the rights to the medicine. You could turn it into a product and set a price.
Yes, of course. What, excuse me, do you mean?
You just knock on Pfizer's door, and they're like, “Take whatever you want for $1 million.” We saw this and built 2 very similar companies, where our main limitation was finding people to value these assets. So I started a small company. Yes, just by reading articles, reading patents, going to universities, to small pharmaceutical companies and large ones, and making connections. In our field, this is called business development.
Our biggest bandwidth problem was reading these articles. So, in 2015, I founded a small NLP company in parallel. We had Watson. It was wonderful. We tried to read every single one. It wasn't great. We tried to read every article published in medicine. That was 40 million articles. Getting people to do this is a very difficult task.
How do you take a molecule and turn it from $25 million to $6 million? What do you do in those 6 or 9 months that you have it?
I think what I did with sparsentan, the drug I named and which is now FDA-approved, was a blockbuster. Bristol discovered and developed it for the treatment of hypertension, which is a pretty dead, sleepy market that nobody really needs.
Hypertension—high blood pressure.
And we moved on to a disease called FSGS, a disease for which no drug had ever been approved. So, you try to approach it in a way that makes sense. In that case, the drug received FDA approval. Now, 30,000 people who have focal segmental glomerulosclerosis and never had medication before have it. I can be a good person, too.
I think that's what art is in this field. This is a great way to make $1 billion.
This is what Royalty Pharma does? Royalty Pharma. But is that also what Royalty Pharma does?
They're a slightly different player—brilliant guys, of course. Their main achievement was that they came to NYU when it was receiving huge royalties from Remicade, and they bought them out at a price that was roughly equal to a year's profit.
Yes.
And then these royalties increased 20 times. They buy cash flows at a greater discount than the market values them.
I know you want to insert a word, so say yes.
Martin, take this structure now. You know the main levers of influence in the market today, right? What is expected from computing power? What is expected from this capital expenditure? Do you spend a lot of time analyzing this? Do you see any pitfalls there? What are you afraid of? What would you sell short? What would you bet long on in this complex?
3. AI, drug discovery, and who actually wins
Yes, I already mentioned photonic computing, because I believe that in 20 years, these H100s, B100s, and B200s will end up in the trash. We will find ways to do it better. I think it's crazy what we think about this space. I think it's crazy that we're thinking about nuclear power.
Sam Altman finally stated publicly on another podcast that he thinks photonic computing could be the future. So I just don't understand why we think it's a smart idea to cover the Earth with computers when we can build faster ones.
This is not to say that NVIDIA isn't a good investment. I currently own shares of semiconductor companies. I think Jevons' paradox always wins in the end. Everyone invokes Jevons' paradox in everything. Alex Wang told me that Jevons' paradox would take care of everything.
This will be a million times more computationally efficient. In Silicon Valley, they talk about better transformers that are 1,000 times more efficient. If a transformer that is 1,000 times more efficient appears, it's bad for NVIDIA. There is no doubt about that.
So, you know, we're not going to have a million times more demand. Well, we can have it in another way, but there could be a million times more people gone. Whether we will have a million times greater need for computing power remains to be seen. So far, I think my guess, like everyone else's, is to play for the upside. But in 2000, I would probably also have bet on Yahoo and Cisco. The situation here is not simple.
So how do you—are you still running a hedge fund?
No, I don't run one.
So you manage your own money, and when you think about portfolio allocation, do you approach it the same way you would when you were managing a fund? Do you look for specific high-alpha events, startups, or disruptive technologies and actually invest in them without trying to compete with the market?
Yes, I'm sure I'm playing for appreciation. I mean, that Dario trade was a great reason to play for appreciation. So I invested in the midst of this fear, but really—
Yesterday's sell-off was just a deception maneuver.
Yes, I think that was ridiculous.
What do you think of his essay?
I want to give him credit for building a billion-dollar company, but I don't like the guy. Building a trillion-dollar company is not easy. I think he's spreading fear.
Imagine that you and I started a business, and I said, “I'll show you Pandora's Box. Here it is.” And you said, “Wow, that's amazing.” Then you said, “Thank you, Martin.” I'll go around the corner and open Pandora's Box, too. What are you doing? I would probably sue you for disclosing a trade secret or something like that. That's exactly what happened.
So you're saying that because he was at OpenAI, and then he and 5 or 6 of his co-founders left, it's like—would you be friends with a guy like that? Would you like that?
Well, I think that happens quite often in Silicon Valley. It's quite difficult to determine what is a patent and what is acquired knowledge. I'm not justifying it or saying I know exactly what happened, but—
The Waymo guy went to jail for it.
Well, I think it's because it turned out that he actually stole the physical documents that were—
Okay, so you have some questions about the history of the project.
This is a ridiculous exaggeration. He says AI will help develop drugs in 5 years. Complete nonsense.
What about effective altruism?
Yes, even worse.
Why? What repels you so much about this group of people? Let's figure it out.
Yes, I helped popularize, and I think I would even say started, the EAC movement, which was a kind of countermeasure.
Sorry, there's EA—
Yes. There's e/acc, which was intentionally named—
EA—it's effective altruism. EAC is—
Effective accelerationism.
Effective accelerationism. What does this mean?
That's a great question. It means something like “move fast and break stereotypes.” It's no secret that Zuckerberg came out the other day and said that we need to move fast in AI. I think fearmongering is a bit pointless, and there will still be time to panic. Just not now.
Basically, AI provides the following capabilities. For now, you can turn off the computer. This is not such a big problem. I think nerds get too excited. If you were ever a nerd or a jock, I was, oddly enough, a little bit of both.
Athletes know what to do when they find themselves in the spotlight. Nerds—no, they withdraw into themselves. Some of them, instead of saying that everything is fine, want to attract attention to themselves. They want to say, “The thing I'm working on is really impressive.”
The reality is that it's just software. This is not the end of the world. It's the most important kind of software ever created, as you said before, but it won't kill us all.
Darryl also said about 2 years ago that we would be completely out of work within 1–5 years. Well, he has 3 years left. He also said that we will cure all diseases. I don't think there will be a single drug related to artificial intelligence in the next 10 years.
AI doesn't help with drug development. If you have ever developed a drug, raise your hand if you developed a drug and received FDA approval. Okay, I'm pretty sure that's it. You know, it's quite difficult. This is quite difficult to do. We've had computers for a long time. I built a Cray cluster for drug analysis. Now I can use GPUs. That's great, but AI can't really help us develop drugs.
But what exactly is the revenue cycle that determines the valuation? Let's look at this from a hedge fund perspective. You're sitting there thinking, “Well, this is a hedge fund. You see these rates, right?” I think Jensen mentioned this yesterday: $400 billion of venture capital has gone into AI. So do you see this as VCs funneling investor money into buying tokens?
Yes, I think so.
And that's your argument? I think you're going to have a valuation trap. Do you think?
I don't think so. I think what you mentioned earlier—productivity—I mean, there's an infinite demand for intelligence and an infinite demand for labor.
When you said labor, I thought, “I think Brad is making a mistake here.” Because if you take the total market size and limit it to just knowledge workers, you're actually underestimating it.
Sure.
And, you know, I'm trying to be conservative.
Yeah, there's no limit for intelligence.
One example that I've often given is my buddy left OpenAI and started a company called Core Automation, and we're really inspired by that. I mentioned it the other day. He invented the reasoning model at OpenAI, and he has a pretty low-profile startup.
I think you're going to see AI start businesses. I think part of his plan is to create, inside OpenAI, OpenAI Pharma, OpenAI Motors, OpenAI Aviation, OpenAI, and so on, and he'll have his own versions of that. You'll see an entire S&P 500 index inside OpenAI, Anthropic, or wherever.
So I think the possibilities are endless. The last thing is that McDonald's can afford to spend $10 million and say, “OpenAI, give me 10 million GPUs. I want to know what the next best sandwich is that we should launch.” That's the net present value with a plus sign, if intelligence is worth it. If it's not an asymptote, then at some point more tokens don't give a better answer. But when the day comes when you say, “Instead of hiring McKinsey or a focus group, I need 10 million GPU brains to think about what McDonald's should do now,”
What's a good analogy for the AI market? You see a handful of providers of a critical resource and incredible valuations, which, of course, attracts a lot of additional capital and generates a lot of competition. There are different solutions. Jensen made the analogy yesterday that it's kind of like water. You still have $8 water, but you also have free water. It's a world where all the flowers bloom. What's your mental model? How do you compare this to previous markets that we could understand?
It has to be like the internet. At least part of it has to be like the internet, where the only winners will be consumers and businesses. They get this kind of tax-free internet. Cisco and other players will get a little bit of a kick out of it, but the main highway that we all benefit from—and we already do—is free intelligence in our pockets.
So I would interpret your words as saying that we have to guess now: A, who will make more than the norm, and B, how long that excess profit lasts.
Sure. And I think it's Meta.
Timing?
I think it's Meta.
You mean Meta?
Yeah. They have distribution through Messenger.
Why exactly?
They have distribution through Messenger. They've shown that they still have a knack for products. I think we're still in the early days of AI. I actually think that chatbots—we'll look back and laugh and say they were the early days.
Would you own—would you go out today and buy OpenAI and Anthropic?
Yeah. We have OpenAI, obviously, through a couple of different avenues. I think Sam's bet on hardware is just brilliant because he's always going to have the best model, I think. But being able to provide that model is probably going to be more valuable, and that's what we saw with io. There's really creative product work that Jony Ive did. It shows a new path for them.
What do you think about Elon's suggestion of peer review as a sort of third way, a pragmatic compromise solution for balancing safety and speed?
They say that the Wright brothers—if Warren Buffett could go back in time, he would go back and shoot the Wright brothers to stop the aviation industry from destroying capitalism. I think Nick Bostrom wrote a book called Superintelligence, which scared Elon to death. Maybe I'm not advocating violence, but it would be better if some of the effective altruists had never developed this narcissistic self-confidence that drives some of this fearmongering.
I don't think OpenAI believes that. I think they're in it because they don't want to be too different from Anthropic. Fortunately, some of the other guys have the backbone. Meta has the backbone to call a spade a spade and say it's nonsense. The president has the backbone to do the same thing. Jensen has the backbone to do the same thing. It's just computer science. It's not that big of a deal.
Some of these effective altruist guys are pretty brilliant. Dario Amodei's not a fool. Dustin Moskovitz, Sam Bankman-Fried—sometimes you can be very smart about one thing and very stupid about another. Let me ask you this question: Do you think that's true?
I actually do, yes. I mean, that's a good example.
4. Brilliance and rule-breaking, the gifted-kid trap & what prison taught him
Well, then this is actually my argument. I feel like some of the smartest people are unconventional thinkers. You're clearly a brilliant person. People who are brilliant are usually brilliant because they're unconventional. They're heterodox to some extent, but at the same time, if you have that kind of genius, maybe you also have a hard time being conventional in what, let's say, is law-abiding.
By the way, I want to ask your opinion on Sam Bankman-Fried. You met this person, right?
I gave him advice.
Yeah, you gave him advice on how to survive. Can you share your impression of this person?
He was early on in Anthropic, at the very beginning. You look at the investment portfolio that he built. All in all, it was crazy in 2021 and 2022. It looked like genius, yeah.
Can you tell us a little bit about what got this guy off track? Was it because he was such an unconventional thinker? That's what got him all this success, but it also created problems. Or was it the speed?
Yeah, I think it's both. One of the problems is that if you grow up as an intellectually gifted person, you have this big burden. I think that happened with Elizabeth Holmes. Raj Rajaratnam is another example—a hedge-fund manager who was a pretty good investor.
In short, you have this self-awareness that you have to be accountable to something. Sometimes the desire to be accountable to this big thing can be very damaging, and failing and admitting that you're not that great at it is not that bad. It was hard for me to admit that I wasn't the best hedge-fund manager in the world, but I was a great entrepreneur. That simple refusal to admit that got me to prison.
How has that changed your perspective on our behavior? What exactly changed about you after prison? What changed? You've changed a lot—you've become sharper and a little more methodical in your approach to life. Tell us what you've learned during that time.
Well, thank you very much. One of the interesting things about prison is that no matter where someone comes from, what they've done, or whatever, there's a lot of commonality that unites us as people. Everyone is upset that their friend isn't keeping in touch, or they miss that person, or whatever it is. It's really refreshing to remember that no matter how high we reach, deep down in our souls, like you said before, we all want to be the best parents we can be, or something like that.
Putting aside your case, whether it's valid or not, if you've been prosecuted for whatever reason, what have you learned while you've been in prison? How has that changed you as a person? You've changed a lot—you've become sharper and a little more methodical in your approach to life. Tell us what you've learned during that time.
Well, thank you very much. I think one of the interesting things about prison is that no matter where someone comes from, what they’ve done or whatever, there’s a lot of commonality that unites us as people. Everyone is upset that their friend isn’t keeping in touch, or they miss that person that, you know, they miss, or whatever it is. And it’s really refreshing, you know, to remember that we, you know, no matter how high we reach, deep down soul, like you said before, we all want to be the best parents we can be, or something like that.
Martin, listen, I think you're just a very independent thinker, and you're also willing to say things and communicate in ways that others don't. That's why we wanted to have you here. I think this was obviously a great demonstration of that. We really appreciate you taking the time to be with us.
Thank you very much.
That's great. Thanks. Nice to meet you.
Nice to meet you.
Hey, man, come by anytime.
Oh, thanks, bro. Thanks.