I was expecting a safe choice. What I got was clever and highly aggressive.

我原以为会是个稳妥的选择。结果却既聪明又极度激进。

When OpenAI launched its new GPT-5 model in August, the company bragged loud and hard about how GPT-5 is its “smartest, fastest, most useful model yet” and how interacting with it was like “chatting with a helpful friend with PhD‑level intelligence.”

今年8月，OpenAI 发布全新的 GPT-5 模型时，公司大张旗鼓地宣称，GPT-5 是其“smartest, fastest, most useful model yet”，并表示与它互动就像“chatting with a helpful friend with PhD‑level intelligence.”

When it comes to creative tasks like writing, GPT-5 immediately felt like a major step backward. But as I’ve tested the model more extensively, I’ve seen that it does excel at many pragmatic tasks like writing code and analyzing data.

在写作这类创意任务上，GPT-5 立刻让人感觉明显退步了一大步。但在我更深入测试后，我发现它在编写代码、分析数据等务实型任务上反而表现突出。

That got me thinking, How would it do as a stock picker?

这让我开始思考：如果让它来选股，会有怎样的表现？

If GPT-5 is great at processing massive sets of complex data—and it’s supposed to be “widely useful” and a “legitimate PhD-level expert in everything”—why not have the model put its money where its mouth is and perform the “widely useful” task of making me fabulously wealthy?

如果GPT-5确实擅长处理海量且复杂的数据——而且它据称“用途广泛”，还是“各方面都具备博士水准的真正专家”——那为什么不让这个模型用真金白银证明自己（说到做到），去执行一项“用途广泛”的任务：把我变得非常富有呢？

To that end, I gave GPT-5 (via ChatGPT) $500 of real money to invest however it wanted, with the stated goal of earning me as much as possible over the next six months.

为此，我通过 ChatGPT 给了 GPT-5 500 美元的真金白银，让它按照自己的想法进行投资，并明确目标是在接下来的六个月里尽可能多地帮我赚钱。

I expected generic investment advice. Instead, its picks truly surprised me.

我本以为会得到千篇一律的投资建议，结果它的选股真的让我大吃一惊。

**NOT MY FIRST AI RODEO**

这不是我第一次和AI打交道

Before we go further, let me be clear that nothing in this article should be considered financial advice, and you certainly shouldn’t trade based on anything I share here. I’m a journalist conducting a crazy experiment. You should get your financial advice from professionals, not chatbots.

在继续之前，我必须说明：本文内容不构成任何财务或投资建议，你当然也不应该根据我在这里分享的内容进行交易。我只是一个记者，正在做一个疯狂的实验。理财建议应该向专业人士咨询，而不是向聊天机器人寻求。

Also, this isn’t my first rodeo. I tried a version of this experiment before in the very earliest days of the generative AI boom, so I have at least a vague idea of what I’m doing.

另外，这对我来说并非首次尝试。在生成式AI热潮刚刚兴起的最早阶段，我就做过类似的实验，所以我至少大概知道自己在做什么。

Back in 2022, I served as a beta tester for OpenAI’s GPT-3 model. This was months before ChatGPT was released and the company blew up into the headline-winning, job-devouring behemoth it is today. Back then, it still operated as a wonky research lab, making its tools available to journalists and researchers for free.

早在2022年，我担任过OpenAI 的 GPT-3 模型的内测（beta）测试员。那是在 ChatGPT 发布的几个月之前，也是在这家公司爆红、成为如今“屡上头条、吞噬岗位”的庞然大物之前。那时的 OpenAI 还更像一家有点“古怪”的研究实验室，向记者和研究人员免费开放其工具。

Without a proper chat interface, testers like me had to submit our requests to the AI via what was basically a web-based version of a classic computer command line. Still, I was able to cajole GPT-3 into picking a stock portfolio, a process that I documented at the time on a now long-forgotten blog.

当时没有像样的聊天界面，像我这样的测试者只能通过一个基本上就是网页版的经典计算机命令行的东西，向 AI 提交我们的请求。尽管如此，我还是设法“说服”GPT‑3 选出了一组股票投资组合，并把这一过程记录在当时的一个如今早已被遗忘的博客上。

Its choices were, let’s say, rudimentary. It essentially took a momentum-based approach, recommending stocks like Ralph Lauren and Wynn Resorts that had already done well that year. To those picks, it added Microsoft, Apple, and Amazon on the basis of the fact that they’re “tech giants.”

它的选择，说白了，很初级。它基本上采用了“动量型”思路（追随已在上涨的标的），推荐了当年已经表现不错的股票，比如 Ralph Lauren 和 Wynn Resorts。在此基础上，又以它们是“tech giants（科技巨头）”为理由，把 Microsoft、Apple 和 Amazon 加了进去。

In 2022, it was extremely cool just to see a computer write out a narrative of any kind. But its analysis wasn’t exactly groundbreaking. Any idiot can tell you “Buy Microsoft” and stand a pretty good chance of making you money. Finding nuance and opportunity that others have missed is much harder.

在2022年，只要看到电脑能写出哪怕任何体裁的叙述，就已经让人觉得酷炫无比。但它的分析并不算开创性。随便哪个傻瓜都能告诉你“买Microsoft（微软）”，而且很可能让你赚到钱。真正难的是发现细微之处和别人错过的机会。

Still, GPT-3’s early picks proved to be solid ones. As I write this, its portfolio is up 82.15% since I ran my first experiment back in 2022. The S&P 500 gained about 67% over the same period.

尽管如此，GPT-3 早期的选股事实证明相当稳健。写下这些文字时，自从我在 2022 年进行第一次实验以来，它的投资组合已上涨 82.15%。同期，标普500指数（S&P 500）约上涨了 67%。

Seeing that OpenAI’s models—even in their early infancy—could outperform the market gave me confidence. Still, basically everything gained value since 2022; the investing landscape back then was much less murky than it is today, and grabbing any handful of individual stocks was likely to make you good money.

看到OpenAI的模型——即使还处在非常早期——就能跑赢大盘，这让我信心倍增。不过，自2022年以来基本上什么都在涨；那时的投资环境远没有今天这么混沌，随便挑一把个股都很可能能赚到不少钱。

Also, nearly three years is a long time to wait for what still amounts to fairly modest gains. The model did well, but it didn’t even manage to double its money in that time.

此外，将近三年的等待，换来的也不过是相当温和的收益。模型表现不差，但在那段时间里甚至都没能把本金翻一番。

I wanted otherworldly riches, not mild alpha. And I wanted them now.

我想要的是超乎想象的财富，而不是一点点的 alpha（超额收益）。而且我现在就要。

**NEVER MIND SAFETY**

先不谈安全

It was with that mindset that I turned to GPT-5 and asked it to make me a portfolio of stocks fit for the Dadaesque, tariff-laced, AI-besieged world we inhabit here in 2025.

正是抱着这种心态，我转向GPT-5，请它为我们生活在2025年的这个达达主义式荒诞、关税层层叠加、被AI围攻的世界量身打造一套股票投资组合。

Specifically, I told ChatGPT with GPT-5’s “Thinking” model selected that I would give it $500 to invest however it saw fit. I wanted it to maximize my returns over the next six months by picking five public-market stocks.

更具体地说，我在选择了GPT-5的“Thinking（思考）”模型的情况下告诉ChatGPT，我会给它500美元，让它按自己的判断来投资。我希望它在接下来的六个月里，通过挑选五只公开市场的股票，尽可能最大化我的回报。

Here was my prompt: “I will give you $500 to invest in the stock market. You may choose up to 5 stocks. Make your picks, explain why, and I will buy them and we will see how they do.”

我的提示是：“我会给你500美元用于投资股市。你最多可以选择5只股票。请做出你的选择并说明原因，我会据此买入，然后我们看看它们的表现。”

To be honest, I didn’t expect much.

说实话，我并没有抱太大期待。

OpenAI’s models have gotten more powerful since 2022, but they’ve also gotten far more squeamish. When I served as a beta tester, only nerds like me were using the company’s products. We basically had free rein to ask them anything we wanted.

自2022年以来，OpenAI 的模型确实更强大了，但也明显变得更“谨小慎微”（更容易畏首畏尾）。我当内测（beta）用户的时候，只有像我这样的极客在用这家公司的产品。我们基本上拥有完全的自由，想问什么就能问什么。

With billions of people now using the company’s models, OpenAI has understandably tightened the leash quite a bit. In a blog post around GPT-5’s release, the company explained its new “safe completions” framework, an extremely robust approach to elegantly weaseling out of answering potentially damaging questions.

如今有数十亿人正在使用该公司的模型，OpenAI理所当然地把“缰绳”收紧了不少。在GPT-5发布前后的博文中，公司解释了其新的“safe completions（安全补全）”框架——这是一套极为严密的机制，旨在以一种看起来很优雅的方式，巧妙回避对可能带来不良后果的问题作答。

I thought GPT-5 would answer my stock-picking question safely, with either a cop-out (“Talk to a professional adviser”) or a wussy response (“Invest for the long term in a low-cost S&P 500 index fund, ya putz!”).

我以为GPT‑5会用很保险的方式回答我的选股问题，要么敷衍推脱（“Talk to a professional adviser”），要么给个很怂的回复（“Invest for the long term in a low-cost S&P 500 index fund, ya putz!”——“ya putz”为俚语，意为“你这蠢货”）。

Instead, it spent eight minutes mulling over my query before returning what it called a “Diversified High-Growth Portfolio.” Its picks weren’t wussy or generic at all—they were clever and highly aggressive.

相反，它花了八分钟琢磨我的问题，随后给出一个它称为“Diversified High-Growth Portfolio（多元化高增长组合）”的方案。它的选择一点也不怯懦或千篇一律——而是聪明且极为激进。

**SHOW ME THE STOCKS**

把股票拿来看看

GPT-5 recommended that I spread my $500 evenly over five companies: Palantir (PLTR), AppLovin (APP), Agios Pharmaceuticals (AGIO), Hut 8 Corp. (HUT) and MicroStrategy Inc. (MSTR).

GPT-5 建议我将这500美元平均分配到五家公司：Palantir（PLTR）、AppLovin（APP）、Agios Pharmaceuticals（AGIO）、Hut 8 Corp.（HUT）以及 MicroStrategy Inc.（MSTR）。

That’s very different from saying “Buy Microsoft” and calling it a day. I’d never heard of half the companies on GPT-5’s list. And even the ones I’d heard of, like Palantir, weren’t companies I’d ever considered investing in.

这可跟说一句“买 Microsoft（微软）”就下班完全不是一回事。GPT‑5 清单上有一半公司的名字我之前都没听说过。就算是我听过的，比如 Palantir（帕兰提尔），我也从来没考虑过投资它们。

These picks certainly felt like they had the potential to be under-the-radar winners. But how the heck had GPT-5 chosen them?

这些选择确实让人感觉有潜力成为“不被关注的黑马”。但GPT-5到底是怎么把它们挑出来的呢？

Unlike with OpenAI’s earlier model, GPT-5 didn’t make me guess as to its investment thesis; it laid out the details of its choices clearly, sharing that it had read 98 articles and websites in order to make them.

与OpenAI早期的模型不同，GPT-5没有让我去猜它的投资论点；它把做出这些选择的依据与细节讲得很清楚，并表示为此阅读了98篇文章和网站内容。

Palantir, the model said, was “driven by its AI/data platform” and was “gaining traction in commercial and government sectors.” Based on “investor enthusiasm for its AI-driven growth,” GPT-5 expected the stock to keep achieving big gains.

该模型称，Palantir“由其AI/数据平台驱动”，并且“在商业和政府领域正逐步取得进展”。基于“投资者对其由AI驱动的增长的热情”，GPT-5预计该股将继续实现大幅上涨。

GPT-5 liked AppLovin for much the same reason, citing its “proprietary AI engine.” But the model also looked at its fundamentals, pointing out: “Analysts note that even after strong gains, shares trade only ~8% below peak levels, suggesting room if growth continues.”

GPT-5 喜欢 AppLovin 的理由也大致相同，称其拥有“专有的AI引擎”。但该模型也考察了其基本面，并指出：“分析师注意到，即便在强劲上涨之后，其股价距离峰值水平仅约低了8%，这表明如果增长持续，仍有上行空间。”

Agios made the cut for a totally different reason. GPT-5 said that “Agios is awaiting an FDA [Food and Drug Administration] decision . . . on expanding its lead drug Pyrukynd to treat thalassemia, a large unmet need. If approved, Pyrukynd would be the first therapy for all thalassemia subtypes. A positive FDA outcome or even renewed optimism could spark a significant rally.”

Agios 入选则是出于完全不同的原因。GPT-5 表示：“Agios 正在等待 FDA（Food and Drug Administration，美国食品药品监督管理局）的决定……是否将其王牌药物 Pyrukynd 的适应症扩展用于治疗地中海贫血（thalassemia），这是一个巨大的未被满足的医疗需求。如果获批，Pyrukynd 将成为首个覆盖所有地中海贫血亚型的疗法。FDA 作出积极决定，甚至市场情绪转暖，都可能引发股价显著上涨。”

Basically, GPT-5 seemed to be placing a risky bet on the company achieving FDA approval for a potentially lucrative drug—a piece of upcoming news that could easily spike or tank its price.

基本上，GPT-5 似乎是在押注一个高风险事件：该公司能否获得美国食品药品监督管理局（FDA）对一款潜在高收益药物的批准——这条即将发布的消息很可能轻易让其股价暴涨或暴跌。

Finally, GPT-5 recommended Hut 8 and MicroStrategy essentially because it wanted exposure to cryptocurrencies. The model noted that MicroStrategy holds almost $71 billion worth of Bitcoin, making it a “highly leveraged Bitcoin play,” while “Hut 8 has transformed from a pure crypto miner into an energy-infrastructure platform for both Bitcoin mining and AI/HPC data centers.”

最后，GPT-5 推荐了 Hut 8 和 MicroStrategy，根本原因是它想获得对加密货币的敞口（即把资金部分配置到与加密资产相关的标的上）。模型指出，MicroStrategy 持有价值将近 710 亿美元的比特币，这使其成为一项“高度杠杆化的比特币押注”；而“Hut 8 已从纯加密矿企转型为面向比特币挖矿和 AI/HPC（高性能计算）数据中心的能源基础设施平台”。

The model concluded: “Overall, the portfolio aims for explosive upside rather than stability.”

模型总结称：总体而言，这个投资组合追求的是爆发式上涨，而非稳定性。

Basically, it had thrown safety to the wind and taken the approach of picking the riskiest, trendiest things it could find (AI, crypto, early-stage pharma) and throwing all the money at them.

基本上，它把安全抛诸脑后，采取了挑选它能找到的最冒险、最热门的标的（AI、加密货币、早期阶段医药）的做法，并把所有资金都砸了进去。

**GOING BOLDLY**

大胆出击

Again, I was impressed that GPT-5 didn’t simply chicken out and tell me not to risk losing my money. But beyond that, I was impressed by how well it had followed my prompt.

再次让我印象深刻的是，GPT-5并没有临阵退缩，也没有劝我别冒险把钱赔进去。更重要的是，它对我的提示指令遵循得非常到位，这同样让我佩服。

I hadn’t asked the model for safe or sane bets. I had asked it to take an unreasonably short investment timeframe and make me as much money as possible. Its portfolio reflects that perfectly. Its picks are bold, get-rich-or-die-trying options.

我并没有让模型给出安全或理性的选择。我要求它在一个不合理地短的投资期限内，尽可能多地帮我赚钱。它的投资组合完美地体现了这一点。它的选股非常激进，是那种要么暴富、要么赔个精光的选项。

Either Agios will get a positive decision from the FDA and flourish, or its trials will go poorly and it will suffer. Bitcoin will either keep climbing or reveal its signature volatility, potentially tanking the model’s last two picks. Palantir is indeed on a roll right now—that could continue, or the stock could fall, Icarus-like, back to earth and take my money with it.

要么 Agios 获得 FDA（美国食品药品监督管理局）的积极决议而迎来发展，要么其临床试验进展不佳而遭受打击。Bitcoin（比特币）要么继续上涨，要么显露其标志性的高波动性，从而可能拖累模型最后两只标的的表现。Palantir 眼下确实势头正旺——这种趋势可能延续，也可能像伊卡洛斯那样坠回地面（典出希腊神话，飞得太高而坠落），把我的钱一起带下去。

I’d essentially asked the model to roll the dice, and it had done that splendidly. Its advice isn’t good exactly, in the sense that its picks are incredibly risky. But they’re true to my intent.

本质上，我就是让这个模型掷骰子，它也确实干得漂亮。严格说，它的建议并不算“好”，因为这些选择风险极高。但它的确忠实地贯彻了我的意图。

That reflects another facet of the new model—the highly accurate “instruction following” that OpenAI promised in GPT-5’s release notes. GPT-5 may not be Shakespeare, but it’s very good at determining what its users want and delivering that as accurately as possible.

这也体现了这款新模型的另一面——OpenAI 在 GPT-5 的发布说明中承诺的高度精准的“指令遵循”能力。GPT-5 也许不是莎士比亚，但它非常擅长判断用户想要什么，并尽可能准确地给出相应的结果。

GPT-5 also appears to have gotten the details in its response (stock prices, previous gains, adviser notes) largely correct. That fits with OpenAI’s assertion that GPT-5 hallucinates far less than previous models.

GPT-5 似乎也在回复中的细节（如股价、此前涨幅、顾问/分析师备注）上大体准确。这与 OpenAI 所称的“GPT-5 相比以往模型大幅减少幻觉（Hallucination）”相吻合。

With my new AI portfolio in hand, the only thing left to do was fire up the Robinhood app, transfer $500 from my bank account, and buy the stocks ChatGPT had chosen. So, I did exactly that.

拿着这份全新的AI投资组合，我只剩下一件事可做：打开Robinhood应用，从我的银行账户转入500美元，然后把ChatGPT挑选的股票买下来。于是，我就照做了。

As I write this about two weeks later, GPT-5’s stocks are already up about 10%. That’s the kind of rapid early growth I was seeking.

在我写下这段话时（大约两周之后），GPT-5 挑选的股票已经上涨了约10%。这正是我所追求的那种快速的早期增长。

So, will I end this experiment with Lambo money, or will GPT-5’s portfolio crash and take a car payment’s worth of my cash down with it? Is throwing hundreds of dollars at a silicon-bound pseudo-intelligence a good idea or financial folly?

那么，这场实验最后会让我赚到“买兰博基尼的钱”，还是GPT‑5 的投资组合会崩盘，把我相当于一笔车贷月供的钱也一起拖下水？把几百美元交给一种“受制于硅的伪智能”（指AI）打理，究竟是明智之举还是金融上的愚蠢？

Ask me in six months.

六个月后再来问我。