最近我在思考一个问题:为什么 AI 几乎统治了所有智力活动,却还没有在投资上打败人类?
本着「先问是不是,再问为什么」的精神,我们看一下全美投资大赛 (USIC) 2025 的成绩,并且把前6名的风格列出来。其中只有 J Law 公开承认用 AI 做投研,即便如此他依然靠自己决策。
| 排名 |
交易者 |
风格 |
是否使用AI |
| 1 |
J Law +252.3% |
Minervini风格 VCP + Multiple Edge Trading System。中期摆动/仓位交易,重技术(70%)+基本面催化剂与主题(30%)。 |
用多个 AI 交叉验证催化剂与新闻,帮助持仓 conviction。核心决策仍是自己做。 |
| 2 |
Christian Flanders +167.5% |
前职业扑克玩家转交易。成长股摆动,事件驱动(episodic pivot)、阶段分析、严格风控与 progressive exposure。 |
无公开证据使用AI做核心决策。 |
| 3 |
Clement Ang +140.4% |
香港交易者,受 Minervini、Gil Morales、Oliver Kell、Qullamaggie 等影响。成长股摆动转向仓位交易,强调主题/相对强度 + 严格回撤控制。 |
无公开证据。 |
| 4 |
George Tkaczuk +86.3% |
RCM Wealth Advisors 组合经理。技术分析 + 宏观/微观趋势,适合 choppy market 的快速止损与轮动。2020年该组别冠军。 |
无公开证据。 |
| 5 |
Tong Wang +73.5% |
公开信息极少,ViZen Management 相关。 |
未知。 |
| 6 |
Steve McIntosh +70% |
动量/成长股摆动(受 Qullamaggie、Minervini、Roppel 影响)。基金明确表示不用算法交易,依赖人工“感觉市场温度”。 |
明确不使用算法/AI执行。 |
实际上 USIC 的确有参赛者大量使用 AI 决策,比如我搜到的这两位:
Dr. Terry Potter(AI Trading Academy,香港)
2026年 Q1 Enhanced Growth 组别($20k+)领先,+1,678.8%,创该赛事任何组别三个月新纪录。官方新闻稿标题即为“AI Aided Trader Sets Record”。他主要交易商品,依赖价格行为。
Dr. Tom Doub(Arcadia Analytics,田纳西)
2026年Stock组别表现极强(例如中期曾领先,+880.7%等)。他公开表示交易完全系统化,策略与参数由自己设计,所有入场由系统自动执行。同时大量使用AI做研究、写代码和agent辅助。
2026 年的结果未定,我们尚不知道任何组别是否会有 AI 选手夺冠。但不论如何,在 2026 年末的这个时间点,「AI 还没有在投资上打败人类」这个论断是成立的。当然,这里的「没有打败」指的是类似 AlphaGo 这种「独立端到端操作」,而非投资中的每个环节都弱于人类——比如搜集整理信息、发掘量化信号等方面,AI 显然早已把人类甩在身后。
投资只是一个引子,我真正想探讨的是,AI 究竟擅长和不擅长哪些工作?
为此,我把人类活动按「输入/输出」「简单/复杂」分成四个象限:

属于每个象限的活动举例:
- ① 简单输入 ➜ 简单输出:打螺丝、翻译
- ② 简单输入 ➜ 复杂输出:围棋、编程、数学
- ③ 复杂输入 ➜ 简单输出:投资
- ④ 复杂输入 ➜ 复杂输出:规划人生、公司治理
① 在 LLM 之前已经被解决了,工业机器人早已接管很大一部分制造业,传统 CV 也足够应付大部分工业场景。
② 在过去十几年间从神经网络 + 强化学习开始,到现在被 LLM + Agent 彻底解决了。为什么说这几项活动输入简单?围棋规则学前小孩都能懂,编程也只需要输入键盘上的字母和符号。数学看起来复杂,其实早已被符号化且有严格的公理体系,这些输入在 AI 看来都是简单的。
③ 的代表就是投资了。很多人说「投资不就是高买低卖么」,的确,投资的输出就是这么简单,简单到人人都以为自己能掌握。然而,这种简单的背后却依赖复杂的输入:量、价、基本面、新闻、宏观,无法穷尽。
④ 即便排除掉那些物理上 AI 还完成不了的事,目前也还无法将这类活动完全委托给 AI。
在「简单输入 ➜ 复杂输出」这组里,不论是曾被视作智力巅峰的围棋,还是自然之母的数学,AI 完全吊打人类。为什么输入简单是关键?因为训练任何 AI 模型,首先需要把真实状态离散化编码。比如棋盘状态本来就有限,因此 AlphaGo 提取特征很容易;语言模型则需要先用 embedding 将词和概念映射成向量。而一旦完成了离散化编码,只要训练数据和算力充足,必然可以训练出足够好的模型。讽刺的是,人类过往最推崇的那些智力活动,比如计算、逻辑、搜索,既是人脑的短板,又恰好是 AI 的特长。人类在演化中并不依赖这些能力,自然大脑也没有朝这些方向进化,物以稀为贵的我们便最为推崇那些恰好拥有这些特质的大脑。现在 AI 解决了②,也促使我们重新审视「智力」这件事,比如八种智能中过往不被重视的那些,从这个角度看不失为一件好事。
说回第三类活动。当输入变得不再有限,人类便逐渐展现出优势。首先,人不依赖真实状态离散化;其次,人可以通过经验和直觉排除掉大量无关信息;最后,可能也是最关键的,人可以在样本极小的情况下,注意到某些”异常“,从而规避掉一些尾部风险。这里必须要提一嘴 Frame problem(说起来我还是当年玩 Steins;Gate 0 第一次知道这个概念的,感谢 Maho🙏),它描述了智能和认知研究领域的一个核心问题:如何让智能系统如何只关注相关信息、忽略无关变化,以及预见自身行为的后果?尽管这个问题的背景是 AI 尚基于规则的蛮荒时代,但对于把预测下个 token 当作 AI 基石的现况,我认为仍然是有意义的。尤其是,如果想让物理 AI 真正落地,那么处理复杂输入的能力势必要提升。我对未来保持乐观。或许我们并不需要完美的理论,甚至不需要真正解决复杂解决输入的问题,只需要限制输入的类型,让 AI 能在特定场景下较好地工作(比如像前文提到的 Dr. Terry Potter 那样只使用价格行为),也就足够了。

2026: AI Editor → Agent Command Center
As of 2026, there are four major forms of agentic coding products:
| Phase |
Product Form |
Representative Products |
| Transition period |
Editor & IDE + chat side-panel |
VS Code, Cursor 1.0, Antigravity 1.0 |
| Transition period |
Coding Agent CLI |
Claude Code, Codex CLI, Gemini CLI |
| 🤖 AI-native |
Agent coworker chat apps |
Slack, ClickUp, Slock |
| 🤖 AI-native |
Agent command center apps |
Codex App, Cursor 3, Antigravity 2.0 |
We can split them into two phases:
Phase 1: Transition from human → agentic coding (2023 ~ 2025):
Products: Editor & IDE + chat side-panel, Coding Agent CLI
- Editors predate LLMs, adding a side-panel enables AI coding without breaking the classic interface.
- Coding agent CLIs were first built as an experiment (according to Claude Code's creator). Nobody thought they would work this well, so the creators chose a CLI over a GUI simply because it was easier to build.
Phase 2: 🤖 Agent-native workflows (2025 ~ now):
Products: agent coworker chat apps, agent command center apps
Compared to Phase 1—where humans still wrote code by hand or constantly steered the agents—the primary user journey has shifted to orchestrating agents to handle work from start to finish. The agent is no longer an assistant living in a side panel; it represents the core functionality of the product.
In 2026, there is a notable trend: companies are redesigning their existing products to be agent-focused. For instance, Google launched Antigravity 2.0 at I/O, replacing the VS Code-based editor with an agent-first layout (i.e., an agent command center). Cursor 3 did this even earlier in April. To quote from their announcement:
We're introducing Cursor 3, a unified workspace for building software with agents. The new Cursor interface brings clarity to the work agents produce, pulling you up to a higher level of abstraction, with the ability to dig deeper when you want. It's faster, cleaner, and more powerful, with a multi-repo layout, seamless handoff between local and cloud agents, and the option to switch back to the Cursor IDE at any time.
This shift is more significant than you might realize. Leading players in the market have thrown away incredibly successful products used by millions of people to create something completely new. Clearly, these are rational, careful decisions backed by comprehensive internal studies and metrics. Is this the final form of agentic coding products? Not necessarily, but for the near future, agent command center will continue to dominate.
What about agent coworker chat apps? I love the concept, and I use OpenClaw from Telegram every day. But the question remains: do I really want another chat app? Maybe not. That's why I think WhatsApp, Telegram, Slack, and Discord will maintain their positions while becoming more AI-native. Besides, a good coding environment still requires features like code editing, debugging, and linting, which chat apps don't provide out of the box.
What's Next?
Today, if you use Codex/Cursor/Antigravity, they look and feel almost identical:
Antigravity 2.0

Codex App

Cursor 3

A natural question is: what's next? I see a few directions for evolution in agent capabilities and product design:
Better Agent Harnesses
Needless to say, agents harnesses will continue to improve. There is a ton of exciting research into dynamic context, safer sandboxes, memory/skill management, and subagents. What I find most interesting is that Anthropic "abandoned" their belief that "You don't need a new agent, just create skills," and instead created agent teams and dynamic workflows. If I understand correctly, this is a loop:
- Previously, when there weren't enough agent trajectories in the training data, you needed a harness layer or prompting tricks to make things work. Now, as LLMs are trained directly on that data, they can complete long-running, multi-tool tasks without much steering, allowing the harness layer to become thinner and more standardized.
- Consequently, people have created more sophisticated architectural patterns like agent teams because LLMs aren't inherently trained to handle them yet. Will LLMs handle this natively someday, creating a harness on the fly to fit a given task? It might not be that far off.
Proactivity: Can Agents Suggest What to Do Next?
People have been talking about proactive agents for a while, but few products actually support them. Jules is one of the exceptions. If you haven't tried this feature, Jules scans your codebase looking for TODOs and performance bottlenecks, generates suggestions, and notifies you. Once you approve a suggestion, Jules automatically creates a PR. I find the performance tips quite useful and have already merged multiple PRs this way.
Beyond that, agents should be able to proactively maintain a codebase by completing tasks such as:
- Fixing code style issues
- Creating tests
- Upgrading dependencies
- Fixing bugs
Some tasks may require external triggers, like a pipeline that automatically converts GitHub issues into user prompts. What is even more interesting—and has yet to reach a consensus—is whether we should let agents proactively suggest new features, let alone submit those changes without human approval. This is very different from standard maintenance work, as the problem space is open-ended with no single right answer. I'm looking forward to hearing stories from teams trying this.
Team Collaboration
Imagine a team working like this: previously, engineers produced 5 commits a day; with coding agents, they produce 12.

This brings two major challenges, which will only intensify as coding agents become more capable:
- Context Isolation: Vital engineering context gets trapped inside isolated agent conversations, leading to a breakdown in shared team knowledge. With engineers shipping code they don't read, even humans lose the full context of how the code functions.
- Coordination Hurdles: Codebases evolve 10x faster with AI. Without a central hub to synchronize intent and progress, teams face broken APIs, redundant efforts, and code conflicts. Existing coordination problems are magnified tenfold.
A possible solution is to build a coordination layer that syncs context across agents and accumulates a team memory. This would act as your team's ultimate PM—someone who knows what everyone is working on and proactively unblocks or helps people when needed. Much like the human brain, long-term memory would store lasting team knowledge (e.g., design choices), while short-term memory tracks immediate team progress and workspace status. SageOx is one startup embracing this exact idea.
Verification: Can an Agent Verify if a Change Actually Works?
I'm putting verification at the end, as it is always the final step in an agent loop (if there is one) 😉.
Many people don't realize how important verification is. However, if you examine agent trajectories, almost all coding agents perform some form of validation at the end. For example, an agent might launch a local server to test a newly added endpoint. If a test fails, it will attempt to fix the issue and continue looping until it passes.
Does this mean verification is a solved problem? Certainly not. The biggest bottleneck remains setting up the test environment. Not all verification is as simple as sending HTTP requests. Changes to a frontend, for instance, typically require browser automation or even recording video and screenshots to be evaluated by an LLM auto-rater—which is a non-trivial process on its own. What if you're fixing bugs for an Android app but don't have the necessary toolchains installed? Yes, you can connect to a remote sandbox containing all dependencies, but then you have to figure out communication protocols and address privacy concerns. While it isn't too difficult to build a custom solution for a specific requirement, turning this into a generic, out-of-the-box component for any coding agent still requires a long way to go.