Editor's Note: As AI begins to write code, optimize code, and gradually take over the software production process, a deeper structural change is approaching: professional division of labor, corporate organization, and even knowledge barriers may be redefined.

The author of this article managed a team of nearly 20 people at a hedge fund but chose to leave during the peak of his career to start a business. In his view, the real signal is not market sentiment but the leap in technological capability. When models can stably generate usable code and possess recursive improvement capabilities, the logic of software development and knowledge production has already begun to change.

From the perspective of quantitative finance, the article analyzes several types of short-term "moats" that may still exist in the AI era, including proprietary data, regulatory friction, authoritative endorsement, and the lag of the physical world. It also proposes a core judgment: in a highly uncertain era, what matters more than accurately predicting the future is identifying the direction and taking action before the window closes.

Below is the original text:

When Models Start Writing Code, the Change Is Irreversible

I first realized the industry was approaching a tipping point during my previous job. It felt like the background music was slowing down, yet everyone around me was still pretending nothing would change.

At the time, I was managing a team of nearly 20 people at a hedge fund, doing what I had been doing for many years. From the outside perspective, it seemed like a steadily rising career path. If I had stayed, I would likely have achieved even greater success. But ultimately, I chose to leave that coveted position to start a company from scratch with just a handful of people. At the time, almost no one understood this decision, and it was even seen as "career suicide."

But in recent months, with large-scale layoffs, voluntary resignations to start businesses, and more people working day jobs while quietly coding on projects at night, that seemingly "crazy" decision doesn’t seem so far-fetched anymore.

During this time, many people have asked me: Where is all this ultimately headed? This article is my current answer.

Frankly, I’m not sure how big the changes will ultimately be. But one thing quantitative finance taught me is: Being directionally correct is often enough.

What truly made me realize the change was irreversible was ChatGPT’s o1 model.

Before that, I had always referred to these systems as "LLMs," not "AI." I didn’t believe they truly possessed anything接近 intelligence. But when o1 emerged, something changed: these models could, for the first time, stably generate code through structured prompts.

The code wasn’t perfect—it still had hallucinations or misunderstandings. But the key was: it could write useful code.

My judgment was simple. Once AI can generate usable code, it will begin to recursively improve its own logic and drive software development at an unimaginable speed.

Whenever I point this out, someone always argues, "This code still has bugs; it’s far from production-ready." But this恰恰 ignores a fact: human-written code also has bugs. We won’t stop writing code only when AI writes perfect code.

The real turning point is when AI’s code has a lower error rate than humans while being much faster. At that moment, writing code will be彻底 outsourced to machines.

After亲眼 seeing o1’s capabilities, I was almost certain: very dramatic changes will happen.

Moats That Still Exist in the AI Era

Initially, I thought AI would gradually erode the quantitative finance industry, but the process would be slow. The reason was simple: institutional-level code has almost no public data for training.

At the time, I imagined software engineering as a pyramid: the bottom layer is basic coding work; above that are senior engineers with architectural capabilities; further up are specialized developers, such as data scientists, quantitative developers, and various industry experts. Theoretically, the deeper the expertise, the safer the career.

My judgment then was: within two years, basic programmers would be淘汰 first; followed by senior engineers; and as models gradually absorbed专业知识, higher-level positions would also be impacted.

But I soon realized something else: leading model companies would eventually directly hire industry experts to input专业知识 into models. In other words, expertise would indeed be a short-term moat, but in the long run, it would also be gradually digested by models.

In that initial assessment, there were several types of businesses that would not be easily颠覆 in the next five years.

First Type: Proprietary Data

Companies with large amounts of proprietary data are harder to replace.

For example, large multi-strategy hedge funds (pod shops), like Millennium, generate massive amounts of data every day: analyst research, investment recommendations, market judgments, actual trading results.

This data can be used to continuously fine-tune models, creating advantages that are difficult for outsiders to replicate. As long as the company’s data sources are not easily accessible to models, it still retains a moat for some time.

Second Type: Regulatory Friction

Any industry requiring significant human approval is not easily quickly颠覆. For example, traditional financial markets.

To enter these markets, you need to: open brokerage accounts, obtain licenses, sign cross-border legal documents. Trading crypto assets is easy, but it’s far less simple for a foreign company to trade iron ore in China.

As long as an industry still requires human signatures for approval, its development speed will inevitably be limited by the approval process.

Third Type: Authority as a Service

Now, it’s not difficult to have AI write a legal opinion. But the reality is, people are still willing to pay tens of thousands of dollars for a lawyer to issue a legal opinion. The reason is simple: AI’s opinions currently lack authority.

Smart contract audits follow the same logic. Technically, AI may already match or even surpass top auditors. But the market still prefers to buy the "stamp" from a well-known audit firm.

Because what customers are really buying is not the opinion itself, but the authority behind it.

Fourth Type: The Physical World

Hardware进步速度 is much slower than software, and hardware problems are also harder to fix.

Therefore,实体 industries that interact directly with the real world are unlikely to be quickly颠覆 by AI in the short term. However, once hardware capabilities catch up, the same logic will still apply: lower-level jobs disappear first, followed by higher-level ones.

These moats do exist. But it must be admitted that they only delay change, not prevent it.

Act Based on Signals, Not on Waiting for Certainty

When the future is highly uncertain and changing rapidly, people typically make two kinds of mistakes.

The first is waiting for certainty before acting. The second is simply applying historical analogies, such as: "This is just like the dot-com bubble."

Both approaches can lead to misjudgment.

When information is incomplete, a more reasonable method is to reason from first principles.

You don’t need to know every detail of the future. You only need to roughly judge the direction and design asymmetric wagers—that is, if you’re wrong, the loss is controllable; if you’re right, the gain is huge.

In an uncertain future, asymmetry is everything.

A practical thinking method is to first ask yourself, "What prerequisite conditions need to be true for a certain outcome to happen?" Then ask, have these prerequisite conditions already appeared?

Looking back at this AI inflection point, it wasn’t hard to foresee. Because the key inputs早已 existed: code that can write code, models that can recursively improve, institutional knowledge that can be bought rather than cultivated.

By carefully observing these signals, one could大致 judge the future direction.

We can even continue to extrapolate.

We probably haven’t truly seen the following scenarios yet: AI can train itself, AI can replicate itself, AI operates完全 autonomously.

If an AI can improve its own capabilities by 0.1% through a series of actions, it sounds small. But as long as that number is not 0, it will amplify. Behind this is a typical power-law effect.

In financial markets, once a signal becomes obvious, the trade is often crowded.

In investing, you exchange uncertainty for early conviction. In careers and entrepreneurship, it’s essentially the same.

So the real question is not, what will happen in the future? But rather, what do I already know? What direction does this information point in? What is the cost difference between acting now and waiting?

Another often overlooked fact is that action itself creates information.

Action doesn’t happen in a vacuum. When you take action on the world, the world gives feedback. This feedback brings new information. Information drives iteration. Iteration produces better actions. This is the basic mechanism of progress.

Remaining stationary amidst uncertainty is a slow decline. Action, however, means exploration.

If I just wanted to continue eating the红利 of the existing system, I could probably维持 for a few more years. But I’ve always wanted to do something truly my own, and I felt this window was closing rapidly.

Of course, the world’s largest hedge funds will still do well; they have proprietary data that is extremely difficult to replicate. Traditional financial markets are also still constrained by regulation and manual processes.

But I believe these institutions will eventually also use AI to replace the vast majority of their employees, even including portfolio managers.

It won’t happen immediately, but it will happen eventually.

My judgment at the time was that I had about a 4–5 year window. Once foundational model companies absorb enough industry talent, it will be very difficult for new startups to enter this field. In some markets, like the US stock market, this trend is already very apparent. How efficient things will be a few years from now is almost unimaginable.

Soon, there will be no room for "second place" in this world. I could continue working for the top institutions, but I prefer to make my move in areas where I still have an advantage.

So I resigned and went all-in on starting a business. Later, this company became OpenForage.

Now, the window is明显 narrowing. The pace of change is no longer gradual. Progress that used to take months now takes weeks.

I don’t believe jobs will completely disappear in the coming years. Humans will still need humans. People are social animals, and humans still don’t trust AI. Authority certification still needs to come from humans.

In the next few years, we might even see AI CEOs, but there will likely still be a need for a human CEO to approve the AI’s decisions. This "human certification" will be transmitted layer by layer through the organizational structure. Human managers will manage a group of AI agents.

But hiring logic will change. If it’s easier for a CEO to give instructions to an AI than to you, then you likely won’t be hired. Basic coding jobs will become increasingly hard to find.

If you want to make yourself irreplaceable, you need to do two things. First, operate on a timescale beyond AI. For example, long-term strategic planning, complex decision-making, multi-year cycle management. Second, operate on a system scope beyond AI. AI’s context is still limited; they know many facts, but they struggle to understand the连锁反应 of complex systems.

If you can think long-term, quickly absorb information, make long-term decisions, and have good collaboration skills—then, for the foreseeable future, you will still have a job.

Before the inflection point arrives, the signals can actually be seen. It’s just that most people don’t look, don’t act even if they see them, or only react when the signals become deafeningly loud. But by then, opportunities have often already been priced in by the market.

Don’t ignore the ground that is moving. Don’t stay in a position that is losing its advantage while telling yourself to wait for a better time. Real opportunities rarely give advance notice. When everyone realizes it, the window has often already closed.

I saw the signals, I made the bet. Now, I am living in the outcome of that bet—for better or worse.