Translated by Claude from the Chinese original.

It all started with a question I was asked today.

1

Today, a recruiter at my company came to ask me: “Do you think OpenAI is a phenomenal success or just a coincidence (with low technical barriers)?” A candidate had told her that OpenAI’s technical barriers aren’t that high—they were just the first to take the plunge. Looking at it now, OpenAI’s progress has indeed slowed down, and from what I know, other companies are catching up fast, even surpassing OpenAI in some models. But in my mind, OpenAI remains special. I believe OpenAI’s value lies not in its commercial success, but in its technology. Besides, while OpenAI may not be extremely successful commercially, it certainly can’t be called a failure.

Those who followed OpenAI early on would know that GPT and GPT-2 were seen as mere “toys” by outsiders, while Google’s BERT was considered the strongest language model at the time. When GPT-3 came out, people found it interesting but still thought it was useless. It wasn’t until GPT-3.5 that people saw the possibility of AGI. Before GPT-3.5, OpenAI had been persisting with the GPT architecture for over four years. Even with substantial funding from Musk, Altman, and others, it was incredibly difficult for a startup to keep investing in a direction that might not succeed.

In my view, OpenAI’s persistence is what truly differentiates it from other companies. Other companies build large language models because of commercial necessity; OpenAI builds them because it believes this path can create a future with AGI. This is also why what other companies do feels “meaningless”—after OpenAI released GPT, if a newcomer wants to contribute to humanity, they shouldn’t choose to work on large language models. Unless you have insights vastly different from the mainstream, this field has already been thoroughly explored by OpenAI. A newcomer should independently think about what other important fields are being overlooked by the market and remain unexplored. Those are where new value and technology will be created. In other words, as a newcomer, you should prioritize doing new things rather than rushing to do old things.

2

Some might disagree with the above—just do whatever makes money, right? But I believe that markets without competition are where the profits are. In the crowded LLM space, unless a clear winner emerges, everyone ends up working at cost.

Take China’s “AI Four Dragons” (SenseTime, Megvii, Yitu, CloudWalk) as an example. Their technology seemed difficult, but they were all highly homogeneous. Eventually, after using their solutions for a while, big companies turned to building in-house, leaving the Four Dragons to pursue ToB and ToG routes. Of course, computer vision also has the problem of limited commercial value, but that’s another story.

Another example is Xiaomi. Xiaomi always enters markets after demand has been validated, which is why they make “heartfelt” high-value products. There’s nothing wrong with this business model, and Xiaomi is a company with great social responsibility. But the reality is that Xiaomi’s hardware profit margins are extremely low, leaving little room for innovation.

So circling back—is building LLMs profitable? Personally, I think it’s a money-losing business. The only exception seems to be ByteDance’s Doubao, because ChatGPT can’t serve China, and Chinese users need a good LLM app.

3

Compared to other companies, what the big tech companies are doing with LLMs is actually the “old thing” that OpenAI already did. What OpenAI did with GPT years ago, with multimodal in the past two years, and recently with post-training—those are the “new things.” I personally prefer doing “new things,” for the following reasons.

First, doing new things is meaningful. OpenAI’s years of persistence on AGI led to today’s AI application explosion. New technology creates new demands and resources, while old things mostly just redistribute existing resources.

Second, maximizing returns. If a new thing succeeds, you become the industry leader with substantial commercial profits, and abundant capital allows you to do many things. In contrast, profits from old things are always limited because markets and user demands gradually solidify, and competitors want a piece of your pie. The success rate of new things is actually quite low—failure is the more common outcome. But failure also brings significant returns. In the 1990s, a company called General Magic tried to create a portable touchscreen internet device—something that looks very much like an iPhone today. As you might expect, 1990s technology couldn’t produce an iPhone. General Magic failed, but many of its employees went on to have successful careers. Tony Fadell left General Magic and eventually joined Apple, leading the design and development of the iPod. Kevin Lynch led the development of the Apple Watch. Andy Rubin wrote an operating system called Android that was acquired by Google.

Third, leaders have a huge advantage in copying others. While the “new things” leaders create will be copied by followers into “old things,” leaders can also copy the “micro-innovations” of followers. Examples abound—from Tencent to Apple to OpenAI. All followers of these companies are merely helping leaders validate demands they didn’t have time to validate. Once validated, leaders can copy at minimal cost, and even take over the followers’ original customers.

Fourth, being misunderstood is actually a huge advantage. Because short-term profit-seekers can’t see you, you can easily filter out employees who don’t share your vision. Because what you’re doing hasn’t become a trend, you can acquire resources from suppliers at very low cost. And because there’s no competition, you won’t have too much time pressure—the rhythm and feel of doing “new things” will be much better.

4

If we agree that doing “new things” is better, what practices can we adopt?

1. Career: Prefer companies that do “new things.” This way, you not only gain more personal growth but also have a small chance of getting significant financial returns.

2. Investing: Prefer companies that do “new things.” These companies have obvious characteristics—they focus on things others don’t want to do, so their stock prices are low. But once these things succeed, they’ll have great commercial value, and stock prices will rise significantly. However, judging things and companies isn’t just talk—I don’t recommend spending too much time or money on investing.

3. Daily work: Shift your evaluation criteria from “completeness” to “innovation.” I used to have a bad student mentality—I could get 80% of the results with 20% of the time, but I’d spend 100% of the time to get 100% of the results. The extra 20% is often trivial or unimportant stuff. Now, for me, quickly delivering results and validating the value of ideas is more important. If your 80% is important enough, many people will be willing to do the remaining 20%.

What you don’t know is always more valuable than what you already know.