As AI came rushing in like a tsunami, I, too, became an “AI investor” in May 2024. As a newcomer with no background in computer science or product, I have spent much of the past year watching lectures, reading papers, listening to podcasts, trying products, and thinking. I lack the background to go very deep into many technical details, and many of my views and thoughts may contain mistaken assumptions. But I believe that anything that sparks thought is worth reading, and I hope to share some of it with you.

For this first post on my WeChat public account, I want to record the themes and questions I have often thought about over the past year. I welcome discussion. Going forward, I will share questions I find valuable, information I have collected, and my own thinking here.

I stole the name Magic Ink from Bret Victor. Writing is a medium that helps me make my thinking visible and work with it, and it is also easy to share. I have long wanted to put my thoughts into words, both as process data for my own growth and to share with readers who happen to find them. But inertia and procrastination are my natural enemies. As I turn thirty, I have finally decided to escape the gravity of mediocrity.

Models and General-Purpose Agents

What is the main strategic direction for model companies?

In a recent interview about GPT-4.5, Dan and Sam returned to an idea from lectures given two years ago by Ilya (“An Observation of Generalization”) and Jack Rae (“Compression for AGI”): compression enables unsupervised learning to truly abstract patterns and produce intelligence. But pretrained model sizes and the amount of data they compress seem to be growing in proportion. Are they really doing more compression?

In theory, the compression ratio is not a comparison between the data and the model’s size, but between the data and the smallest program that can generate the model’s weights from scratch. When the data is available in the environment to learn from, even storing the weights is theoretically unnecessary. As Dan mentioned in the interview, what we ultimately arrive at through compression is, in fact, a learning rate.

Pretraining is searching for ways to achieve a higher learning rate; o1 is searching for ways to implement reflection and backtracking. Shunyu Yao’s ReAct, too, points to reasoning plus action as something that can generalize across many kinds of problem-solving situations.

In the DIKW framework, a model’s weights may contain a great deal of wisdom and cognition. But the problem OpenAI is trying to solve is metacognition, rather than cognition.

So what is the main strategic direction for a company like OpenAI? My answer is to look for a path to AGI at a higher level of abstraction. Such a path is slow in the short term but will accelerate rapidly in the long term. A higher level of abstraction means stronger generalization, especially the ability to handle more corner cases in general-purpose settings. General-purpose agent companies are still largely expanding their features and use cases in a linear way. If the model companies reach an inflection point, those agent companies could face considerable pressure.

Taste and Vertical Agents

Where do the opportunities lie for most application startups?

Two recent essays by Sutton and Shunyu both mention that once the magic of RL really starts to show, collecting and evaluating data becomes one of the most important things to do in this era, whether through self-play or by hand.

RL brings the generator’s abilities closer and closer to those of the verifier. It also opens up more possibilities for products, because we now have entirely new infrastructure that does not require us to possess a full, complex set of capabilities ourselves. Models let us convey our taste with a very low barrier to entry. But the competition that follows will also be about taste—or, in technical terms, how we do evals.

My intuition is that what comes next is the era of taste.

In an interview, Steve Jobs spoke about bringing together people with the best taste in the world in their respective specialties to build a product. He also famously said that users often do not know what they want. A user’s stated wants may be only a 60-out-of-100 reflection of their needs. In the future, there will be more products that can meet those needs at a 95-out-of-100 level, drawing on the founders’ and teams’ 95-out-of-100 taste and first-principles thinking.

Organizations That Go from Zero to One

Why is OpenAI worth so much?

I think it comes down to its organizational form, built for going from zero to one, and its extraordinarily high talent density. The two are intertwined.

I think Sam Altman is remarkable, chiefly because he has succeeded in building this kind of organization as CEO. When I spoke with friends at DeepSeek last year, I asked them what they thought of Sam. At the time, I had no idea that DeepSeek was operating with a similar organizational form, one that would eventually produce outstanding results.

Why is it so difficult to build an organization like this? I see three main reasons:

1. Founder mode. Greatness cannot be planned. Ideas that take us from zero to one emerge from employees on the front lines, in research and in product alike. They cannot be planned or predicted. The best way to contend with this risk is to build an organization that develops every employee into a potential source of emerging ideas, relying on the possibility of emergence across the whole system rather than on isolated breakthroughs from a handful of people. As a CEO, you need founder mode to obtain information directly from front-line employees and make decisions, because great ideas emerge from that work. Founder mode also seems to have been passed down through Silicon Valley: from Steve Jobs to Jensen Huang, Elon Musk, and Peter Thiel, then to Brian Chesky, Sam Altman, and Dario Amodei. There are countless connections among great CEOs. Brian Chesky has also said, “Jony Ive and Hiroki Asia from Apple inspired me to adopt founder mode.”

2. The Stochastic Mindset. A recent Sequoia blog post said: “The stochastic mindset moves us from having minimal leverage on a task and 100% certainty of its outcome to 100% leverage on a task and far less certainty on the exact manifestation of its outcome.” The vast majority of our organizations do not have this mindset today, yet it seems to be a prerequisite for great ideas to emerge. Historically, only a very small number of academic institutions, companies, and VC firms seem to have possessed this rare ability. In the future, though, models may help more organizations and individuals adapt to this way of thinking.

3. Organizational atmosphere. Why would top talent choose to stay? Viewed through theories of human needs, they are more likely to be seeking growth and self-actualization. Perhaps an organization should create an intellectual atmosphere with high standards, diverse perspectives, and depth, where everyone runs hard in the same direction but in different lanes. Running in the same direction is healthy competition; running head-on into one another is conflict.

In China, the nature of the capital backing such an organization is another enormous challenge.

Emotional Companionship

I once saw Wang Ning tell Li Xiang that if you are building a character or an IP, you should not add functional utility: once it provides a function, it loses its emotional value. I had this in mind while playing Death Stranding. The game tells the story of an emotional bond between the protagonist and a baby whom everyone in its post-apocalyptic world treats as a tool. What moved me was the idea that each of us has a channel connecting us to an object. Our first impression acts like a filter, screening out many of its other attributes. A person or object that offers emotional value, for example, may find it hard to offer functional value to that same person. Of course, over time, people will also perceive functional value in products that offer emotional value. It simply will not be competitive on that dimension. Hence the claim that “once it provides a function, it loses its emotional value.”

The second point is that when this channel is only about taking, it is really just a product meeting my surface-level needs—something like c.ai is, in practice, content consumption. At the heart of emotional companionship is not that it keeps me company, but that I keep it company. Emotion needs to flow in both directions through the channel. When we give emotionally and help the other, we feel a joy that comes from within. This may be an innate form of self-actualization for humans as social animals: an internal need being met. From this perspective, I think the core of companionship is being able to receive the emotions we give and respond appropriately. That is what distinguishes a close friend from a therapist.

Content Platforms

How will new relationships of production be built in the content space?

There is broad agreement around content generated in real time, whether we call it interactive video or games. All of that concerns the reward, R, that people get from consuming it. The less widely shared view is that the cost of interaction, I, may be the key determinant of audience scale. To surpass short-form video, I think that cost needs to approach the cost of interacting with short-form video. Innovation in content formats and interaction experiences may give rise to new relationships of production.

Following the same logic, goods may also develop new relationships of production in the future.

HCI

We have all noticed that today’s interfaces have taken us back to the command-line era. So it is worth putting ourselves in Steve Jobs’s shoes when he saw the GUI at Xerox PARC: what was he thinking about?

I think there are two core differences between a GUI and a language user interface, or LUI:

  1. Humans are visual creatures, and the brain has greater bandwidth for receiving images. Thanks to its parallel architecture, we can instantly register colors, structures, and spatial relationships in an image, whereas language must be processed sequentially.
  2. Language is discrete; graphical interfaces are continuous.

So I believe the interface of the future should also be a GUI, enhanced by AI. On one hand, it could generate customized interfaces in real time to address our needs; Haijun Xia’s research at UC San Diego is worth looking at here. On the other, for creative expression, continuous interfaces give us more room to explore than discrete language. With models as a medium, we can freely combine and extend ideas in vector space, enabling more diverse creative expression. Research into model interpretability may also deepen our understanding of art; Linus Lee’s blog is a useful reference on this.

Vibe Coding

To my mind, the biggest difference between consuming content and consuming software is that content lets us experience unknown worlds. Demand for content can therefore extend almost endlessly through a boundless virtual space. Software, by contrast, meets needs we have here and now, and those needs are constrained by the physical world.

So I have not fully worked out how many new needs there will be in the future—even long-tail ones—that require new software to address them.

And now that AI can also help existing software companies solve the problem of low ROI in serving long-tail needs, I am unsure how much value these kinds of vibe-coded software will have.

Infrastructure for Meaning in Life

In this era, the cost of intelligence keeps falling, as does our comparative advantage as humans in creating economic value. In the short term, we can look for comparative advantages in the new economic structure. But this trend will still force more and more people to think about the meaning of life and their own mission.

The meaning of life, or what shapes who we are, has more than one dimension, and each person assigns different weights to them. To me, the inward dimension consists of every thought and decision we make throughout our lives, and every feeling we experience and express. The outward dimension consists of our connections with others and acts of altruism.

I was inspired by something Allen Zhang said: that WeChat’s greatest significance lies in the ID it gives you. My thought is that a person can have many IDs in virtual worlds, but only one identity in the real world. As the real world becomes digitized, that identity needs a unique counterpart in the digital world. WeChat has taken on the role of that unique digital ID because our social relationships, to some extent, define who each of us is in the real world. WeChat is infrastructure for meaning in life along the outward dimension of connection. I suspect more software will become infrastructure for other dimensions of meaning in life—perhaps by recording our thoughts and emotions?

What Matters Most to Individuals in This Era

Emotions and cognition are woven through our everyday lives. In my understanding, what Eastern and Western philosophy hope we can acquire is an emotional stance toward our emotions and an understanding of our own cognition. These are also what matter most in this era.

An emotional stance toward our emotions: equanimity.

An understanding of our own cognition: mature mental models, letting go of ego and bias, and reflection.

For groups, what matters most is empathy. All of the above may also be the most meaningful education of our time.

Benfen and Equanimity

My understanding of benfen—doing what is yours to do—is long-term thinking and unity between knowing and doing. Equanimity is a powerful inner strength: not being elated by external gains or saddened by personal losses.

Everyone knows these are good things. The question is: how do we acquire them?

Benfen and equanimity both express Chinese philosophy, a philosophy of life that requires enough lived experience and reflection. For benfen, the question to consider is whether you have found your mission in life. That mission may have no apparent origin; the explanation I find most convincing is even that it comes from karma accumulated in a previous life. But finding your mission matters: it guides your direction over the long term.

As for equanimity, having experienced significant highs and lows in life has made it more tangible to me. Many things I once called “wrong decisions,” “wasted time,” or “painful emotions” instead brought unexpected insights and shaped who I am today—and I am happy with who I am now. So I firmly believe in connecting the dots. As long as you remain committed to your goal, the difficulties, challenges, and even low points you face today will pass. The experience of this period will, in time, become fuel for your next ascent. “Reversal is the movement of the Dao; weakness is its way of working.” Just as someone who has lived through an economic cycle can be calmer when the next one arrives, having lived through a cycle of life can bring equanimity toward the peaks and troughs of the next.

Growth

Growth brings me many good things: an understanding of the world, taste, the experience of life’s emotions, and, one level deeper, an understanding of my own cognition, an emotional stance toward my emotions, and so on.

A year ago, I thought growth was an end in itself and one of the most important sources of meaning in life. Now, I increasingly believe that its purpose is to give back to the world the good things I have gained through growing. There are two reasons:

  1. To borrow from Steve Jobs, we eat food grown by others, wear clothes woven by others, and use products invented and knowledge created by others. Almost none of these things are what we ourselves have brought into the world. It is only right that we give something back.
  2. Humanity is one vast, interconnected web. I truly feel that what keeps moving us toward a better world is the flow of good things through that network, whether knowledge or emotion. These things help us progress and bring us closer together.

Everyone can pass on the good things they have learned and felt through growth to others within their circle of influence. Investing and entrepreneurship simply let us use companies as leverage to create a much larger circle. That immense influence is both a capability and a responsibility. It also asks that investors and founders begin from good intentions.

Venture Capital Investing

I often ask myself three questions: Do I enjoy VC investing? If I were financially free right now, would I keep doing it? If today were my last day, would I want to spend it doing this?

So far, VC investing really is my best choice for realizing meaning in my life, for two reasons:

  1. There is still so much in the world that could be better, and investors have more options to choose from.
  2. As a one-star investor, I can support five-star founders in realizing a shared dream. If I started a company myself, I might only be a one-star founder and have far less impact on the world.

The paths I envy most are Sam Altman’s and Daniel Gross’s: as five-star investors, they found something they were willing to go all in on, and committed themselves completely as five-star founders. I hope that one day, when I find my own mission in life, I too will have the inner strength to start from zero.

I will end with a line I recently came across and especially liked:

As long as you have the courage to begin again, your youth is not over.