Jiu-jitsu is honest

Jiu-jitsu is honest, man. Or sports is.

you have to work for it. A stripe in your belt. Details of a move, direction of your toes. You saw issues in practice, you thought to fix them. There is no shortcuts of it.

consistency is the basic key. Show up. Practice when you don't feel like it. You need the time in gym. 

You earn the respect in compete. 

It's honest as fuck, maybe an antidote to too many other things in life. I am so in love with it.

Presence. 



The Mary of Gold moment / AI, R&B, and Merchandise

“Hey darling, check out this song,” I forwarded a Spotify song to my girlfriend, “I love it.” Two weeks later, I shuffled to this song, fell in love with it and brought it up again to her. This time, we started an interest on the artist and googled her.

There was nothing.

Well, there were some reddit notes mentioning that the artist (Mary of Gold), could be an AI artist because no picture, no history can be found about her.

I didn’t want to debate about AI copyrights in this blog; I am clearly not an expert. What intrigued me was my reaction. Clearly, the song was still very good, and I did enjoy the voice and the progression. But knowing that I got personally engaged with a song that unexpectedly turns out to be generated by AI - it all felt different to me. 

I started to wonder: should I not be listening to this song? Is that “discrimination”? But clearly it was so good. If there was a R&B version of Turing test, this would have passed, maybe. 

If there are two songs, one written by a human and another written by AI, how should we judge them. Should they be judged by the same standards? I think eventually they will be. It is just like man-made vs machine-made redwood furnitures - eventually the market will judge its value. Maybe there was never really one standard of “judgement” - there have always been tiers, sometimes non-conforming standards to price things.

I later read a social media post from one of my non-technical friend, sharing her first experience of “vibe-coding.” It became clear to me that execution, or making a demo-ready app is now going to proliferate. People would no longer be bogged down by, “I have this wonderful idea, I just don’t have time / know how to put it into work.” Now everyone can do it. 

What’s going to be more valuable? 

  1. An easy answer: “better” idea is going to be valuable. Ideas that really solve people’s problems and helpful, are going to be important
  2. Resources, like capital, distribution, infrastructure, political powers, energy: those are still going to be scare and will impact the influence of your app.
  3. Laws of economics will still rule: people may run to publish their app online - but is it going to generate cash flows that feeds its operations? If not, a lot of the development will end up vanishing. 
    1. Caveat: will the concept of “money” still stand in an extreme version of merchandise proliferation?
  4. Laws of physics will still rule: to assemble things, you still need to put a bunch of articles together. That’s a lower bound of resources you will need, and the resources are not infinite.  

I tried to think of a metaphor in previous history where there was such a proliferation of merchandise and individual “makability”, the Industrial Revolution comes into my mind. We are, in the very beginning of an industrial revolution. A lot of new things will be built, many of it will not sustain the laws of economics, but human creativity will take us to an era that we couldn’t imagine before.

I am excited about it.




i believe in AI boom

Using Claude, I was able to prototype my idea of putting several financial metrics in a dashboard in under 2 hours. I probably had 4 or 5 versions.

It does unleash productivity. Code is the first tool that it learns well. People who understand code obtain initial, albeit huge, leverage (the concept of the 10x engineer). Productivity increases for these folks.

Eventually the UI will get better, and it will learn more tools people are using. The AI inference need there is real.

When building becomes easy, deciding what to build becomes more important. “Taste” and “design” will differentiate.

everyone writes prompt; value-add as a data scientist

[drafting]


We can teach subject-matter experts prompt engineering—but then what is our value-add as a data scientist?

  1. Technical connector. We serve as the bridge between SMEs and engineers. We build PoCs, run MVPs, and prove technical viability.
  2. Safeguarding CI/CD. We act as stewards of the CI/CD process, enforcing best practices that may be unfamiliar to SMEs, such as proper version control.
  3. Fluency in frontier research. We stay well-versed in cutting-edge research, understanding how the latest developments can help us iterate on models. This requires deep familiarity with deep learning and neural network literature.
  4. Disciplined model development. Evidence-driven development, edge-case handling, and robust test cases are essential to ensuring feature stability. Today, almost anyone can build a demo—but systematizing it reliably remains a core expertise.


the music in state capitalism: working with the regulations in China

In his book Money Machine, Shan Weijian describes the ups and downs of working within China’s regulatory system in the 2000s. I found it insightful to observe several recurring behavioral traits.

Unless otherwise specified, “the bank” refers to Shenzhen Development Bank, which Newbridge acquired a controlling stake in 2004 and sold to PAIG in 2011.

1. Principles of social engineering

I had the impression that officials adhered to strong, often top-down, social engineering principles and used them to guide even the finer details of transactions.

For example, when Baosteel (a highly profitable state-owned steel company) expressed interest in acquiring the bank, regulators shut it down because, in principle, an industrial raw-materials company was not allowed to own a bank. While such structures exist elsewhere (for example, Mitsubishi owns MUFG in Japan), current regulations in the US and EU prohibit this to prevent banks from becoming “internal treasuries” for industrial firms.

2. Intense leverage and even threats

When the government pursued the share reform in 2005, what began as voluntary participation quickly became coercive. Out of concern that the release of legal-person shares would flood the market and depress prices, a common arrangement required participating companies to grant additional “free shares” (about 30% of the traded amounts) to public shareholders.

Newbridge chose not to participate, worrying that this arrangement would significantly dilute its roughly 20% stake. In response, stock exchanges—under pressure to advance the reform—effectively blocked the bank from releasing its quarterly financial statements, forcing it to violate disclosure requirements for listed companies.

Note: China’s 2005 share reform aimed to make historically non-tradable legal-person shares tradable on stock exchanges.

3. Multiple, often competing, regulations

To obtain approval for the deal with PAIG, both parties had to navigate multiple layers of government—from municipal to provincial authorities, and ultimately Beijing—while also dealing with sector regulators such as the China Securities Regulatory Commission and the China Banking Regulatory Commission.

At best, these regulatory requirements were overlapping; at worst, they were contradictory. For instance, when the bank sought to issue new shares to raise capital, banking regulators supported the move because it strengthened the bank’s capital base and reduced risk. Securities regulators, however, opposed it, arguing that an increased supply of shares would depress the stock price.

4. Publicity, “face,” or pride.

Anticipated public perception is a critical factor in decision-making, because unfavorable perceptions could be detrimental to one’s “face,” or career. When Newbridge filed an arbitration in Paris against the Shenzhen government for breaching the binding transaction agreement, it cracked the government’s once-firm stance on not proceeding with the case. There were signals suggesting that Shenzhen was willing to resume talks in exchange for a withdrawal of the arbitration.

I’d like to note that this does not mean a simple “talk-or-sue” strategy works when dealing with Chinese regulators. In fact, the most important factor may have been that certain top-level officials were still interested in seeing the case go through. Leveraging that momentum, together with the pressure created by the publicity surrounding the arbitration, contributed to the thawing of the negotiation impasse.



First draft: Jan 11, 2025

Second draft: Jan 14, 2025

a Korean lesson: arts of negotiation

The Art of Negotiation
Recently, I’ve been reading Shan Weijian’s memoir about leading Newbridge Capital’s acquisition of Korea First Bank. I was deeply inspired by the techniques and artistry he demonstrated in negotiation. Below are a few principles I’ve distilled to help form a more systematic understanding.

1. Always Think in Terms of Bottom Lines
All of Newbridge’s proposals were built on one fundamental principle: the investment had to make commercial sense. Based on this, clear red lines were established—rules that could not be crossed or compromised. For example, when the government demanded that Newbridge’s investment return not exceed 12%, Newbridge flatly refused.

2. Understand the Consequences of Concessions Through Clear Calculations
Should non-performing loans be priced at 95% of face value or 96%? What does that 1 percentage point concession mean for overall returns? Only with a clear, quantified understanding can you know how many cards you hold and how much room you truly have to maneuver. An operational, flexible financial model is a must.

3. Play Your Cards One at a Time
If you can afford to concede three dollars, start by conceding one. Save the remaining two as bargaining chips for when you need leverage later in the negotiation.

4. The Role of Agents as “Messengers”
Agents and intermediaries serve two key purposes:
They help establish trust between previously unfamiliar parties and bridge differences in position (for example, diplomats or ambassadors acting as intermediaries between foreign investors and financial regulators).
They create an informal “buffer layer” through which tough messages can be delivered without triggering immediate or irreversible consequences.
For instance, Newbridge’s “ultimatum” in September 1999 could not be formally delivered to the government—politically, a government yielding to a foreign investor’s ultimatum was unacceptable. Instead, the message was conveyed via an ambassador: “This is the last opportunity. If it’s not signed now, there will be no chance to sign later.”

5. Managing Public Opinion and Lobbying Politicians
Winning the media battle can be important, but it often creates only localized pressure, and the direction of public opinion is hard to control. Shan maintained regular contact with journalists, which not only helped shape public perception but also provided insights into political dynamics and the progress of a parallel bank acquisition. At the same time, this engagement helped establish Newbridge’s image as professional, rational, and commercially driven in the eyes of the media and the public.

6. Understand Your Counterparty’s True Interests
For financial regulators, the political and regulatory implications of certain agreements may matter more than the precise monetary amounts involved. Once Newbridge recognized this, it could adjust the form of the terms—without crossing its own bottom lines—to better satisfy the counterparty’s political needs, while leaving its financial returns intact.

7. Be Patient: Sugar-Coat Your Punches and Threats

Expressing firmness without breaking the negotiation is an art, especially when the other side makes unreasonable demands.

Instead of saying, “You’ve gone back on your word and broken your promises!”, say:

“We are confused and disappointed by your new proposal, which represents a significant step backward from the Hong Kong discussions that took place just one day ago.”

Instead of saying, “We’ve invested so much time and resources in this negotiation—we won’t walk away empty-handed, and we’ll sue if we have to,” say:

“We entered this process in good faith based on the government’s commitments, and therefore invested substantial time and resources to complete the transaction. Our investors will not allow us to walk away quietly. We will do our utmost to protect our rights, though we hope to avoid a mutually destructive outcome.”

Negotiation, at its core, is not just about toughness or compromise—it is about clarity, patience, and the disciplined use of leverage.

why supervised fine-tuning is infeasible

We hoped that fine-tuning is best-suited for a niche purpose. However I don’t see it as the future of adapting generative AI in industry. Here’s why:
1. Dependency on high-quality annotation. Cliche, but it is hard and expensive to get. The process is not always controllable, and often very time consuming (e.g. the annotation team may mis-interpret the annotation guideline)

2. The foundational capability of LLMs are only getting better. With appropriate prompts, it could solve a large chunk of problems.
3. Fine-tuning can make a model worse. If the training data contains misleading / bad signals, it could actually mislead the models and “erase” some alignment / intelligence originally possessed by the foundational model. A version of “garbage in, garbage out”.
4. Human cost is much higher compared to the computational cost. With tools and techniques (e.g. LoRA, llama-factory), the computational cost of fine-tuning is actually low; however, the human time involved in updating the annotation guideline, debating the wordings and standards, etc are the larger cost.
5. The marginal cost of obtaining new annotations is just much higher than updating the prompts. Fixes an imperfection in annotation guideline could take weeks to turn around. Fixes on prompts, by contrast, could be much lower.
6. It is much easier to manage a library of prompts compared to a library of fine-tuned models.