Goal​

Here we want to share some of our anecdotal observations of what’s happening in the internet market post ChatGPT. We want to discuss some of the market disruption that we have observed and our predictions for the future.

Long tail search market disruption​

There are so many companies like CourseHero, StackOverflow, reddit, wikipedia, Quizlet, and Chegg that used to receive lots of organic traffic from Google. However, we have noticed these websites' traffic has gone down since the inception of ChatGPT. See this article on stackoverflow showing 14% traffic loss in 2023. Also see this article on Chegg stock. Our hypothesis is that lots of younger users who have adopted ChatGPT and other AI assistant tools visit that website to get answers to their questions as opposed to Google in order to see links.

We argue that Google traffic has gone down for these niche long tail searches. One anecdotal evidence for this is this article explaining the Code Red situation at Google. Also Google released an AI Overview for these niche searches in May. Note that they don’t really make ad revenue from this segment.

Some Predictions​

• LLMs are getting better and they will eat more queries/questions over time. That means all apps leveraging AI will keep growing and the Google market will keep shrinking. Innovator's dilemma!
• Lots of internet companies have to think about what the internet is going to look like in the future otherwise they will disappear in 3 to 5 years.
• Companies like OpenAI and Perplexity need to figure out how to turn this to a scalable business model. They might fail but regardless LLM which is essentially a compressed internet will not go anywhere. For example there might be a chance that a new company will figure out the business model outside of the existing AI companies.
• Post ChatGPT the distribution channels are going to be different!

Scaling law and future of LLMs over the next year​

A month ago I heard Elon Musk had built a cluster of 100k GPUs for the first time in Memphis. It seems no one could think this is possible including Microsoft. Some VCs claim this could put x.ai ahead of OAI. Apparently he also is working now on a cluster with 1mn GPUs. Some VCs are arguing that with these larger size clusters he might be able to get ahead of OAI and beat GPT-4 over the next few months with Grok 3. There are obviously so many other variables including if the scaling law really holds as well as if his team is able to pull this off. Interestingly I have seen in their job postings for x.ai they are saying they use JAX & Pytorch.

Overhype vs under hype in the LLM market: I think it’s good to track what is over hyped and under hyped in the market: OAI improvement and its impact on LLM startups: This bitter lesson from Sutton in AI progress is a good way of thinking using history. I think any technique that gets better with more data/compute is the way to go. So one test is to check if a startup product will benefit from this or not. I have seen companies saying that oh we are not competing with OAI and as their model gets better, our system also gets better. One issue with this way of thinking is that if OAI model improvement improves your bottom line (e.g. coding assistant), then what will stop OAI from taking your market? We have seen this with ChatGPT taking a big portion of the writing market already. LLM startups overhype: We see a lot of investments in LLM startups. But, I think we should define what kind of startups could win and which ones have less chance. For example, considering bitter lessons, do startups working on prompt optimization/monitoring, hallucination detection will win? For example we know prompt optimization has become less relevant as models are getting better. A prompt optimization that helped GPT-3 was almost not important for GPT-4. Also the same for startups working on agentic solutions. This gets less relevant as models become more reliable. And historically we have seen models have been getting better since BERT. So I would not bet against the model's improvement.

How AIGC Revolutionized Customer Acquisition in the Digital Advertising Space

When AI-generated content (AIGC) first started gaining traction, the technology seemed like a playground for creatives and technologists. But savvy marketers quickly saw its massive potential in one of the most competitive industries: customer acquisition.

What started as experimental AI-generated images and videos quickly evolved into highly realistic content. This content was indistinguishable from human-made creations, and people couldn’t tell whether they were interacting with a real person’s account or a purely AI-powered entity. Enter Digital Advertising Affiliates—a new, wildly effective approach to acquiring customers at a fraction of traditional costs.

The Rise of AIGC-Powered Affiliate Networks​

In practical terms, here's how it worked:

1. AIGC Meets Social Media
   Affiliate marketers and MCN (Multi-Channel Network) companies began mass-producing social media accounts across platforms like TikTok, Twitter, and Instagram. These accounts leveraged AI-generated videos and images—polished, hyper-realistic, and tailored to niche audiences—to amass followers. Think “hot girl” Instagram accounts or ultra-luxurious TikTok creators, but all powered by AIGC.

2. Cost-Effective Customer Acquisition
   These AIGC-powered accounts didn't just gain followers—they became prime real estate for startups. Startups signed deals with these affiliate networks to drive traffic to their apps and products. Unlike running traditional paid campaigns on Google or Facebook, where every impression costs money, this model replaced expensive SEO or Media Buys with self-owned audiences.

   The result? 20x cheaper customer acquisition costs compared to traditional platforms like Google Ads.

   Let’s break this down with a real-world example. Apps like Rizz (a GPT-powered AI chat assistant) were spending $10 on Google or Facebook to acquire a single customer. With AIGC affiliates, that cost plummeted to just $0.50 per customer—an insane margin boost.

Why AIGC Is a Game-Changer for Startups​

For startups like Rizz, the economics of this strategy are transformative. Here's why:

1. Lower Cost Per Install (CPI)​

Instead of spending $2 or more to acquire a user, this AIGC-driven affiliate strategy brought CPI down to $0.10 per customer. That’s a 20x reduction in cost. The equation changes drastically when you’re not at the mercy of Google or Facebook’s ad pricing.

2. Reduced Risk in Monetization​

With costs so low, profitability becomes far easier to achieve. For example:

• At $2 per user, an app would need at least 20% of users to convert to subscriptions to break even.
• At $0.10 per user, even a 1% conversion rate makes the app wildly profitable.

To put this into perspective: the average app subscription rate in the U.S. is 5%. As long as your app isn’t crashing on the welcome screen, you’re almost guaranteed to win.

3. Eliminating Dependence on Big Ad Platforms​

AIGC doesn’t just lower costs—it frees startups from the monopolistic grip of platforms like Google and Facebook. Instead of paying for every impression, the costs shift to the creation and maintenance of these AIGC-powered affiliate accounts, which have a compounding return over time as they gain followers and engagement.

Why This Strategy Works​

AIGC content has two distinct advantages that make it perfect for this affiliate-driven model:

1. Scalability
   AI-generated content can be produced in bulk at a fraction of the cost of traditional content creation. Whether it’s an AI-generated TikTok influencer or a sleek Instagram reel, AIGC allows marketers to scale campaigns effortlessly.

2. Authenticity (or the Illusion of It)
   People on social media engage with accounts that feel authentic or aspirational. AIGC can replicate the aesthetics and personalities that resonate with audiences, making these accounts highly effective without requiring a human behind them.

How Startups Can Leverage This​

If you’re a startup founder or marketer, this strategy is not just a hack—it’s a playbook:

1. Partner with AIGC-powered affiliate networks to drive users to your app or website.
2. Focus on low-cost content production and account management instead of paying ad giants for impressions.
3. Lower your CAC (customer acquisition cost) to the point where even a modest subscription or purchase rate makes your app profitable.

The beauty of this strategy lies in its simplicity. As long as your app has a decent retention rate, it’s nearly impossible to lose. By leveraging the unique scalability and cost advantages of AIGC, startups can rewrite the rules of customer acquisition, bypassing traditional advertising giants and building a direct, sustainable growth engine.

Definition of Liberal Arts: Liberal Arts is intended to provide chiefly general knowledge and to develop general intellectual capacities (such as reason and judgment) as opposed to professional or vocational skills.

Growing up in two different countries, China and Canada, I encountered a common trend: faculties were often divided into the Faculty of Arts and Faculty of Science. However, my experience working as an engineer at Presence, a pioneering AR tech startup, has taught me that what is often underestimated in the tech industry is the value of liberal arts education.

In some extreme cases, engineers believe that hard skills like coding are the only skills that matter, while liberal arts education is dismissed as irrelevant or impractical. However, I argue that this is a flawed perspective. In fact, liberal arts education can be just as valuable as hard skills for engineers working in the tech industry.

Limitation of Engineering Education​

Throughout my academic and professional experience in the technology industry, I have come to recognize three major issues that are rarely discussed.

1. Fixed Reward Mechanism​

In academia, technical interviews, and in the industry, the standards for evaluating and rewarding engineers are often fixed. Engineers tend to obsess over code cleanliness, optimization of memory and computation usage, and test coverage. While these standards may contribute to the development of better engineers, they may result in less creative problem-solvers overall. In fact, some experts in the field, like Dan Abramov, have highlighted how an obsession with clean code can be problematic. Although there is value in these standards, they prioritize certain skills over others, and consequently, limit engineers' capacity to be well-rounded creators.

Examples of these reward mechanisms include getting an A in a course because your exam answers were elegant, or landing a job offer because you wrote a perfect algorithm that solved a Hackerrank problem faster than anyone else. Additionally, building a better API product than Stripe does not necessarily mean that people will abandon Stripe and use your product.

2. Narrow Field of View​

Engineers often have limited interaction with customers. Typically, they focus on writing code, allowing product managers and designers to define the features that need to be implemented. This can result in a bottom-up working model where engineers remain in the background. However, numerous thought-leaders, like Paul Graham and Clayton Christensen, have emphasized the importance of understanding and interacting with customers. In his essay "Do Things That Don't Scale", Paul Graham remarks that:

A lot of startup founders are trained as engineers, and customer service is not part of the training of engineers. You're supposed to build things that are robust and elegant, not be slavishly attentive to individual users like some kind of salesperson... They'd rather sit at home writing code than go out and talk to a bunch of strangers and probably be rejected by most of them. -- Paul Graham

3. Limitation of Experiments​

Although science is rooted in experiments, many liberal arts subjects such as design, calligraphy, and music produce results that are not always unequivocal. In music, for example, different compositional genres (romanticism, atonality, jazz) will attract and repel different audiences. This subjectivity within certain fields means that experimentation produces more open-ended results.

Overall, recognizing these three limitations of engineering education can encourage engineers to develop broader skill sets so that they can become better communicators and creative thinkers.

Universal Understanding​

The liberal arts were the continuation of Ancient Greek methods of enquiry that began with a "desire for a universal understanding." [1]

I am a strong believer in "Orthogonal Learning", an approach to gathering inputs as much diversity as possible in a learning process. In Computer Science, there is Problem Reduction, which is transforming one problem into another problem. There are many examples of studies in other industries that have helped to explain and solve computer problems:

• For people familiar with Unity or video game development, the quaternion is a common technique for object rotation used in character movement. I was not aware of this terminology but it is no stranger to folks who studied Physics or Aerospace engineering.

Quaternion (Physics and Aerospace)	Leg rotation (Unity)

• Donald Knuth used a lot of real-life problems to describe algorithms problems in his book "The Art of Computer Programming", the example of using train track design to explain dequeue data structure is fascinating:

Track Design (Civil Engineering)	Dequeue (Computer Data Structure)

• And as a music minor student, composing with chord progression is simply building up a finite state automata:

Chord Progress (Music)	Finite-State Automata (Computer Science)

This is an ongoing blog...

Observations​

• Humans unite when there is a common enemy
• Core values and organization cultures can increase productivity of a team
• When people are not confident, they tend to use the bigger identity to hide themselves:
  • I work for a big corp
  • I am from a developed city xxx/country xxx/continent xxx
  • I went to xxx University

About Ego​

Understanding that our identity is not fixed but is rather a collection of ideas and interpretations about who we are can give us the capacity to, when necessary, think critically about what our ego is telling us particularly when the instinct is there to preserve our identity, or ego, at all costs. The ego has gotten a lot of bad rap, but the truth is that the ego is neither negative nor positive. It is merely a sense of self that allows us to operate and function in the world, giving us an idea of who we are and where we belong. The ego is only problematic when it takes control over us rather than our true selves being in control, and when the things it tells us to do or how it is interpreting events becomes damaging to us and/or others.

References​

• Paul Graham's Keep Your Identity Small
• Build your organizational identity
• Understanding the connection between the identity and the ego

For the past year, I have met a lot of entrepreneurs and engineers. I found a common pattern: these knowledgable people not only read more than me but are very smart in choosing the right information to read and digest.

In terms of intelligence, everyone is similar. What matters more is whether a person wants to learn/read and from which source the person learns (the information input). Just like what people always say in the gym -- 'You Are What You Eat', this applies to information and learning.

Information Hygiene 👩🏻‍⚕️​

The first thing to practice is to identify the quality of information. For the modern era, a lot of online media sites and content creators tend to tune their content to maximize the impression rate or view count (if you want to learn how to write news to catch eyeballs 👁 here is a link) a better term to describe the quality of information is Information Hygiene

The following chart is my understanding of information distribution:

Definition of Goals​

• Subjective vs Objective
• Goals can be gamed by blurring its measure of success:
  • ❌Wrong: I made a good product because I think people will like it
  • ✔️Correct:
    • Physical Fact: Humans landed on the moon
    • Economic Fact: People like to pay millions of dollars for a software license per year

First Principal thinking​

• Get to the fundamentals than learning the buzz words
  • ❌Wrong: Learning definitions of Cloud, SaaS, PaaS, Metaverse...
  • ✔️Correct: Understanding FSM "finite state machine", Turning Completeness, Gödel's incomplete theorem... (This is why getting a degree in Business is so less valuable than getting a degree in Engineering. If I give you a book that has the same business strategy as Airbnb, you probably won't build another Airbnb; Whereas, if I give you a snippet of code, whether its 0 or 1, you can justify it in hours)

Why China?​

At the end of 2021, I decided to relocate back to China due to both family and personal reasons. Extended WFH was a huge one: I was losing a ton of passion and efficiency. After evaluating a few opportunities, I decided to join Airbnb, and work out of their Beijing office!

I was talking to my friend who does meditation on a routine the other day. I have always wondered why someone enjoys spending the time to meditate. His theory is that the idea of meditation is to focus on living experience because not everyone is experiencing living. He gave me a few examples:

• When you are enjoying an extremely fancy dinner that costs you 300$, but you know that you are having a technical interview tomorrow morning, you have to go back and sleep early. Are you enjoying the dinner, and tasting every single spice in your food?
• When you are drinking an amazing espresso shot, and someone is knocking on the door which draws your attention away from the espresso. Are you tasting the coffee?
• You can stay in the library for 12 hours and chatting with friends on Facebook. Does that count as studying if your library experience is not studying?

You can first experience the difference between React Fiber (v17+) and pre-Fiber (v16) by playing with the playground below:

You can tell a huge difference between the two versions, the old reconcilor is very slow, but the fiber one is very smooth.

Diff Calculation of React​

React's diff calculation is based on the virtualDOM, which is a tree structure that represents the UI. When a state change happens, React will calculate the difference between the old virtualDOM and the new virtualDOM, and then apply the changes to the realDOM.

However, calculating the edit distance of two trees is NP-Complete, and the standard algorithm needs at least a runtime of O(n^3). There are papers that show that the problem is NP-Complete.

React uses a heuristic approach that compares nodes level-by-level rather than doing an exhaustive node-by-node comparison. While this may sometimes update more nodes than strictly necessary, it ensures no required updates are missed while being much more efficient than a full tree traversal. The algorithm trades perfect accuracy for speed and predictability.

You can see the heuristic approach of diff traversal in the gif below - React compares the tree level by level instead of node by node.

Problem with Recursion​

What’s wrong with doing a full tree traversal for diff in the above animation?

Well, there are two problems with traversing the tree using recursion, we all know in computer science:

1. For any tree recursion, the call stack is O(n)
2. It is impossible to pause the traversal and stop the stack from growing while you are doing recursion.

Here is an interactive slide show of the call stack of the recursion stack:

(You can navigate back and forth using Previous and Next button)

PreviousNext

Solution to the problems - Morris Traversal​

Morris Traversal is a way to traverse a tree without using recursion. It is a linear time algorithm that uses a single stack to store the nodes.

Here is an interactive slide show of traversing the tree with Morris Traversal:

PreviousNext

With Morris Traversal, the call stack is constant O(1) space. Runtime is O(1) for each evaluation. And you can pause the traversal anytime!

This is exactly what React Fiber is doing in their code - adding more path between the nodes to turn a tree into a graph like Morris Traversal.

React Fiber with concurrency​

What you can do with fiber once you have O(1) time and space for each evaluation and being able to pause the traversal is concurrent rendering.

This is how React Fiber achieves concurrent rendering, you can see the animation below:

PreviousNext

Why​

• Your question begins with “Why”, Problem description

Requirements​

• functional requirements - hard requirements:
  • must achieve
• non-functional requirements:
  • soft, but user experience oriented like latency, search speed, autocomplete…