What being technical means with AI

I think what being technical fundamentally means is to know the possibilities of
technology. My favorite example to illustrate this is Bitcoin founder Satoshi, who
I think is highly technical.

But by technical, I don't just mean Satoshi is good at writing code, which he
obviously is. More importantly, I mean that Satoshi saw a purely technical way to
solve digital scarcity without introducing third-party trust, which is what Bitcoin is.

Satoshi is more technical than I am in the sense that if you asked me in 2009
how to allow everyone to print money, I would've said you need an army and a
state to make sure currency's legal validity is protected.

Whereas Satoshi would've said, this is a technical problem that can be solved
by solving digital scarcity. I think this is the core of being technical: seeing a
world that currently doesn't exist but with certain technological breakthroughs,
can exist.

This is increasingly the way I determine if someone is technical or not:
understanding what's possible, rather than asking them technical questions I
can easily get an answer to from AI.

How a Small Company Can Take Down a Monopoly

Examples include Amazon, Instacart, and e-commerce, which share a
similar business operating structure. Another example is Google and
Perplexity, or YouTube and TikTok.

The reason for this phenomenon is often because the original market
is served by a monopoly. However, within that market, there exists a
smaller section—let’s call it a submarket—that is doing okay under
the monopoly but has untapped potential. If this submarket can be
served better, customers often have a strong demand and are willing
to pay for improved service.

A small company often enters the market much later than the big
competitor and decides to target this relatively small submarket.
They aim to provide better service to users within this niche.
Over time, it turns out that this so-called small submarket is
actually quite large and lucrative.

For instance:

  • Instacart is competing for grocery delivery, which is a
    subsection of e-commerce.
  • Perplexity is competing with Google in a specific area:
    when users want to ask a question in a more natural way
    rather than using keywords, and they want direct answers or insights.
  • TikTok is competing with existing creator platforms
    but has focused on shorter video content,
    which is cheaper to produce and potentially disruptive.

These examples illustrate how small companies can find and exploit
gaps in the market, delivering better experiences in areas overlooked
or underserved by larger competitors. Over time, these small submarkets
can grow into significant opportunities, challenging the dominance of
established giants.

Future of Reading

I spend lots of time reading a variety of books: history, politics,  
economics, biography, poker, physics, computer science...

I think the goal of my reading is to help me:  
1. Satisfy my curiosity.  
2. Understand reality better.  
3. Make better decisions to further my impact.

Reading generally helps me understand what to accomplish and  
why it's worthwhile. Another way of looking at reading is that it  
provides good questions for me. With these questions, I can find  
a more interesting way of interacting with reality (Weltanschauung).

The process of reading books is different from simple  
information extraction. The difference is:  
1. Information extraction expands what I know.  
2. Reading helps me compress what I know, based on new insights.

The second point is interesting because it mirrors how LLMs work.  
Compression  seems to be a key to intelligence: information is  
subjectively interesting when it offers a better way to compress  
one’s worldview.

So, the difference between learning and understanding, in my view, is:  
1. Learning expands the knowledge base.  
2. Understanding compresses that knowledge into deeper insights.  
3. Learning helps in specific situations, while understanding  
    provides a systematic framework.

Now, I've talked about the difference between information extraction  
vs. reading, and learning vs. understanding. Next, I want to address  
the core of this essay: is it possible to create a new, better way  
of reading in the 21st century?

I believe the answer lies in:  
1. Understanding the challenges of the 21st century.  
2. Identifying the essence of reading books—what they aim to achieve.  
3. Exploring the possibility of creating a new, better way of reading.

For the first question, the main issue today is the information explosion:  
information on any subject grows exponentially, doubling every 10-15 months.  
This trend has continued since the 1950s. The traditional solution has  
been labor specialization.

School emphasizes specialized skills—how to do X—over independent  
thinking about the WHAT and WHY. Now, with Gen-AI, there is a chance of  
replacing those who only know HOW. Meanwhile, the What and Why will matter more than the How.

"The present growth of knowledge will choke itself off until we get  
different tools. I believe that books which try to digest, coordinate,  
get rid of the duplication... will be the things the future generations  
will value... In the long-haul, books that leave out what's not essential  
are more important than books that tell you everything... You just want  
to know the essence."  
– Richard Hamming, You and Your Research

For the second question, as discussed earlier, I think the goal of reading  
is to understand and come up with better questions. We should think of  
books as a medium that helps us generate good questions rather than just  
providing good solutions.

For the last question, I think GenAI has potential to make books better—  
more effective at achieving the goals of understanding. Here are some  
thoughts on why:  
1. LLMs may struggle with generating original content but excel at remaking existing  
    existing content in simpler, more concise, and clear words. This alone is powerful.  
2. LLMs offer personalized conversation and can answer questions with 80%  
    accuracy at any time—a rarity among real-world teachers.  
3. GenAI could make complex concepts easier to understand with effective,  
    low-cost visualizations, often leading to new insights.

Why make better books? Books are tools that help us understand and ask  
better questions. Can we create a new product with GenAI that succeeds  
in education, focusing on What and Why, rather than How? Can we make books  
that every leader and CEO reads to think better? And is it possible for  
all kids to have a CEO-level teacher?

The more important questions

AI will solve the HOW question completely. Chat interface is a
beginning. Already, we can see that it's more efficient to code
by giving AI objectives and context. Then debug by giving it the
error message. These steps will further be automated by jumping
from Chat to a Sandbox interface. AI will become an agent with
tools and can make decisions and take actions.

I believe over the long run, AI will need more exposure to reality,
more freedom to make decisions, more resources, and more power to
accomplish more things. Of course, during this process, we need to
make sure AI is accomplishing things for us rather than against us.

So if we are very lucky, i.e., we prevented AI from deviating from
us, it eventually will solve the HOW question for us completely.
But I think the remaining What, When, and Why questions are even
more important. Throughout history, these are the questions people
fight over each other; it was never the HOW question that got us
into deep trouble.

I think AI exposes today's schools' deep flaw in focusing on
training rather than education: 
"Education is What, When, and Why to do things, Training is How to do it."

Recently, I have been thinking about how to utilize today's AI to
teach myself, as well as everyone with curiosity, new knowledge—
but this time focusing on the more important questions: What,
When, and Why. In some sense, with today's AI, we all have a
teacher we never had but wished to: all-knowing, present all the
time, and personalized.

I think if we could build a teacher that truly inspires us to think
about the most important questions, we can be better prepared for
a world with AI and limitless ways of doing new things.

Why product fascinates me

Product is how a company interfaces with its users, at scale.  
Wrapped within a product is a message from the founders about  
a specific aspect of the world: a problem, an ideal, or a new  
way of doing things. The clearer the message, the easier it is  
for people to resonate and become users. In a sense, it's a bit  
like religion. The message can present a different, better way  
to do things.

Code essentially dictates a new way of doing things, and the  
internet helps distribute this way. As more people adopt the new  
way, the way itself improves through the network effect. Product  
building, in this way, feels like having a conversation with an entity,  
call it the market or something else. It's about guessing what  
that entity wants and needs.

However, a product can't make an impact without understanding  
something fundamental about the world or the current way things  
are run. It has to address an existing pain or provide something  
people have already proven to need, but in a more effective way.

How to achieve feats fast

Scaling in the right order.

First, figure out what to scale: people who are fast in executing
understand that proof of concept doesn't need to take that long. A
lot of POCs in many directions in a short amount of time allows them
to see the future in a predictable roadmap. This is why OpenAI is
making significant progress very fast: they tested on directions to
scale—model size, training compute, and recently, inference compute.
These are the scaling laws. Once knowing the scaling law, one knows
where to scale effectively. Each scale brings a multiplicative effect
on the overarching goal.

I think doing things slowly is often caused by doing one thing in 10
different ways, i.e., scaling redundancy. To succeed, there is often
no more than 2 things to focus on to scale at any given time. A useful
mindset is to think about "how to achieve 80% of the effect with 10%
of the effort"; this helps bring focus down to what really matters.

Second, it's worth thinking about in which order the scale should
happen. Scaling in the right order means that doing things in a
specific order can make doing the following thing more effective.
Think about why language was the initial entry point of AGI: it's
the lowest cost abstract storage of information about the world,
and we have plenty of it already, and can create more easily.
Whereas images and videos are less supervised, and they are more
costly for representing information. Text has the beautiful property
of filtering out what doesn't matter and only focusing on what
matters. So making language models really well pushes forward other
directions. It's like a brick that is interlinked, and pushing one
direction moves the others as well.

Moore's Law

The focus today should be on leveraging better hardware, thanks
to advancements from Moore's Law, rather than trying to compete
with this trend through software tweaks. We should think about
utilizing the massive computational power we’ll have in the future—
potentially 1000 times more than today—instead of attempting to
squeeze small gains from current hardware.

This approach requires us to rethink how we train models, even if
they may not seem efficient now. Over time, this shift will pay off
significantly. While short-term gains come from optimizing domain-
specific aspects, long-term success relies on broader, more adaptable
capabilities (Chap.28: system engineering). The “bitter lesson” is that
real progress doesn’t come from perfecting small tricks, like making AI
only better at writing Python code, but from developing general skills—
understanding, reasoning, learning, planning, and predicting.

Moore's Law suggests that not only is high-end computational power
improving, but the cost-performance ratio is also decreasing more
dramatically than for lower-end hardware. This is why the iPhone,
despite being expensive, becomes more accessible over time and
outperforms cheaper alternatives on the market. High-end hardware
benefits from exponential growth in performance while becoming
more affordable, whereas low-end hardware sees only incremental
gains and limited cost reduction. This explains why there are few
high-quality consumer electronics priced below $100—performance
improves slowly at this level but accelerates significantly beyond it.