From the airport, you can take the light rail directly into the city, unlike many major US cities. For just $3, you can take the train directly to Pike Place and see the world-renowned fish throwing inside the market. A few more steps outside and you’ll hit the beautiful gum wall, which pictures truly cannot capture.

Inspired by the adventurous outdoors, you walk 1 mile to the Space Needle, which is just a unique design and architectural achievement.

After admiring the Space Needle, you venture north to the University of Washington campus.

The natural beauty around campus and across greater Seattle is breathtaking due to the persistent rain. You’re ecstatic to find a dedicated vista of Mount Rainier from the main square. You’re lucky enough to be there during the spring cherry blossom season and take the requisite Instagram photos.

You head down to Dick’s Drive-In and order one of their famous burgers. You look through the line to see if Bill Gates is there, but you came on the wrong day. The people in line and the employees are so nice to you, a welcome change from what you’re used to in New York.

You viscerally understand what the 12th man embodies: being a part of a community that cares about what’s happening in the world. As you become ensouled by the 12th man, you have a revelation.

Before you even touch down in Seattle, you realize this city was built by idiots. Fly from somewhere else on the West Coast and you’ll find yourself spending 15% of the flight in Canadian airspace beyond SEA-TAC before turning around to land facing south. Of course there’s no secondary airport like every single other major city in the US, so you’re stuck being routed to an airport with one of the worst layouts in humanity.

With Seattle traffic being some of the worst in the nation, you’re forced to walk 25 minutes through the adjacent parking garage to the light rail station. You buy your ticket from an antiquated machine from the 80s and head up the escalators to the train station. There are a lot of other passengers waiting for the train as well; curiously there was no line at the ticketing machine. You get on the train that comes once every 30 minutes and realize that you are the only schmuck actually paying for the ride. The Seattle government thinks that somehow they’re actually in Singapore and that fares can be enforced through an honor system without turnstiles. After reading about the competence of Seattle city officials on the train, you’re just thankful that you aren’t stabbed on your ride to the city.

You go to Pike Place and realize that the main attraction is two people disrespecting fish by throwing a fish across the room every 15 minutes while yelling and a wall made out of other people’s chewed up gum and fermented saliva.

Then you walk over in the rain to the Space Needle. Like any sane person, you use an umbrella to keep yourself dry. You receive dirty looks from the Seattleites who look down on the use of umbrellas and prefer to stay miserable in their Patagonia jackets.

When you finally arrive at the Space Needle, you learn that the Eiffel Tower built over 100 years earlier still clears, and that the free glass exhibit at the Bellagio is better.

It finally dawns on you that this entire city has Stockholm Syndrome from Tom Brady breaking everyone’s hearts in Super Bowl 49. All the 12 numbers you see around the city are actually a tribute to Tom Brady, begging for forgiveness.

For too long, it was difficult for me to articulate exactly what’s wrong about Seattle. The word that accurately and concisely describes Seattle is dishonest.

Every city curates a myth, which drives downstream culture. New York tells you that you should make more money. Boston tells you that you should be smarter. Los Angeles tells you that you should be more famous. Silicon Valley tells you that you should leverage technology to become more powerful.

Seattle tells you that you should care more. Care more about other people, the environment, and the world. Of course this is a noble message, but Seattle is only concerned with the appearance of caring. Because there is no legible metric for caring, the entire city and culture is constructed around a system where feeling good about yourself is indistinguishable from actually doing good. These poor incentives predictably set a suboptimal cultural equilibrium.

The city’s surroundings are objectively beautiful. But if you listen closely to the raindrops, you’ll find that Seattle consistently lies to its residents, purporting to be a city of depth, seriousness, and substance. In reality, it’s a city that has outsourced its identity to its geography and its virtue to branding.

Popularity prediction hash: 9ef965eedfc5cc296b95343b6b8d7284de3dfc12116707b6952b0adf5c83cdf5

The real story of how Axiom was able to achieve such growth in less than a year since launch has not been told, as they are a discreet team and do not publicize much about the team, culture, or processes behind their product.

For context, the following is a timeline of the saga:

Developments are ongoing and this post is not meant to serve as investigative journalism. Rather, it explores how a two-person team with no prior crypto experience broke into the industry, built the fastest company to $100M in revenue in YC history, and what their trajectory reveals about the shape of elite startups going forward.

Axiom (YC W25) is a trading interface primarily targeted towards people trading long-tail tokens (polite term for memecoins) on Solana. Axiom was founded by two individuals and hit $100M in revenue in just 117 days. This is over twice as fast as the most hyped AI and financial applications such as Cursor, Lovable, and Ramp that consistently dominate headlines and mindshare.

Interestingly, one of the few public advertisements for Axiom is this campaign run by the Solana Foundation and a VC firm who isn’t even on their cap table. The ads aim to legitimize crypto as a viable industry to prospective founders in San Francisco:

Axiom represents a new company archetype: one that sits at the intersection of relentless focus, monetary incentives, willingness to enter competitive markets with mediocre talent, and a third-derivative product.

In this post, I explore how Axiom became the fastest YC company to $100M in revenue and one of the highest revenue per employee companies ever.

1. Background on Axiom

   1. The Team Behind Axiom

   2. Axiom’s Business Model

   3. Axiom’s Revenue Numbers are Misleading

2. Just Win: Inverting Consensus Advice

3. Third Derivative Products

4. AI Compresses the Insight-to-Execution Pipeline

Following this analysis, I provide commentary on Axiom and the shape of this company archetype moving forward.

1. Background on Axiom

The Team Behind Axiom

Axiom was started by two co-founders, Henry Zhang (ex-TikTok) and Preston Ellis (ex-DoorDash). Axiom has only raised capital from YC, and has not sold any more of the company.

Unlike other crypto teams who have historically sought to de-risk themselves by raising large seed and Series A rounds before launch, Axiom embodies the YC ethos of shipping, iterating, and talking to customers. Axiom ships and communicates frequent updates to their users via Discord and Twitter.

Axiom has since grown its engineering team in Austin to scale out its product. The majority of the core infrastructure has been built by Henry and Preston.

Axiom’s Business Model

Axiom makes money by charging trading fees and selling order flow. [1] Like Robinhood and other traditional brokerages, Axiom is a frontend and does not own any of the underlying infrastructure.

Axiom currently processes $50M - $75M of daily volume. Trading fees are charged based on volume. Axiom charges 1% on every trade at execution. Depending on your account’s volume, Axiom pays users back via a rebate, lowering the net effective fee. Their public cashback fee tiers range from 0.05% to 0.25%. This means that a user’s net fee ranges from 0.75% to 0.95%.

For context, users trading US equities typically pay 0.05% (5 bps) or less per trade. Axiom is able to charge an order of magnitude more because memecoin traders are extremely price inelastic. Axiom users operate in an asset class with an extremely convex payout profile (users either expect assets bought on Axiom to increase in value by 10x or go to 0 within hours, sometimes minutes). Because of this dynamic, users are willing to pay high fees for a best-in-class user interface and trade execution.

By default, trades on Axiom are sent with 20% slippage (the majority of users do not change this setting). This means that the user is willing to accept a 20% worse price if it is able to be executed. If an asset has a notional value of $100, users express a willingness to accept $80 as long as they can buy or sell the asset. While slippage theoretically works both ways, users almost never receive positive slippage (receiving more than they initially anticipated). Positive slippage is typically captured and retained by a sophisticated party at some point in the supply chain.

Like Robinhood, Axiom also makes money from selling user order flow (PFOF). Axiom’s order flow is extremely valuable because it is non-toxic (originates from normal retail users) and high volume. The majority of the order flow is being sold to Temporal, a Solana-native research and trading firm founded by individuals with backgrounds from Citadel and other traditional trading firms. Axiom routes order flow to Temporal, which is responsible for actually executing the order flow. Temporal abstracts the execution complexity of actually landing the transaction on-chain for developers. Running infrastructure to execute order flow is a highly competitive business in crypto with high technical, operational, and economic barriers to entry (large amounts of tokens are required to run validators).

The entities that Axiom sells order flow to are responsible for executing the order flow in the same way that Citadel is responsible for executing the order flow on Robinhood. PFOF, implemented correctly, increases user welfare, as users receive better prices. However, in the unregulated crypto arena, the little to no oversight by any regulatory committee has translated into recurring abuse of users. While Robinhood users enjoy price improvements from order flow segmentation, Axiom users often get maximally exploited by actors throughout the supply chain.

Axiom’s revenue numbers are misleading

Axiom’s business model of charging 1% upfront and variable rebates leads to misleading numbers. Charging someone $100 but immediately giving back $50 should not be marked as $100 of revenue.

Axiom’s documentation site outlines the variable cashback tiers based on trading volume. However, this is misleading as select large and early users have non-public side-deals with Axiom that increase their cashback rate.

Many of their early users are receiving 50%+ cashback for an effective fee of much less than 0.50%. Because trading terminals like Axiom are so top-heavy (with a few users generating much of the volume and fees), Axiom’s top-line $100M revenue metrics are an overestimate.

Axiom’s public revenue numbers do not include additional revenue from selling order flow to Temporal. While the details of that arrangement are non-public, this is likely in the high seven-figure range depending on volume.

Because Axiom’s business has many obfuscated components, it is difficult to accurately assess their top-line revenue from an outside perspective.

2. Just Win: Inverting Consensus Advice

The consensus VC view is that one should only invest in companies with structural moats, or at least in companies with a clear pathway to creating one. They believe that moats are a necessary component of great companies because they allow companies to operate as a monopoly, enabling them to extract rent via their differentiated products and services. Companies that operate in competitive markets have no structural moat and are viewed as uninvestable and low status.

Ceteris paribus, one would much rather operate a company with a structural monopoly on some product or service. However, most people do not have the technical capabilities or social capital to have a real monopoly. If a company is telling you that it operates a monopoly, they almost certainly do not.

Barring extreme outliers, the sharpest college graduates are incapable of creating a durable monopoly business. Building a monopoly business requires high context, a specific thesis, and political intelligence that requires many years of networking and life experience. This has led to many top graduates chasing empty prestigious titles and careers. No one is really working.

In this world of status-chasing and people losing the plot, actually talking to end-users has become a moat in itself. While anybody can do it, it requires interacting with people many see as below their social status. This has led to many influxes of people only wanting to build companies where they are only selling to people they can bear to interact with (B2B, typically selling to engineers at other companies), if they even talk to customers at all.

Axiom inverted this advice and consciously entered an established market, outcompeting the existing players:

• Built a company in crypto with strong engineering and product backgrounds, where the talent bar is low due to crypto’s deserved reputation for scams.

• Entered the memecoin space, a sub-sector of crypto with some of the lowest talent and integrity levels across tech. Many competitors did not have traditional Silicon Valley experience and lacked the institutional knowledge of how to build applications. Furthermore, many early memecoin developers got rich from the previous wave of memecoins and checked out, leading to significant adverse selection.

• Built a frontend, a thin layer around existing infrastructure and protocols with effectively zero barriers to entry and pulls from a global talent pool.

• Entered the game late (post-Trump election in 2025) when there was sufficient market demand and existing products were already generating millions of revenue (Photon, BullX, BONKbot).

Axiom brought Silicon Valley product experience to a set of users willing to pay high fees. Much of their initial growth can be attributed to just doing the basics: focusing on a niche, talking to users, and rapid product iteration. They penetrated the few high-volume Telegram trading groups that matter, scaling via word-of-mouth and a referral program with Axiom-watermarked graphics.

3. Third Derivative Products

In a previous post, I outlined a theory of “derivative levels” as a way to understand why certain roles and types of work are highly compensated.

In the same way that third derivative workers identify the problems to work on and allocate human capital towards solving those problems, third derivative products enable users to identify where to build new tools and decide what to work on.

When you visit TikTok or YouTube, you are submitting a bid to the global video marketplace. Every time you open one of these apps, you are asking, “What should I watch right now?” The platforms fulfill their side of the trade by serving the content based on what they think you will enjoy.

When you visit Axiom, you are asking the platform, “What tokens should I trade right now?” The context is already set as memecoin traders, so even though the front page says “The Gateway to DeFi”, users have context that they are there to trade high-volatility memecoins. Axiom fulfills their side of the trade by surfacing the user new tokens and allowing them to buy and trade any memecoin. Axiom users are also presented with Twitter overlays and a wallet analysis tool, which Axiom actively curates to display to the user the information they want them to see. The platform offers a unified interface that enables others to identify which tokens to trade and how to reason about the probability distribution of them making money off this token (third derivative). This is packaged in a system that also provides the means to trade the token in the same interface (second derivative).

Third derivative products are high leverage, and overwhelmingly represent the highest revenue per employee companies. Axiom, Hyperliquid, Claude, ChatGPT, Twitter, and other third derivative products enable very small numbers of people to have extreme outsized impact due to zero-marginal-cost software. Because so few companies are able to execute the product in a way that resonates culturally, they build immediate brand trust and create a self-perpetuating network effect. Many products also are able to recursively improve via proprietary model training based on user data. Once brand trust is established, these platforms become upstream of culture as people consciously and subconsciously outsource the creation of their own thoughts to these platforms.

Third derivative products tell you what to make, what to consume, and what to trade while providing unique tools and infrastructure that provide those services. They operate on the meta-level of attention and intention, providing tangible answers across domains to the question: “What should I do?”

Building a successful third derivative product is incredibly difficult. You have to provide value to the user to guide them to what they should do next in contexts you can’t anticipate but have to generalize. These tools must not only be technically sound, but more importantly be epistemically correct in the context of the current culture.

4. AI Compresses the Insight-to-Execution Pipeline

While vibecoding has become commonplace among the most productive programmers today, many did not utilize AI tools to this extent in the early parts of 2025 when Axiom was initially being built. The earlier models were not as good for code generation, requiring much more engineering context to steer in an effective manner.

The founders of Axiom are extremely sharp and hardworking senior-level engineers. Their experience meant that they knew what the shape of the code should look like and where they could cut corners, enabling them to build a feature-rich application with extreme product velocity. Combined with their differentiated product insights derived through experience and tacit knowledge acquired via Silicon Valley firms and the low talent level of competitors, this meant that they were able to outcompete competitors across the board.

The combination of prior experience at traditional product-led companies (TikTok and DoorDash) and novel tooling enabled them to express their differentiated insights without lossy interpretation through other employees. Previously, founders were not able to express these insights at scale, but did not have the bandwidth to be close enough to the ground to manage all small details. This enables much better customer understanding and creates much tighter feedback loops.

Knowing What to Build and What to Optimize

Axiom is able to achieve extreme product velocity because they operate as a simple web app wrapper. Axiom avoided building complex or high-security infrastructure that would have required months of code review and auditing. Instead, they utilized third-party providers such as Turnkey, which manages user logins and wallets. The transaction landing infrastructure is handled by Temporal and actual asset transfer is handled by Solana. Axiom’s edge is orchestrating existing infrastructure in a way that is engaging for end users. They knew that their moat was applying differentiated product insights to an underserved set of users.

Unlike many other teams who either naively use third-party products off the shelf or try to build everything in-house, Axiom made targeted optimizations that actually improved the end user experience. While other Turnkey applications naively use their SDKs, Axiom performs browser-specific private key optimizations. This enables them to cut latency by saving one round-trip to Turnkey’s hosted infrastructure and to send transactions directly from the user more efficiently.

Dark Patterns

Axiom’s product insights range from better onboarding flows for referrals to implementing select dark patterns. Dark patterns are user interface designs that are designed to trick or manipulate users to do things they didn’t mean to do.

To be sure, many companies employ dark patterns and aren’t strictly deceptive or bad. Common dark patterns include making it difficult to cancel subscriptions and disguised ads. The deceptiveness of these patterns exists on a spectrum from being fairly innocuous and well-accepted to straight up lying to their users.

Some of Axiom’s dark patterns include:

• The cashback rebate system

• The lack of transparency around order flow management

• How they present reclaiming state rent back to the user [2]

• Affiliates and sponsorships

The inventiveness of the dark patterns implemented throughout the app is a subjective cultural judgment and is left as an exercise to the reader.

The Next Iteration of Elite, High-Growth Startups

Regardless of your opinion about crypto and the durability of Axiom as a business, Axiom is a clear harbinger of how AI is reshaping elite, high-growth startups.

I have a separate post in the pipeline about how AI changes workplace dynamics, founder incentives, and salary justifications. Across the board, Axiom illustrates many changes on the horizon:

• AI increases the value of judgment, derived via secrets and differentiated outside knowledge, as speed becomes largely commoditized.

• AI enables lower headcount and more leverage per employee, increasing the blast radius of decisions. In turn, companies will pay more in equilibrium as an insurance premium and to decrease the likelihood of spinouts.

  • Axiom had some of the highest salaries for software engineers, advertising salaries well over $500k per year and covering all expenses.

• AI increases the visibility for the company to accurately assess worker productivity, increasing turnover.

• AI increases the value of social and political intelligence and intuition.

The success of Axiom has naturally legitimized trading terminals as a business, leading to much higher-caliber talent entering the space and building competitive products. The enterprise value of Axiom on February 25, 2026 likely hovered around $1-3B. As of February 26, 2026, due to the ZachXBT exposé and a multitude of other factors, their enterprise value is certainly much less and likely trending downward.

Young, immature founders lack the social intelligence and intuition to navigate the political minefield. It is clear that the 2010s and 2020s uniquely favored the one-dimensional technical founder, and that era is coming to an end. Over and over, we see this founder archetype fail to implement standard security or accounting measures and make clear hiring errors. Often, this stems from being too removed from understanding the human condition and becoming far too trusting of others. While they have been able to compensate via sheer technical prowess, they have not been able to create durable and cohesive companies that stand the test of time.

There is no single predetermined path to building a great business. Successful third derivative companies are stewarded by people with differentiated personalities and life philosophies (Sam, Dario, Elon, Jeff, etc.). Navigating the hyper-online, main-character meta is now a requirement, and many people are still not able to assess or develop the traits required to transform an early-stage elite, high-growth startup into something greater.

[1] Axiom also makes money off of staking income. Their fee profit is denominated in SOL, which they then stake on Solana to earn staking income. This is public, as they operate their own validator. This is a common setup for application teams on Solana.

[2] When a user interacts with Solana, they have to pay rent to store data on-chain. When a user closes an account (e.g. an account associated with a specific memecoin or token), they receive their money back.

Popularity prediction hash: 1bd610aa51cc3a510435ceb483b613032309067632b433611ae4b34ab6c5c213

November was actually my most prolific month of blogging. Before November, I published 13 posts in total. I published 30 blog posts in November alone, over double what I had achieved over the course of the previous nine months.

Logistics

All posts are live and available on humaninvariant.com. You may have seen some posts already on MR and HN.

Across the next month, I will be posting select posts from the past 30 days to the Substack mirror and the mailing list. Others will be published to the Substack mirror without a post notification.

Retrospective

I published 28,427 words by posting every day over the course of November. Some topics I covered include:

• Third derivative work (building off of No One is Really Working)

• The negative externalities of agency-maxxing “you can just do things” posts

• Blame as a Service

• Prediction markets

• the lowercase aesthetic

• Why are conversations and my productivity better late at night?

• The market microstructure of YouTube

• Observations from in person interactions with successful bloggers

• Understanding the scale of the internet

To be sure, not all posts I wrote during this stretch are good. Some days I really did not want to write, and likely would have quit without external pressure. But I learned just how good of a forcing function publishing a post every day is. It enabled me to explore more ideas than a typical month of one or two posts, and now I have many more ideas I’m excited to explore.

Many of these posts were specifically enabled by being in person and talking with other people I had never met before. I spent the month with people who have been thinking about prediction markets for years, including someone who ran internal prediction markets at Google in 2017 and an academic who was an early author of many prediction market papers.

I also learned what some of my favorite bloggers are like in real life. Many are true to their online selves and are incredibly generous with their time. I am extremely grateful for each and every one of them.

November was also my most productive month of reading. I don’t have stats on this, but I read more books and blog posts than I ever have, by an order of magnitude. I discovered new blogs and posts from old blogs I hadn’t yet read, enabling me to cross-pollinate ideas and connect previously invisible dots.

Updates

I’m extremely happy and grateful for the audience I’ve attracted. I’ve received countless instances of thoughtful feedback and emails from people, including many people I did not know, with zero negative externalities thus far. Readers are not fungible, and the audience this blog has attracted thus far consists of the exact type of people I’m targeting.

To preserve what’s working and advance the blog into its next phase, I’m implementing the following:

• Exploring a single topic across multiple posts

• Spending more time offline to derive truly novel insights

• Increasing idea velocity by getting more feedback

• Turning off subscriber notification emails and blocking access to my analytics dashboard

Many of my posts up to now have been quick takes, especially during the past month of posting every day. While I want the blog to have diversity in topics and post formats, I want to write more well-formulated, research-driven posts that generate novel insights and push the idea frontier forward.

I won’t be posting every day going forward, but I’m not going anywhere.

Let’s get it on.

Popularity prediction hash: 3fd2777698fba1a7753b485cac734b85b20fac80d81e21da084770f199ef5a1f

The illegibility prevents you from seeing the countless intermediary drafts, weird quirks, and rituals that enabled the final work product to come to fruition. The illegible inputs are an amalgamation of a rich contextual history that is only learned offline, in person.

Physical presence exposes the illegible because it gives you access to people, and people are the generators of culture, systems, truths, and opportunities. The real updates to your life don’t come from reading words online, but rather from being exposed to stimuli in the real world.

Culture

When you show up in person, you feel like you don’t belong. You quickly learn that others have a deep and rich shared cultural history that has spanned over the better part of a decade, while you are a newcomer. You try your best: you connect with people during late night conversations you would have never connected with otherwise, and maybe even plan to throw a joint party together in New York.

You learn via oral history that you are standing in a place that was going to be turned into a soulless Marriott for consultants, but somebody cared enough to prevent that from happening. You learn that even the most rational people operate under the same universal constraints as you.

People

You learn about the entire production function of others and what makes them tick: from the conversation topics that pique their interest even well past midnight to how they take their coffee. You notice that one person always sits at the same outdoor table facing north, how impervious some are to the cold, and how many Diet Cokes people are drinking well past 9 pm. You learn that one of the leading social psychologists has a sick Ness. You learn about the Cosmic Crisp industrial complex, the only good thing to ever come out of Washington State.

Systems

You learn that these same dynamics are present in blog posts, which turn into movements, companies, and political change. You learn the shape of successful people: some of them are nicer and some are meaner than you could have imagined. Some completely strip away the Straussian veneer while others continue to talk in coded messages with multidimensional meanings.

You learn that rational explanations of subcultures are overdetermined. People, not algorithms, are at the frontier of ideas. And that people respond directly to incentives, which are omnipresent and predictably determine downstream behavior.

Truths

You learn that adverse selection exists and that trust doesn’t scale. You learn that you can meet everyone that matters across generations. The world is more binary than what you imagined.

You realize that you can surround yourself with people who are actively conspiring to help you succeed. That endless opportunities are available for you. And that you can do the same for others.

You learn that you can’t reason about everything from first principles. But you have specific ideas about the world that are differentiated and worth publicizing. You learn to generate your own worldview of idea-driven and data-driven ideas. You realize that your breakthrough idea from a week ago is actually incorrect and deeply flawed. You know that being wrong is the nature of operating at the frontier of ideas, and part of life.

Becoming

You learn more about the shape of the life you want, the types of relationships you want with certain people, and the sacrifices you are willing to make to be ambitious. You learn that ambient ambition is real, and what it really means to love what you do, when you witness someone over twice your age work into the night and enjoy every moment.

You learn that your true friends are the ones willing to cross the Bay Bridge and pay the toll to visit you on a random Saturday. Having the courage to trust and share intimate moments with your close friends is worth it. That the only life worth living is the one you can share.

Caring

You learn what it feels like to use a typewriter because somebody else cared enough to give you the opportunity to experience that feeling and made it happen. You use that opportunity to ponder the market structure of writing implements and to write an endearing note to the people who have given you the opportunity to learn more about yourself. While doing so, you learn what the sound of a bell means when typing on a typewriter.

You learn that having the courage to care is the scarcest resource in the world. You can predict the future you want by caring enough to build it.

Most importantly, you learn that the ability and desire to care is built through presence.

Popularity prediction hash: 0d49acef8f704d3f4fd61376a1d5cf4def810be952733c1320f62e0630ecde4f

In this post, I attempt to provide a better intuition for contextualizing audience sizes across orders of magnitude, using visualizations and anecdotes.

Following the contextualizations, I provide reasoning for why our inability to fully grasp audience scale is precisely what makes it possible to create authentic work on the internet.

Click here for the interactive slider

1 - 10

1 handshake per second = 10 seconds

Physical corollary: A dinner table

While it may feel like you are posting into the void, it is akin to giving a speech to a table of ten people in meatspace. You’re not “posting online”, you’re giving a thoughtful toast to a small table of people who know your name. Every sentence feels intentional because every person is close enough to see the look on your face as you say it.

10 - 100

1 handshake per second = 1 minute 40 seconds

Physical corollary: A large classroom

You’re giving your 15 minutes of fame presentation to your peers in a high-school-sized room with one to three dozen desks, and people standing up in the back as overflow. Some people may be distracted on their phones, but everyone is present and listening to your words.

100 - 1,000

1 handshake per second = 16 minute 40 seconds

Physical corollary: A large university lecture hall

You’re giving a speech in a tiered lecture theater with the lights shining in your eyes. You can technically make out some faces in the front row, but you can’t track reactions in real time. This is the first scale at which stage fright becomes a real physiological event, as your body recognizes the stakes even if you try to mentally downplay them.

1,000 - 10,000

1 handshake per second = 2 hours 46 minutes 40 seconds

Physical corollary: A sold-out medium-sized concert venue

This is likely the upper limit of humans you’ve ever had in your field of view at once. You aren’t able to see people individually anymore, but rather as a larger organism. If you bombed in front of this many people, you wouldn’t sleep normally for days. If you crushed it, you’d feel invincible.

10,000 - 100,000

1 handshake per second = 1 day 3 hours 46 minutes 40 seconds

Physical corollary: A football stadium

At this scale, the audience stops behaving like humans and becomes terrain. This is the domain of athletes, heads of state, and A-list musicians. The crowd seems like texture, as faces are impossible to discern. Most people live their entire lives without ever being acknowledged by this many humans at once.

100,000 - 1,000,000

1 handshake per second = 11 days 13 hours 46 minutes 40 seconds

Physical corollary: A metropolitan city

You’ve exceeded the capacity of any single synchronous physical space. The entire population of San Francisco stopped what they were doing and paid full attention to your creation: every barista, every commuter, every tourist on a cable car.

1,000,000 - 10,000,000

1 handshake per second = 115 days 17 hours 46 minutes 40 seconds

Physical corollary: A small country

This is approximately the population of Hong Kong and Singapore combined. This is what it means for something to go truly viral on the internet. This is more people than you could meet across a lifetime of dedicated socializing, compressed into a single notification.

I once said that the early days are the best. It’s still early days for this blog, but it is now less early. Some days, as I’m responding to thoughtful emails, I stare at my view count and attempt to contextualize this audience. Even once you think you intellectually understand them, it still takes constant reminders that these numbers are real. Most days, the view count is just a number, and I am able to freely write as if it is just me and the page.

Other days, I feel the entire weight of this audience staring over my shoulder. The number that I typically interpret as bits on my screen suddenly represents a collection of real people, each encountering these words in the context of their own lives. What is just a number suddenly becomes legible, and I feel the weight of all your hands I have yet to shake.

Popularity prediction hash: fb99551bce2217061756bfa2e9003d921be03c44efd61d912ea127af8aca364d

• Idea-driven ideas: an idea where you start off with a high-level philosophical frame and deduce a concrete insight using logical reasoning

• Data-driven ideas: an idea derived from pure data with no preconceptions

Building on this framework, industries can be characterized as idea-driven or data-driven.

Idea-driven industries:

Data-driven industries:

Similarly, movements can be categorized as primarily idea-driven or data-driven.

Idea-driven movements:

Data-driven movements:

Enabled by technology, data-driven ideas operate at higher velocities than their idea-driven counterparts across industries and movements. Data-driven ideas are not more rigorous in nature, as evidenced by the many industries and movements built on falsified data.

In this post, I outline how the legibility of data-driven ideas creates coordination, which in turn, creates velocity.

Data as a Schelling Point

The post-Moneyball era demonstrated that everybody must think in data or risk obsolescence. The current era of statistical models and graphs aggregating thousands to billions of data points is a very recent phenomenon that is only now gaining widespread cultural acceptance.

The key feature of data is that it is focal, not that it is necessarily better. This leads to people gaining a sense of transcendence to data for its promise of unbiased representation, even if data is often misleading at best and fabricated at worst.

Charts and statistics compress an arbitrary number of data points into something everyone can see and cite. Often, these charts and statistics transcend language barriers, requiring only very basic knowledge of English and often zero additional context.

Legibility enables quick ascents

Data-driven ideas convert attention into conviction faster than idea-driven ideas. This enables individuals, industries, and movements to experience extremely large growth rates in a short period of time:

• Individuals: Aella is a scientist and sex worker who has leveraged her following to gather some of the most comprehensive datasets on relationships and sexuality in human history. One of her surveys, which takes over one hour to complete, has over 850,000 responses. Every single post by Aella is backed by evidence, typically a graph derived by her own data from one of her surveys. Sex and relationship discourse is stale and vibes-based, while Aella’s data provides legibility.

• Industries: Quant funds out-fundraised discretionary managers not purely on returns, but on the ability to cite legible backtested data over vibes-based hedge fund managers. Growth engineers beat product visionaries in internal debates by A/B testing every aspect of the product experience.

• Movements: EA scaled within years based on the premise that impact can be numerically measured. Lives and charities can be directly measured in impact per dollar, enabling the comparison of interventions on a spreadsheet and the ability to optimize your altruism.

Idea-driven ideas have to continually persuade, while data-driven ideas just have to point.

Falsifiable claims create fragility

The same property that enables quick ascent also enables quick collapse:

• Individuals: Public intellectuals who stake their reputation on data-driven claims rise and fall with that claim. The prediction market pundit who calls an election correctly becomes a genius; the next miss makes them a cautionary tale. The half-life of a data-driven pundit is much shorter because their credibility is indexed to something falsifiable.

• Industries: The replication crisis put the entire field of social psychology into question overnight. A quant fund faces the same dynamic: the backtest that attracted capital becomes a liability the moment the model breaks.

• Movements: Climate activism coordinated globally around a specific number: 1.5 degrees. Websites now display countdown clocks to 1.5 degrees, currently at 3 years and 235 days. This creates a failure mode if there are no long-term externalities after reaching the number. What happens if we pass 1.5 degrees and the predicted externalities don’t materialize on the expected timeline?

Idea-driven ideas are low-beta and don’t face the same fragility because they are tied to unfalsifiable ideas.

Open Questions

• How do the aesthetics of data-driven ideas relate to their popularity?

• The half-life of many public intellectuals is quite short. How much can be attributed to staking their reputation on a specific data-driven idea and going down its respective rabbit hole?

• How correlated are idea-driven / data-driven ideas and missionary / mercenary, respectively?

• Where do LLM-derived ideas fit?

• What’s the optimal ratio of data-driven to idea-driven claims for a movement that wants both velocity and durability?

Popularity prediction hash: 9fa5947ff5a5745a1291997d6f7d81edcf0913ff6f2239ac79fc43d992b08237

While I think that the current equilibrium of the diffusion of bloggers creating a set of active practitioners is a better allocation of human capital, I recognize that this is hypocritical as I am a direct beneficiary of the broader blogosphere. I did not enjoy university and was not on a path toward social and intellectual fulfillment. At the right moment of my life, I discovered the broader blogosphere network, which enabled me to engage with ideas in a way that school never clicked for me at school. Most importantly, it provided me with a network of earnest peers and mentors that provided the substrate for confidence and courage, and growing my ambition at a critical point in my life.

Much of this can be attributed to the fact that the blogging community is made up of an incredibly small number of extremely generous and earnest individuals. Capital and time are viewed as sharable resources to invest and bestow upon the next generation. While most institutions and other aspects of life are typically very antagonistic and transactional, the blogging community provided one of the first glimpses into a culture that can make a legitimate claim to be earnest and positive-sum.

The handful of people provided mentorship and capital that definitively changed the course of my career. Many of my friends and other bloggers have had similar experiences. All of us are eternally grateful to our friends and benefactors who enabled us to chart a different path in life beyond the default life trajectory. These people, some of whom I met because I read their blogs growing up, acted as our omniscient entity. They showed us firsthand what it takes to be great at something, raised our aspirations, and made the right connection at the right moment in our lives. More blogs would enable more people to have similar experiences.

In this post, I provide multiple tested and untested mechanisms to incentivize new blogs and help existing blogs scale.

Incentivizing New Blogs: 0 → 1

The barrier to entry for bloggers is effectively zero. Infrastructure is free via platforms such as Substack. Most of the time, bloggers are stuck in the valley of believing they have nothing interesting to say or don’t think the opportunity cost is worth it.

The following are steps to incentivize the creation of new blogs:

1. A richer culture of encouragement around starting new blogs to provide an easier on-ramp.

More new blogger Slacks, core meetup czars in different cities, etc.

2. Proactive funding to get blogs off the ground.

While starting a blog might be free, opportunity cost exists and proactive funding helps alleviate these concerns. More importantly, capital provides a subtle, in-group signal that their work is meaningful and worthwhile to be pursued.

3. Retroactive funding to provide the incentive that tangible blogs will be rewarded.

Not every blog will be legible enough to receive proactive funding. Retroactive funding enables someone to start their own blog, prove themself, and receive retroactive funding after accruing a rich body of work.

Scaling Existing Blogs: 1 → n

While one of my favorite activities is reading the early blogs of people before they became successful, one of the saddest activities is seeing a blogger I enjoyed stop posting. Good blogs don’t fail because the blogger runs out of ideas; they fail because the blogger is isolated, under-resourced, or determines that the opportunity cost of writing a blog is too high. Practically, most bloggers find themselves with demonstrated traction while being too small to justify a serious career out of blogging.

The following are ideas to help scale existing blogs:

1. Better tooling

The best bloggers write to say stuff worth knowing, not just for the sake of writing. They have countless great ideas but lack the time and resources to craft the idea into a legible blog post. Furthermore, many are extremely neurotic and only publish blog posts when it is perfect.

LLMs and other general writing tools decrease the time from idea to post, therefore decreasing opportunity cost. Existing blogs with a corpus of existing posts and other internal documents can better leverage LLMs to assist in each stage in the writing process as needed. Most of the workflows that exist today are primarily coding and general purpose-oriented, not writing oriented based on personal training.

2. Grant matching

For every dollar a writer receives in reader funding during a given period, a sponsoring organization matches it up to a cap.

3. PIPE

A writer with existing income (Substack, ads, sponsorships) can sell a percentage of the next two years of earnings for upfront capital. This is analogous to a PIPE round for individuals rather than companies, which is useful for funding projects with upfront capital requirements (e.g. research projects, travel, etc.)

4. Implicit rights of the blogger’s future endeavors.

For a specific set of bloggers, the real asset is the associated network and deal flow of the blogger. The blogger can guarantee that creations for the next n years have x% reserved for early supporters.

5. Invest in bloggers via lump-sum capital.

Bloggers receive capital into an escrow, which can be pulled out after achieving certain milestones. In the success case, the blogger is able to generate an income stream that outpaces the amount withdrawn, so the capital can be returned back to their owners. In the failure case, backers lose their initial investment.

6. Add the ability to buy future gated content at a discount.

Popularity prediction hash: f9060f0f351d67d38b3de52025bb17a8b6a6f9f8d5c39802f82cbc24961b5837

The blogosphere attracted a certain archetype that was overwhelmingly interested in ideas and willing to argue about a variety of topics with random strangers on the internet. Because of the blogosphere’s centrality and small size, a formal status kingdom among internet intellectuals naturally emerged and a network effect around the blogosphere developed.

Scott Sumner described the distinction between “blogosphere economics” and “academic economics” as:

To me, the term ‘blogosphere’ sounds dumber than ‘academic,’ even if you didn’t know the meaning of either term. Say them both out loud. I’d prefer “the ungated, free entry, merit-driven, competitive, methodologically eclectic, wisdom of crowds-based global hive mind,” vs. “ivory tower, methodologically dogmatic, elitest, academic world where you either data mine to publish empirical papers or invent endless permutations of theoretical models.

The blogosphere was one of the primary intellectual counter-cultural movements that legitimized internet-based discussion and science. There is still a long way to go but there are clear signals that the tides are turning:

• SlimeMoldTimeMold are mad scientists trying to solve obesity via a terrific series on how to run one-person studies.

• Aella is a scientist and sex worker running one of the most comprehensive datasets in human history, contributing to our understanding of relationships and sex with novel data-driven insights.

• Adam Mastroianni, a Harvard social psychology PhD, choosing to not work in academia, electing to run a Science House for researchers to live together and work on science, while also writing a Substack.

One by one, elite talent is being vampire attacked into the ungated, free entry, merit-driven, competitive, methodologically eclectic, wisdom of crowds-based global hive mind. But while internet science begins to gain legitimacy and solidify, the blogosphere that spawned them has been on the decline.

The blogosphere decline

Since its peak in the early 2010s, the blogosphere has largely faded away due to some combination of the following factors:

• Social media: Twitter offered higher engagement, more career upside, and more universal status markers for a fraction of the effort. Why spend three days crafting the perfect 2,000-word blog post when you can pump out 280 characters every time you take a shit?

• Stagnation in blogging market microstructure: besides Substack, which started in 2019 and provides infrastructure to deliver posts directly to inboxes along with a monetization layer, nothing else has changed. Existing blogosphere participants were already plugged into RSS feeds, meaning that there was little expansion of the scene.

• Higher opportunity costs: The 2010s were a decade-long bull market, where people rationally chose to become a practitioner over a Straussian internet blogger.

• “Real jobs”

• Kids

Regardless of the causes, like many multidimensional emergent scenes, even a sense of stagnation can lead to cultural decay and exodus.

Where is the present day blogosphere?

Every few years, a blogging renaissance seems imminent, only for core contributors to defect to do something else.

The would-be bloggers of 2025 have far more outlets to flex the same intellectual muscles their predecessors once expressed through blogging. Capital is abundant, barriers to other mediums have collapsed, and the pathways for earning internet social capital are more egalitarian and multidimensional than ever.

Becoming legible to existing players, once largely achieved through blogging, can now be accomplished in far more lucrative and impactful ways:

• Working at a research lab and writing deep dives in company Slacks

• Building alternative institutions for science

• Running anon accounts on Twitter

• Running YouTube channels

• Starting a hedge fund

• Starting a podcast

In many ways, the blogosphere resembled the high school experience of hyper-optimized children: a concentrated, artificial, out of touch environment where status competition was the main event. The dispersal into other professions with positive externalities should be interpreted as graduation, not retirement.

This isn’t to say that a rich blogging scene isn’t useful or will never be seen again. Blogging residencies are connecting successful bloggers with newcomers. The beauty of the blogosphere is that it is emergent, egalitarian, and not capital-intensive: you only need a handful of core individuals and one hit post to spark an entire movement.

Maybe one day the next blogosphere will make it out of the new blogger Slack.

Popularity prediction hash: aa2ad20ce00d0a395586901bf2f784e96a115c3d89c114b4968c4f08de0b8d0e

• Julian Sanchez’s The Straussian Case Against Gay Marriage

• Richard Hanania on why attractive women are still high status

• Any Bryan Caplan book, blog post, or podcast

Oftentimes, I learn that the author has a largely anti-Straussian reputation.

Rarely, I come across an anti-Straussian moment from an individual with a Straussian reputation:

• Chamath speaking to Stanford GSB students, stating that the world is run by only a handful of men, and that it is your moral imperative to make money to impart your worldview on the world.

• Thiel’s The Diversity Myth: Multiculturalism and Political Intolerance on Campus, which he later softened and reframed.

• Sam Altman on his blog and his advice to YC founders that contradict the effective mechanics of building durable companies.

In this post, I explore the reasons why there is little insightful anti-Straussian content and the forces that shift people toward Straussianism.

Why is anti-Straussian content so rare?

The following are possible explanations, in no particular order.

1. You can have more impact by being Straussian and applying insights to real world activities.

Those able to generate insights rationally opt to apply them to the real world to capture the upside instead of writing about it.

• A consultant who uses their knowledge of interpersonal dynamics to increase efficiency across a wide array of companies

• A VC who uses their insights into talent formation to invest in the next outlier founder.

• A founder who starts a company based on some key insight.

2. There aren’t many people who can generate these insights.

Most people aren’t interested in generating insights about the world. Most impactful insights require a rare combination of life experience, hard skills, and curiosity.

3. The number of people willing to say truly unpopular things is near zero.

While this may seem like an easy “arbitrage”, going against the social order is one of the most difficult things to do, especially because the costs are often permanent.

The Straussian Equilibrium

While technology has enabled faster feedback loops and makes it easier for the long tail to publish Straussian ideas, the social dynamics that underpin human behavior have not changed.

People typically first cultivate their reputation by generating anti-Straussian content. This operates much like a credential, and serves as an initial filtering mechanism. Today, this manifests in relationships that begin online and transition offline. This enables people to accrue offline social capital and generate high eigenvector centrality in certain networks.

These high-variance individuals are quickly subsumed by institutions that instill Straussian behavior (elite media, VC firms, academia). For the institution, this serves as a selection mechanism to determine who is most suited to continue its legacy. For the individual, this grants them a ticket into the in-group and, more importantly, an everlasting omniscient entity.

Additionally, Straussian content creates a call option on any future interpretation. Done correctly, social capital is bestowed on the creator, as they are viewed as extremely interesting and multidimensional to outsiders. There is nothing more bullish than a person being able to regularly generate downstream cultural impact based on their Straussian content.

Options

1. Become a practitioner

2. Go anon, a la the Founding Fathers

3. Be anti-Straussian

Popularity prediction hash: ef5cd6acf584f6f166705c55b8e909f841506cca723c6012a43b39e7923da9f8

Over the past couple of weeks, I’ve had the opportunity to meet numerous internet-famous bloggers in person. Approximately half of my discussions happened in small groups, while the other half occurred in a one-on-one setting. My conversations with these bloggers have ranged from how they generate ideas to the personality characteristics of their kids. If you’re reading this post, you’re the type of person who has read many of their posts as well.

While I think I can predict the success of a given post, I don’t think I can predict which bloggers will be successful. [1] Some of this can be attributed to the fact that a large component of blogger success is future post frequency, which is an unknown metric. That said, meeting them in person has left me with the beginnings of a pattern I can’t yet articulate.

The following are some of my assorted notes from my in-person interactions with successful bloggers.

1. They do at least one thing that is weird and makes you double-take.

Some anecdotes:

• There’s a well-known blogger who insists on only browsing the internet on incognito mode while on a VPN and not using the trackpad. It’s not that the trackpad is broken, but they use some combination of the keyboard to move their cursor around.

• One blogger uses a single monitor in vertical orientation.

• Another blogger has a specific three-monitor setup: one horizontal monitor with one vertical monitor on each side.

2. They work extremely hard and love it.

I have been working physically adjacent and in the same room as a successful blogger. Every single day without fail, he leaves after I do at night and gets there before I do in the morning. And I’m not leaving that early. I leave at around 10:30 or 11 pm, and get in at 9 or 9:30 am. More importantly, it looks like he’s enjoying every single moment of being there.

I’m not very good at estimating ages, but if I had to guess, he’s somewhere around 50. I’m younger than 50, and I don’t have this level of stamina.

3. Most or all claim to write for themselves.

If I were to estimate, about 40% are saying this earnestly, 40% are saying this because the other 40% are saying it, and 20% are not telling the full truth.

4. They are overwhelmingly low-ego and interested in ideas.

5. They all have smarter friends than them and often wonder why they are popular bloggers.

6. They read every day. Then they read more.

7. They are better dressed on average than what I expected, and not in an eccentric sense.

8. When talking with them, you can hear the blog come out of them in real time.

9. In aggregate, they skew towards being introverted.

10. Many of them use Androids.

This could be a function of generational differences as successful bloggers are older, on average.

11. They are irrational in one weird way, at least from my perspective.

12. They all lead very interesting lifestyles or are ~interesting people.

13. They are spiritually children in at least one dimension.

14. For some, their blogs give off the impression that the blog is the best instantiation of their life’s work, for others it feels like a footnote.

15. They are comfortable with being wrong on the internet.

While some started their blogging career as being deeply uncomfortable with being wrong, many got over it by consistently publishing.

Many had a significant incident or feud that served as a canonical moment for them, demonstrating that there’s really nothing to worry about.

16. Many are able to contextually place their comment section in a segregated frame of mind.

With one notable exception.

17. They’ve met everyone. And I mean literally everyone.

Popularity prediction hash: 0e1f93ac058e1d0f88659fb55ac44b1a583e71474f3c0c311efdb41e9a1cf9b8

[1] Maybe bloggers are such an idiosyncratic bunch that there is no generalizable model. I’m fairly confident that this isn’t true though.

Flaws:

1. When there is text present in the search bar, the closing “X” overflows into the surrounding region vertically and on the right side.

2. The microphone icon is slightly misaligned with the search bar on the vertical axis.

3. When the search bar is in an active state, there are two redundant search bar icons (this one is more subjective).

For all the memes that working at Big Tech is just moving buttons around, even the almighty Google has yet to implement a flawless search bar. Keep in mind that Google has a strong incentive to perfect every aspect of the user experience for its 122 million+ daily active users.

As a recovering efficient markets believer, this feels like Google is choosing to walk by a 100 dollar bill every single day and refuse to pick it up.

In this post, I outline possible explanations for why the YouTube search bar is visually flawed, in order of least likely to most likely.

1. It’s a talent and prioritization issue (extremely unlikely).

Everybody on the YouTube team at Google knows that this is an issue. It’s sitting there as a ticket for somebody to fix but is low priority and is continually buried.

The highest quality talent on the YouTube team is only concerned with optimizing the algorithm, both for the primary click and watch portion of the website and YouTube Shorts. Somehow this visual fix is so complicated that you cannot fix this issue without looping in one of these engineers with the context and competence to fix this.

This feels extremely unlikely based on my anecdotal experience as a software engineer and from conversations with engineers who work at Google. There are plenty of agentic designers and engineers who should be able to fix this in a day.

2. It’s an Easter egg for identifying talent (very unlikely).

Google is known to put Easter eggs throughout their sites for engineers to source talent and for fun.

The YouTube search bar design flaws are another iteration of an Easter egg that sits in plain sight that over a billion people gloss over. I certainly did.

3. YouTube has to be backward-compatible with legacy screen types (very unlikely).

Military and medical websites are often visually unappealing due to the practical requirement to always remain backward-compatible with legacy screen types. To be fair, I only tried looking at the search bar on relatively modern monitors and all had the same visual defects.

Perhaps if I tried it on a much older model screen, this would look better. Even so, there is no reason why they can’t have specific renderings for screen types. This is something that most major frontend libraries have built-in.

4. YouTube is countersignaling to others that they are so dominant that blatant design mistakes don’t matter (moderately unlikely).

Google is blatantly countersignaling to the world that they deliberately put bad design into their search bar, and it remains the number one video platform by far. The “mistakes” they make are so minor that any half-decent Silicon Valley startup with good engineering and design principles would fix them.

The YouTube network effect is going on 20 years strong. Google not fixing these flaws serves as a middle finger to all startups trying to A/B test every pixel on their site to displace YouTube.

5. YouTube is signaling to users that they don’t want you to use the search bar (possible).

YouTube does not care about the aesthetics of the search bar because they want to slowly steer users away from using the search bar. This could be because they are developing an entirely new UI/UX interface for search and the search bar is going to be deprecated in short order anyway.

The best evidence for this is YouTube taking a more opinionated view on search and discovery. YouTube has begun displaying non-sequitur results when you search for a specific video under headings such as “You might also like” or “Top news”. Additionally, when a video is full-screen on desktop, scrolling down only shows recommended videos instead of the current video’s information and comments.

While users expect to use the search bar to find a specific video, YouTube actively believes that it can find a more engaging video.

If one were to subscribe to Occam’s razor, the most likely scenario is that somebody implemented this incorrectly and it sits in the graveyard of tickets that are not high priority enough to allocate engineering cycles to it.

Note: Multiple emails and conversations with Google engineers confirm that there is not a bug or mention related to the “X” or the microphone misalignment issue in the Google internal bug tracker.

Nothing about the design flaws are necessarily deliberate ex ante, but the energy and resources required to fix the problem are not worth the engineering cycles.

This feels intuitively correct while simultaneously deeply unsatisfying.

If you work on the YouTube team at Google or have any better explanations, please do get in touch.

Popularity prediction hash: adbfeae57d7a242fe2fe0a4ad74e6bac39aef2bed78606bb52ceb45e80e388c9

EDIT (2025-12-04): Added “Note: Multiple emails and conversations with Google engineers confirm that there is not a bug or mention related to the “X” or the microphone misalignment issue in the Google internal bug tracker.”

Algorithms like the YouTube algorithm seek to maximize platform engagement. Platform engagement is highly correlated with the amount of interesting material a user receives. Early versions of algorithms were heavily weighted to serve new content from large existing content creators. On average, a post from a creator with a large following is much more likely to be engaging versus the marginal creator with no track record. This is a tractable signal of a creator’s quality that is easy to implement as a model feature.

Since writing that piece, I have learned much more about YouTube microstructure thanks to many discussions with Justin Kuiper. Justin has spent over a decade as a YouTube writer and creative director, working for MatPat’s Theorist channels (Game Theorists, Film Theorists, Style Theorists, and Food Theorists). His videos have amassed over half a billion views.

The following is an interview with Justin.

Background

• How did you get into scriptwriting?

I got started with YouTube scriptwriting in 2017 when a big YouTuber named MatPat found a high-effort Reddit post I made about Pokemon lore and sent me a DM asking if I wanted to write for his channel.

But my scriptwriting career began before YouTube, I actually got my start in commercial scriptwriting around 2014, when I started writing visual novels. When I looked at the market for English visual novels I saw a lot of demand that seemed underserved, and there were a lot of developers who would pay a few hundred dollars for a week of work, so it was a great way for me to get my foot in the door.

• You’ve been thinking about this idea of a “promise delta”, can you briefly explain what that is?

It’s how I think about the difference between expectations in reality. Think about situations where you were promised the world, but the content failed to meet expectations. That’s negative promise delta. You got clickbaited. That sucks.

So it’s tempting to run in the opposite direction: And people then run in the opposite direction: if negative promise delta sucks, then positive delta must be good! “Under-promise, over-deliver.” It seems like a good thing to get something better than what you were promised. But actually, I’d argue that it’s not good: you shouldn’t undersell what you have, or else people will miss out on the good content because you failed to properly hype it.

So the goal should be to get the promise delta as close to zero as possible. You don’t want to over-promise, but you don’t want to under-promise, either. If you have something that’s amazing, don’t be shy about telling people that it’s amazing.

On Platform Mechanics

• Does the video medium have a chance at supplanting text?

No, text is too useful for looking stuff up. In fact, I find myself frequently making use of YouTube’s transcript feature so that I can control+F to a specific part of a 2-hour interview.

It’s also much lower bandwidth for conveying raw information. Average talking speed is around 150 WPM, average reading speed is around 200-300 WPM.

But there is a sense in which video can be higher bandwidth. I mean, it’s higher bandwidth in the literal sense, in that in order to transmit it to you, I need more bits; I’m literally delivering more data into your eyeballs and ears. And sometimes that part is actually signal. Because you don’t just care about my words as a robot would transcribe them, but you care about the intonation, the emotional valence, and so on. And then there’s every other part of the presentation. The music, for example, can set a tone. And as they say, “a picture is worth a thousand words.” It’s true: oftentimes just looking at a graph or diagram gives you a better sense of a thing in 2 seconds than a paragraph that takes a minute to read.

This is one of the things I’ve had to learn as a scriptwriter. When you’re writing an essay, there’s all these extra words that you add to give your audience a sense of the emotional valence of what you’re writing. But when you’re writing a script, and actually delivering it well, you don’t need all those extra words, because the emotional valence will hopefully be communicated by your face when you’re talking, or the BGM, or other parts of the audiovisual presentation.

• Is the video content market efficient? In a previous post, I argued that the written content market is efficient. Does this extend to the video content market?

I’d argue that the video content market is a lot more efficient than what exists for written content, because YouTube does such a great job of surfacing related content. “People like you watch videos like these.” If you watch enough Wendover videos, it will start recommending Polymatter and RealLifeLore.

You will open up your YouTube app to watch the latest Wendover video, and the homepage will recommend similar videos. And then off to the side, you will see “suggested videos.” That’s way beyond anything that exists for written content.

I think that Substack is trying to do this – trying to be the YouTube of written content – where they’re not just the place where everything is hosted, but they’re also the recommendation algorithm. It’s pretty good, but I don’t think they’re there yet in terms of personalization.

• What are some video content economics that most people are generally not aware of?

A lot of things are downstream of the fact that AdSense does not really pay that much. Like, a creator who is getting 10,000 views per video, that’s pretty decent, in a sense. But that’s only like $20-30 in ad revenue. That’s not a sustainable income.

But what you see among a lot of those smaller creators is that a lot of them are actually content marketers. Like they’ll say “I’ve spent 10 years working as a book editor, here are my tips for aspiring authors…and by the way, if you’re an aspiring author in need of an editor, here’s where you can hire me.” Or “hey, you watched my video on how to deadlift, here’s where you can book me to coach you.” Or “hey, I’m an indie author, you come here to watch me review fantasy novels, now read my fantasy novel.”

And it’s not all “little guys” who are doing the content marketing game saying “buy my Kindle novel for $5” or “pay me $100/hr for a coaching session.” Some of them are B2B SaaS platforms saying “buy our service.” And they’re happy to spend five figures on a video that might get less than 100k views. Because for a regular YouTuber, 100k views might be just a few hundred dollars in AdSense rev, but their customer LTV is so high, a single conversion might be worth thousands of dollars for them. These were some of my favorite clients to work with.

But you also have a lot of people who are just in it for the love of the game, they love making YouTube videos, and they keep uploading videos and investing hours of effort for something that only pays $20 in expectation. And what often happens to that smaller YouTuber is that a big YouTuber will spot them, and say “Hey, if you work for me, I’ll pay you more than $100 per video, a lot more than you’re making working on your own channel.”

And the end result is that you have a lot of specialization, because a lot of times when someone was getting a modest viewership on YouTube, it’s because they were good at only part of the job. Maybe they are a great writer, or a great editor, and getting hired lets them focus on the one thing they were good at.

• How long does it take you to notice a change in the algorithm? Who, if any, sees these changes before you?

I notice changes in the algorithm way more based on my own viewing habits. I use it every day, and see like a hundred suggested videos every day, so that’s the place where I notice trends like “Oh, YouTube is recommending more 20+ minute videos this week”, or “1-minute videos are back.” That being said, I don’t always know whether that’s representative of broader platform-wide trends, so it’s useful to have multiple YouTube accounts that each sort of represent different categories of viewer. Like if I’m starting an educational YouTube channel, and I am, then I should have a google account that I just use for watching the types of videos that I think my viewers will be watching.

I don’t spend a ton of time looking at the dashboard except sometimes in the days after a new video goes live, when of course I get invested in how the most recent video has done, but I don’t think I’m learning anything generalizable about how the “algorithm” might have changed based on that.

I know some creators who are a lot more obsessive about the numbers, there are group chats where they talk about these sorts of things. And those group chats are how you get a bunch of people who realize, “Hey, all 5 of us suddenly saw our view counts dip by 15%, all starting during the exact same week, except for Bob whose uploads are in a totally unique vertical, so clearly YouTube is doing something different.”

• Are thumbnails still undervalued by YouTube creators? What’s the limiting factor as to why there aren’t more creative thumbnails?

I think thumbnails are correctly valued by most successful YouTube creators, I don’t think they’re undervalued or overvalued.

One of the limiting factors is that YouTube thumbnails have to look good on a 5-inch phone screen, a 55-inch TV, and everything in between. And you’re juxtaposed there with a bunch of other thumbnails, the goal is to stand out at a glance. And you have to create a curiosity gap. In a sense, the thumbnail has to be “unsatisfying” but with the promise that you will get satisfied if you click.

I think there are some people who get really creative with their thumbnails. Nerrel, this guy who makes video game mods, is the one that comes to mind. He doesn’t have a clear “formula.” A lot of his thumbnails break every rule of “good thumbnail design.” A lot of them are intentionally kind of ugly and off-putting. I think that fits because that’s kind of his personality, he’s got an abrasive personality and that’s part of the appeal, he’s very punk rock, so when he does it, it feels very authentic, it’s congruent with his entire approach.

Synthet is also very skilled at thumbnails. It’s a music channel, how do you use a visual thumbnail to convey auditory concepts? And the answer is “pretty creatively.”

• Is there an ideal time to publish new videos? Is it specific to a channel’s target region?

Timing of the upload doesn’t really have much of an effect, but you can train a tiny fraction of your audience to expect uploads at a certain time. “New videos every Saturday morning”, or whatever. Those people are probably going to account for a tiny fraction of the audience. From the analytics I’ve seen, it seems like around 5% of subscribers enable push notifications, and of those, maybe 1% actually click on the push notification, and subscribers are already a fraction of total viewers. So we’re talking about a tiny, tiny fraction of the audience that actually cares when the video is uploaded. There’s an argument that having early engagement from “superfans” in the first hour or two helps to improve your early metrics, and the theory is that this would get YouTube to push your content out to more people, because in the first hour you had people who were leaving a comment and “liking” the video. I can’t confidently say it has zero effect, but it’s minuscule in comparison to overall clickthrough rate (CTR) and retention.

Overall my more general heuristic is that you don’t need to treat the launch of a new video as a “make or break” event where if you fumble things on day 1 things will be ruined forever. This has become even more true with thumbnail A/B testing, where now the first 24 hours or 72 hours are less about maximizing views, and more about gathering data so that YouTube can select the best packaging for the video to do better in the long run. Unless you’re doing very timely videos where you’re talking about the daily news, you should expect your video to have a longer tail.

On Content Dynamics

• Are most successful channels these days “marketing-first” instead of “video-first”? Is it always obvious when videos are “marketing-first”?

So to explain this for the sake of your audience, I think there’s two separate approaches to video-creation. One is more the “artist’s” approach, where you make the video that you want to make, and then you try to come up with a hook for it and a title or thumbnail for it. The question of, “How do I get people to watch the thing I made” is almost an afterthought. You’re David Lynch, you’re Werner Herzog. That’s what an artist does.

The opposite of the artist’s approach is the “content creator’s approach.” You start by thinking, “What does my audience want to see?” And then you make that video. In some cases you start with the title and thumbnail before you’ve even made the video. And, not surprisingly, these videos tend to do better.

And by the way, this is how a lot of media is made including IP-driven blockbusters. You have studios that start by deciding, “Okay, we have the rights to make a Spider-Man movie, we’re going to make a Spider-Man movie”, and then they find a screenwriter and a director and actors. Maybe they have posters and marketing before the movie is even done filming.

I think describing these sorts of videos as “marketing first” is accurate. The creator is concerned with clickability, or in the case of the blockbuster, they’re thinking about whether they have an IP that will get people to buy movie tickets. I think it can have a bad connotation, because people think of marketing as gross, and it feels crass and commercial.

But if you view it in a different light, in a sense, the content creator has to have more empathy for their audience than the artist does. In a way, it’s more humble. Like, I’m not so smart that I should assume that everyone else is interested in the same things I’m interested in. I should listen to my audience and let them tell me what they want to see. And okay, it looks like my audience likes this type of video, okay, I’ll make that.

It’s possible for this to go too far. Again, you see this in Hollywood. It’s cool that they gave Sam Raimi $140 million to make a movie because they were confident that Spider-Man would sell movie tickets. But if you care too much about the marketability, the execs will decide “Venom has to be in the movie, audiences want to see Venom”, and then Spider-Man 3 kind of becomes a jumbled mess with three villains. It’s the Hollywood version of audience capture.

• Have there been any videos where you just can’t understand how they are popular or why the algorithm is boosting it?

There are lots of videos that have made me think, “okay, why has the algorithm decided to show this to me? What made YouTube decide that I’m the type of person who wants to see this video?”

But I don’t think there’s ever been a video where I truly thought, “Okay, 2 million people watched this video, and the reasons for that are inconceivable to me.” I might disagree with those people, I probably don’t share their taste, but I can still look at a thing and understand, “Oh, yeah, this is what some people are looking for, and YouTube has successfully found the audience of people who want to click on that video.”

On Talent

• YouTube is unique in that it pulls from a global talent pool. Where in the cycle of professionalization is YouTube currently? How does it compare to where it was 5 or 10 years ago?

The talent pool for thumbnail designers and editors is pretty liquid and lots of people are making a living as “full time freelancers.” I don’t think we’re really headed toward a world where the majority of YouTube people become W-2 employees who work for a few big channels, but I think one of the benefits of YouTube for workers is that they can have more leverage. They can skill up quickly, find bigger clients, learn a lot in a few months, and get themselves a raise.

I will say that it’s really hard to hire good writers. I realize that’s kind of self-serving, because I’m a writer, but everyone I talk to feels this way, and it was true for me when I was trying to hire more writers as a creative director. Sam Denby from Wendover Productions has talked about how it’s just really hard to hire good YouTube writers, and from what I understand he goes for journalism school grads and trains them to write YouTube scripts. The market for writers is higher-friction than for editors or thumbnail designers, there is much more of a process of learning the “voice” of the channel.

I think a big part of the challenge of trying to hire a “YouTube scriptwriter” is that writing across verticals isn’t super fungible. You have some YouTube scriptwriters who are basically copywriters for verticals where you’re aiming for something close to “LinkedIn-speak”, and then you have educational verticals where you have writers who are spending a week reading academic research papers and books and summarizing what they’ve learned, and then people who are doing deep-dives on hobby stuff, or celebrity drama. All of those are kind of different skillsets.

• How often does YouTube talent get poached? Are there generalist superstar people who work behind the scenes of various channels that are highly sought after?

“Poaching” in the traditional sense doesn’t happen so much, because you don’t have a lot of W-2 employees to begin with. It’s not like when Adobe would go to engineers and say “Hey, we’ll give you money to quit your job at Apple and work for us.” There are some channels that hire some full time employees, at the very top levels, but a lot of the ecosystem is freelancing.

That’s where you have a sort of “soft poaching” that happens. Like, “I’m going to find the editor who works for that guy, and hire him to be my editor.” And if enough people are bidding for that editor’s time, eventually he drops the clients that are paying him the least, or that he enjoys working for the least, and reserves his time for the clients who pay him the best or are the most fun to work with.

And by the way, there’s this open question on YouTube of how to credit people in video descriptions. Because if someone credits their editor in the video description, then you can get in touch with that person and hire them. If they don’t want you poaching their editor, maybe they don’t credit them. And if you’re that editor who is not getting a video credit for that reason, my advice to you would be to charge your client a premium for that. You don’t even have to give up the right to feature that work in your portfolio.

Back when I was doing work for B2B SaaS clients my rate was something like:

1. You can pay $X and credit me and put my social media link in the description

2. You can pay me 1.5 * $X and not credit me in the description but I’ll feature you in my portfolio

3. You can pay me 2 * $X and this will be 100% ghostwritten.

And basically all of them opted for the more expensive rates.

• 30% of kids want to be a YouTuber. Do you think this is a good equilibrium of talent allocation in the long-run if this continues to be true?

No, I think a big part of this is “kids name a profession that is legible to them.” Like if you ask kids what they want to be when they grow up, a lot of them will say “teacher” because that’s the adult profession that they interact with the most often.

Kids say they want to be an athlete, doctor, vet, firefighter, or a YouTuber because those are legible professions. Their teacher says, “you have a talent for soccer” and they think, “Oh, I’m good at soccer, maybe I’ll become a pro athlete like my hero, Messi.” Nobody kid is saying “I want to be an actuary” or “I want to be an auditor.” And, you know, the number of kids who can actually grow up to have a job related to YouTube is actually way higher than the number who can grow up to be a veterinarian. Not everyone can be a “famous YouTuber”, but you can make a good living as a video editor.

On Market Structure

• Will click and watch platforms (YouTube) ultimately lose to the get served content platforms (TikTok, Instagram Reels)?

I don’t think so. YouTube gets a ton of high-intent traffic, because when you see a YouTube video, you chose to click on it; it didn’t just start auto-playing because you swiped up.

That high-intent traffic is way better for community-building, and having a fanbase, where you show up to see your favorite creator, as opposed to just watching whatever video starts auto-playing when you open the app. That’s what you get with a “click-and-watch” platform like YouTube.

And of course that high-intent traffic is valuable to advertisers. I talk with advertisers about this, and TikTok is cheaper for sheer reach, but YouTube is higher conversion for anything beyond an impulse-click like “install this free app” because it is higher quality and higher intent.

• Since good writing is one of if not the limiting input factor, should we see more scriptwriters start their own YouTube channels? Why haven’t we seen more of this yet?

I think part of it is that if you’re really good at scriptwriting, and bigger YouTubers recognize that, then working on your own channel probably isn’t the highest-leverage thing you can do. Like, Paddy Galloway has his own channel, where he could work on a video that will be seen by a million people, or he can work for MrBeast, whose videos reach 100 million people. And if you’re a Paddy Galloway, your comparative advantage isn’t being the on-screen talent or providing capital. If he’s one of the best on YouTube and his goal is just to “make cool videos”, it makes way more sense for him to work with Red Bull and work on videos with drones and F1 cars rather than try to bootstrap his own channel.

• You make the distinction between “creators” and “artists”. Because you describe YouTubers as “creators”, does this cap their potential influence? Are we at a global maximum for their cultural reach/impact?

I think I sort of got at this a bit earlier, where the distinction between “creators” and “artists” actually has a parallel in the world of Hollywood, where if you think about it, the people making big IP blockbusters are more like “content creators” who are paying attention to what the audience wants, rather than “artists” like Robert Eggers or Yorgos Lanthimos who are just going to make their thing and say, “This is the thing I wanted to make, take it or leave it.”

And if you look at that dichotomy, it’s obviously the big IPs that have the biggest influence, and the same is true on YouTube with mega-channels like MrBeast. The thing that both Marvel and MrBeast have going for them, too, is that they’re global brands. You can watch MrBeast dubbed into German, French, Brazilian Portuguese, or whatever language you want. So if anything, I would think that “content creators” would have more potential reach than “artists”.

I guess the limitation comes in where content creators aren’t really coming up with new IP. Like MrBeast can make a Squid Game video that gets nearly a billion views, but to actually get a Squid Game in the first place you need to find someone like Hwang Dong-hyuk. Again, parallel to blockbusters. It’s super rare for a new IP to show up on the big screen. I’m not sure how you solve that.

Miscellaneous

• Why is there only 1 long-form video platform like YouTube?

I’ve written about this before, the thing I’ll say in summary is that there are other long-form video platforms out there like Nebula, the Dude Perfect app, and Dropout. Those were all started by YouTubers, by the way, and they don’t exactly compete with YouTube, because they are selective. You have to get invited to Nebula. Dropout is its own separate thing. And they rely on YouTube for a lot of their discovery.

But nobody wants to be YouTube, the site that anyone can upload to. And that’s because it’s expensive to let anyone upload! And then you have to make sure they’re not uploading anything illegal. And then you have to host gigabytes of video forever, that’s getting seen by no one. Even Twitch, which is owned by Amazon, doesn’t want to do this, they prune Twitch VODs.

I think there’s room for a lot more Nebula-tier services, where it’s more selective than YouTube, but less selective than Netflix. But YouTube is always going to be the place where beginning creators start out.

• If you were a VC and were looking to deploy into the YouTube ecosystem, where would you invest?

Oh man, I don’t think this is a question I’m qualified to answer. You know, I think my problem is that I’d go after the things that look like “safe bets” and not the higher-EV moonshots. All of the ideas that come to mind are things that I myself would do if I just had a tiny bit more risk tolerance. The thing that I think has maybe a 30% chance of succeeding and becoming a lifestyle business for the next 10 years is probably content marketing, like trying to do what Jed Herne is doing. And that’s not what the VC wants, right? They want a 1% chance of funding the next Mark Rober, not a 30% chance of being the next PolyMatter or Jed Herne.

So if I were going to play VC and take the moonshot bet, and I’m thinking about a vertical where I know enough to do due diligence, I’d take a bet on VTuber agency. It’s a vertical I’m reasonably familiar with and understand the culture and the scene. I personally know several people who are VTubers. I co-designed and scriptwrote a game that was pretty popular among VTubers. Starting an agency that’s trying to mimic the model of the biggest Japanese VTuber corporations would be pretty capital intensive, which is probably why more people haven’t tried to do it, but that’s what the VC bucks are for.

When VShojo imploded, and then left all their talent out to dry, that would have been the moment for someone to swoop in and say, “Hey, come join us, we’re going to be the big US VTuber Corp that doesn’t suck.” So it feels like there’s unmet demand for this.

• At the far right tail, is one’s sense of virality learned or innate?

Definitely learned. If you look at MrBeast’s early videos, it’s very obvious that he did not have all of the skills and knowledge that he has now. You can see he has a few correct intuitions, but doesn’t exactly know how to apply those intuitions. Like he knew from the start that “bigger is better”, but that’s not a unique insight. The reason his videos go viral in 2025 and didn’t go viral in 2015 is that it’s really easy to say, “I need an eye-catching thumbnail that teases a thing that people really care about” and it’s really hard to actually execute on that without a lot of specific and deep knowledge.

• If you could wave a magic wand and make changes on YouTube, what would the top 3 things you would do?

YouTube has this “channel membership” thing where you can pay a monthly subscription to channels in exchange for access to videos that they’ve gated behind a paywall. I like the concept, but the execution kind of sucks. It’s too hard to actually get to a screen that just shows you all the paywalled content you just paid for, and the UI for navigating that content isn’t great. I have paid for a channel membership several times, and it was not an experience that filled me with joy, delight, or enthusiasm at the prospect of doing it again. So they definitely need to change that.

And a related issue to “making it suck to pay for content through the app”, YouTube also has a feature where you can sell courses, and I am shocked by just how much it sucks. I think it’s literally broken. Like I’ll go to a 30-video course where several of the videos are paywalled, then it just shows the text, “This video requires payment to watch.” And it doesn’t prompt me to pay. There’s no button taking me to a page with an interface where I can pay. It literally feels broken. I don’t know why YouTube is allergic to me trying to give them money through their website or app.

And lastly, I would like more control over the recommendations. I know that this is dangerous to mess with because this is the “secret sauce”; clearly they shouldn’t trust my stated preference over the revealed preference over millions of users, but I think it would be nice to have a “temperature slider” where I can say, “I’m in the mood for something totally different today, feel free to get creative”, or “right now I’m watching a specific type of video and I want my recommendations to look very similar to this.” You can kind of do that with the clickable categories but I’d like more ability to fine-tune it.

Popularity prediction hash: b4f1f9d78560319493df5add7ecf4729fa24bdf8f8ad8ff477736de49c2f7797

Late-night productivity seems to be a common pattern, especially among the writers and programmers I know. Most blogs and research online cite circadian rhythm and innate biology as the primary reasons for being more productive at night.

I want more control and agency over when I’m productive. I want to do something else at night that’s not just lounging back in my chair by myself typing away. Having a consistent and normal sleep schedule would also do wonders for my social life.

I’ve experimented with countless strategies to force myself to be more productive during the day. Some things I’ve tried include:

• Blocking certain websites and even Wi-Fi

• Changing exercise times

• Restructuring my meal schedule

• Scheduling social plans later at night

None of it worked.

In this post, I provide six explanatory mechanisms for why I am more productive at night, in order of anecdotal explanatory power.

Following these explanations, I describe the only universally agreed upon solution I’ve heard for permanently shifting productivity back into the daytime.

1. Late nights are associated with weakened mental filters, leading to less overthinking and easier flow state access.

In my last post, I claim that one reason late night conversations are better is because people are more honest due to weakened mental filters:

People’s inhibition levels are lower due to fatigue and possible inebriation effects from other late night activities such as alcohol. This weakens our mental filters and leads to more direct communication, as people say what they actually think rather than a coded version they would say with all their mental faculties.

Lower inhibition levels also lead to faster response times, which are correlated with signaling social connection in conversation. When we’re saying what we think without processing second and third order effects, people respond quickly (< 250 ms) and the conversation flows.

Furthermore, our willingness to be honest creates a virtuous cycle where people are continually willing to ask more personal questions, allowing us to connect on a deeper level with others.

The same mechanism that makes late night conversations more honest also applies to individual work. The frontal and prefrontal cortex are primarily responsible for regulating and coordinating other brain regions, with much of this function involving inhibition of impulses and thoughts. This inhibitory control requires finite cognitive resources that become depleted throughout the day (ego depletion). This state enables greater access to our stream of consciousness without the typical blocking mechanisms.

The most compelling evidence comes from working while jet-lagged. When my body thinks it’s midnight but local time says noon, I can access that late-night cognitive state during normal working hours.

2. There are fewer distractions and obligations late at night, so this is when I can grind without interruption.

The fewer distractions and lower opportunity costs can be direct (actual interruptions or missed events) or indirect (the expectation that you check your phone or a sense of missing out on something).

3. Late-night work is more productive because you’re more likely to have thought about the task throughout the entire day, which allows the work to metastasize.

Although humans can only sustain about 4 hours of deep cognitive work daily, the remaining hours serve as important “incubation hours”. The 10+ hours of conscious and subconscious incubation enable my work at night to draw from that pre-processing.

4. Physical tiredness from daytime activity makes evening indoor desk work feel natural.

Sitting down for focused work feels restorative rather than restrictive after a day of physical movement. I am much more restless during the day when I haven’t burned 2,000+ calories of energy yet, making it harder to concentrate while sitting at a desk at a computer.

5. I’ve Pavlov’ed myself into being the most productive at night due to my late-night work habits during high school and college.

Years of last-minute assignments and exam cramming have hardwired my brain to associate nighttime with serious work. Throughout high school and college, my brain became used to daydreaming through daytime classes, socializing between lectures, then actually learning the material alone at night.

6. It’s easier to get into a flow state in the dark when my brain knows it should be dark outside.

I’ve tried recreating this with blackout curtains during the day, combined with white noise and AirPods to create a controlled sensory environment.

While this helps marginally, my brain still knows that it should be bright outside, and sometimes I end up peering outside my window. The artificial darkness ultimately throws off my circadian rhythm even more, pushing my biological clock even later.

I’ve only heard one solution that universally works: have kids.

Your body simply doesn’t have the energy reserves for midnight work sessions anymore after multiple middle-of-the-night crying sessions.

If late-night productivity is secretly just training for those 2 am baby bottles, well, at least it’s preparing for something.

Popularity prediction hash: 2578e800fc76a460adeec5ebc109405aa56e2092c5446d055a7d1a64a161f6e7

Whenever I mention this phenomenon to friends, they overwhelmingly agree. Almost all of the canonical conversations that defined our relationship – the ones that we still reference years later – occurred late at night.

In this post, I provide five explanations for what makes late night conversations better, in order of explanatory power.

1. Late night conversations serve as a multidimensional commitment filter, leading to less ambiguity around intentionality.

Late nights are socially understood to be personal time. All parties are signaling that they’re willing to use their non-work hours to be present at that moment. This creates a peer-to-peer context with far less ambiguity around whether any networking motives are in play.

Because the discussion is happening late into the night, anyone can decide to leave whenever they want with a legitimate reason to go to sleep. When people elect to stay, they are collectively demonstrating their willingness to sacrifice their sleep to participate in the late night conversation.

2. Late night conversations typically occur with people we genuinely enjoy being around.

The people we have late night conversations with tend to be those we naturally get along with better. These can be old friends with pre-existing context or new friends that all parties expect they might get along with.

3. Late night conversations are better because they are longer and more focused.

Most conversations during the day are time-constrained and littered with distracting messages we feel obligated to respond to. Good conversations are a series of doorknobs, and the most interesting parts of a conversation occur at the second or third hour mark. Additionally, people are more mentally engaged at night, as there is less of an expectation to respond to messages. This allows people to detach from their phones, a major source of distraction.

While conversations don’t end when people actually want them to, longer conversations also lead to higher satisfaction that people got what they want. On average, people’s desired conversation time differed from their partner’s desired time by seven minutes. As conversations get longer, the seven-minute preference gap effectively shrinks to zero as a percentage of the total conversation.

4. Late night conversations are better because people are more honest.

People’s inhibition levels are lower due to fatigue and possible inebriation effects from other late night activities such as alcohol. This weakens our mental filters and leads to more direct communication, as people say what they actually think rather than a coded version they would say with all their mental faculties.

Lower inhibition levels also lead to faster response times, which are correlated with signaling social connection in conversation. When we’re saying what we think without processing second and third order effects, people respond quickly (< 250 ms) and the conversation flows.

Furthermore, our willingness to be honest creates a virtuous cycle where people are continually willing to ask more personal questions, allowing us to connect on a deeper level with others.

5. Late night conversations tap into an evolutionary trust window tied to intimacy and vigilance.

For most of human history, the only people awake with you after dark were tribe-mates you trusted with your life. This created a predictable context: low light, fewer interruptions, and a small circle of familiar faces.

The hours associated with sex, whispering, and shared vulnerability naturally manifest into a rich late night conversation. They shift our minds from performance mode to connection mode.

It’s also worth playing devil’s advocate here. It’s possible that late night conversations aren’t objectively any better, but rather a combination of being tired, inebriated, or some other mechanism that alters our recollection of the conversation.

Whether the effect is causal or retrospective, remembering these conversations as unusually better is itself evidence that something notable is happening.

Popularity prediction hash: 4dc19916e1f154437638731e4cb4e994f580d3a4154e2e53e6f780e3ec7dbe47

Thanks to Ben Pace for posing this question in a discussion and sparking this blog post.

The Black Box Model of Outsourced Human Thought

The black box model of outsourced human thought can be understood as a deterministic function where any arbitrary input (diversity of humans) produces a specific output. The inputs can be any question such as “who to trust for medical advice” or “what the meaning of life is.” Each black box espouses a certain worldview that they seek to disseminate.

Black boxes are necessary for any functioning society. Any given person simply does not have the time or expertise to reason about everything they encounter from first principles. Black boxes are proxies to scale trust. When you don’t understand the mechanism or what is really going on, you outsource how to think and feel to various instantiations of black boxes.

Religion, institutions, and AI have been the primary instantiations of different types of black boxes. In many cases, the mechanisms of the black boxes are difficult to understand even for the small in-group elites themselves, as they may consist of too many moving pieces or be too technically difficult. Often, the black boxes resemble the complexity and evolution of a living, breathing organism.

Religion

Mencius Moldbug (Curtis Yarvin) describes religion as an organization or movement which tells people how to think.

Every major religion has a storied history of being written by or for divine figures. These divine figures represent a small set of in-group elites who tell a story about the world.

Religion has been passed down through people. People from diverse backgrounds and life experiences enter religion (often through a church) and come out shaped into what all major religions aim to produce: productive stewards of their communities.

Institutions

Institutions were the next chronological evolution of the black box. Instead of being primarily based on faith, institutions generated social capital for themselves, nominally through scientific reasoning and rationality. Nevertheless, the deterministic function is the same: transform a diverse set of people and impose the institution’s worldview on how their constituents should think.

These institutions can be public institutions run by the government (e.g. NIH, FDA, CIA), private institutions run by civilians (e.g. Arc, McKinsey, Cato), or somewhere in between (e.g. Harvard, Stanford, Mayo Clinic). These are the same gating mechanisms that pervade other professional industries including law and medicine. The administrators who run these institutions are a small in-group elite who typically rise to power through appointments by the existing in-group.

Because of this circular nature, these institutions are kept in check by other elites and typically decay from the original founder’s vision. They are able to accrue and maintain their social capital by creating an equilibrium where they connect ambitious, talented people with the networks of existing power players.

AI

We are in the very early innings of AI, specifically LLMs, as the next development of a black box that people utilize for outsourcing thought. The pervasive religiosity around certain models is early evidence that these are more than tools.

Just like religions and institutions, a core in-group elite controls the means of production of AI models. Unlike religion and institutions, the levers the AI research labs can control are technological in nature. They can unilaterally change model weights with outsized rippling effects on the economy.

In this way, AI is the ultimate centralizing and communist force.

The (New) Straussian Moment

COVID provided a unique insight into the capabilities of the people across various institutions. While some institutions responded effectively, many demonstrated their incompetence and orthogonal ulterior motives.

These institutions and people are rationally guided by their personal incentives, which leads to extractive behaviors that are not always aligned with the long-term health of the country (e.g. housing policy).

While rapid news cycles and technological disintermediation (e.g. prominent figures going direct via X and YouTube) theoretically accelerate the creation of new public institutions, it has mainly eroded our collective faith in existing institutional capabilities.

Prediction Markets as the First Bottom-Up Black Box

Prediction markets are the first black box mechanism guided bottom-up without direct elite mediation. Furthermore, they are the first permissionless black box where everybody has legitimate direct input into changing the output.

Prediction markets achieve this by aggregating distributed intelligence of all interested parties. If built on crypto rails, prediction markets are an uncensorable mechanism for human coordination.

Having been theorized in libertarian circles for decades, prediction markets represent the ultimate libertarian market mechanism. Pundits and executives can be quantitatively judged based on performance, creating a permissionless environment for anybody to reach the top of their field, regardless of prior connections with the existing elite.

Taken to their logical extreme, a world guided by prediction markets looks something like a Balajian Network State, where all global dark talent is discovered and unleashed in their ivory towers. However, there are many local equilibria between the current state of the world and the world we can push for.

Prediction markets, in their current form, are far from incentive-compatible. Just like religions, institutions, and AI, the market structure will evolve to fit the wants and needs of its constituents.

Popularity prediction hash: ec56ba2587baf40a59aa69d525aa4bab6377d437db59126434fffd01b39a9444

1. Market order: takes liquidity off of the book and execute at the best available price

2. Limit order: places liquidity on the book which will be matched if there is a requisite matching order

When a user sends a market or limit order, they are instructing the frontend to build a transaction with their exact parameters. The frontend then sends the constructed order to the orderbook matching engine. The user must decide exactly how their trade should execute, even if another venue, route, or execution path could deliver a better price.

In this post, I outline why intent-based aggregators will be the preferred retail prediction market interface and provide specific examples of novel orders enabled by intents.

A brief primer on intents

An intent is a set of constraints that allow a user to outsource transaction creation to a third party without relinquishing full control to the transacting party. They allow the user to tell a third party “what” they want without caring about “how” it is achieved.

These intents are typically sent to an aggregator. The aggregator is responsible for figuring out how to give the user the best price. Aggregators achieve this by asking a set of fillers for a quote and routing to the filler with the highest quote.

UniswapX was the first major aggregator to implement intents. Today, all the major DEX protocols operate as an intents-based aggregator, albeit with slightly different implementation nuances.

Intents are a meta-order type that allows for more flexibility. This enables novel user interfaces that serve as wrappers around different intent orders. One such novel user interface is an LLM-based interface that accepts user orders in plain English. These LLM-based interfaces can then easily translate and encode them into an intent order.

Intent-based prediction market aggregators will provide the best UX

Equities and tokens are fungible assets. 1 share of META = 1 share of META (within a given share class). 1 BTC = 1 BTC.

Prediction market assets are non-fungible, as each individual market may have different resolution criteria. In the simplest case, the same event might point to different resolution sources. For example, crypto markets on Polymarket resolve based on the Binance spot price, while Kalshi markets resolve based on CF Real-Time Indices.

Other markets such as the “US National Bitcoin Reserve in 2025?” have different resolution criteria across prediction markets (Polymarket, Limitless, Kalshi). Polymarket and Limitless resolve YES if the US government holds any amount of bitcoin, while Kalshi resolves YES only if there is a designated National Bitcoin Reserve (Appendix C). While these markets exhibit high covariance, they could resolve differently, adding another layer of complexity for users to navigate.

If an aggregator were to treat a given market as canonical, this would shift the responsibility for oracle resolution to the fillers. Fillers would be responsible for managing risk and their covariance to accurately manage the nuances of the non-fungibility of markets. Intents provide the incentives to manage these markets at scale to give the users the best price.

This transforms illiquid, segregated, and non-fungible prediction markets into a liquid and fungible interface for users. By abstracting away market-specific complexity and providing superior liquidity through cross-venue aggregation, intent-based aggregators will capture the majority of retail prediction market volume.

Intents enable intuitive and novel order types

The following are some examples of prediction market orders that leverage intents:

1. Buy to price orders: a user buys shares of a given market until they cannot get any more marginal shares at $X per share.

2. Buy to average price orders: a user buys shares of a given market until their average share cost basis is $Y per share.

3. Conditional orders: a user buys shares of market A when market B crosses some price threshold $Z per share. This can be chained with simple market orders or chained with other order types such as “Buy to price” orders.

Popularity prediction hash: 171e5c1af4a900e4e164c01d004e7d985cd6da11a2a7e6c5a50065dc5aeb9544

This is what Polymarket and Kalshi looked like in 2022:

Three years later, their interfaces are almost identical:

While this made sense initially for early user acquisition, the lack of any meaningful changes directly affects market efficiency, liquidity provision, and price discovery. Opinionated frontends and aggregators should differentiate themselves along different axes to develop specific tools that people want.

Third-party frontends have similarly failed to innovate thus far, despite the clear opportunity. Prediction markets are still awaiting their native trading interfaces.

In this post, I outline six problems with prediction market interfaces today. For each problem, I offer commentary and a solution.

Note that the scope of this post is limited to the specific market event pages, not the global interface.

Core Interface Problems and Solutions

There are six main problems with the interfaces today:

1. Problem: There is no way to view current or historical market depth or understand the confidence one should have in the market.

Prediction market platforms typically use the midpoint price as the price they tout as the probability of an event occurring, regardless of the spread:

(Prediction market platforms are incentivized to deceive users in this manner because the vast majority of markets are not liquid enough to output any meaningful information.)

Solution: Offer a historical chart with liquidity levels at a specific point in time.

2. Problem: There is no intuitive context mapping between prediction market prices and real world events.

While Polymarket has started to do this for select markets, it is unclear and not intuitive for most users. Price changes should be contextualized with specific news events.

The graphs can be user-generated where users are able to add annotations to the graphs where everybody else can vote on.

Solution: Overlay specific news events on the market graph.

3. Problem: Every market uses identical YES/NO interfaces regardless of underlying question structure, forcing unnecessary cognitive translation.

The market pages themselves are completely standardized. A market showing a sports game shows the same YES/NO feed as the market for the number of Elon tweets and AGI timelines.

Solution: Display standardized interfaces across question types: price predictions, ranking interfaces, calendars, percentage ranges, geographies.

• What will Bitcoin’s price be by the end of the year?

Similar specialized interfaces can be made for other question types:

• What will inflation be in Q3?

• Map-based UI for geographical questions

• Ranking interfaces (1st, 2nd, 3rd)

• Calendar interfaces for release dates

4. Problem: Portfolio PnL is displayed based on midpoint price.

When a user holds shares across event markets, the USD value displayed on the UI does not match what would occur if the user were to market sell their positions at that point in time. While many other traditional and crypto exchanges operate in a similar fashion, the illiquidity of prediction markets make this especially pernicious.

Solution: Allow a toggle for users to incorporate a mark-to-market portfolio instead of using midpoints.

5. Problem: Insufficient funds trigger multi-screen portfolio liquidation flows.

If a user has $80 of cash and wants to submit a $100 buy order, the user is forced to navigate to their portfolio page and decide which position to sell.

Solution: Offer a pop-up that allows the user to confirm selling their shares in the highest liquidity, lowest slippage markets automatically.

6. Problem: Chat is unusable given the signal to noise ratio.

Solution: Include the ability to sort comments by aggregate PnL and category-specific PnL.

Popularity prediction hash: 25c3236575b1d12d5da4fa8a530b00363f2dd64ef408f22c10d29db4a4ed1f56

Many play money platforms, such as Manifold, have strong communities that persist to this day. However, the rise of Polymarket and Kalshi provides concrete incentives for sharp traders to trade on real money prediction markets instead of play money platforms. Many early Manifold traders became early active users of Polymarket and Kalshi.

In this post, I examine the performance of play money prediction markets compared to their real money counterparts. Following this analysis, I outline several advantages in their market structure and describe a future where they function as important actors in the prediction market industry.

Quantifying prediction market performance

The Brier score is a scoring rule that measures the accuracy of probabilistic predictions. It ranges from 0 to 1, where 0 represents perfect accuracy and 1 represents the worst possible predictions.

The score is calculated as the mean squared difference between predicted probabilities and actual outcomes. For a single binary prediction, if you predict probability p and the outcome is o (1 if an event occurs, 0 if not), the Brier score = (p - o)². Lower scores indicate better calibration and resolution of predictions.

Brier.fyi uses Brier scores to evaluate prediction market accuracy across four prediction market platforms: Polymarket, Kalshi, Manifold, and Metaculus. Polymarket and Kalshi are the two largest real money prediction market platforms. Manifold and Metaculus are the two largest play money platforms. Manifold uses “mana”, which is play money currency. Metaculus operates as a forecasting website with slightly different prediction mechanics, but can be interpolated into many of the prediction market aspects.

Polymarket, Kalshi, Manifold, and Metaculus have Brier score as follows:

Unsurprisingly, real money prediction markets outperform their play money counterparts. This makes intuitive sense: if price discovery occurred on play money prediction markets, someone could easily arbitrage between play money markets and real money markets.

Nevertheless, they perform better than one might expect and actually outperform real money prediction markets on certain axes such as science markets (Metaculus vs Kalshi).

The advantages of play money prediction markets

1. They do not fall victim to numeraire effects.

In my last post, I described one reason why prediction market prices are not always equal to their respective probability of an event occurring. This is due to numeraire effects, as the utility of the currency may change under certain cases.

Play money eliminates these frictions as traders are not incorporating any real money valuations. The artificial currency becomes a pure scoring mechanism rather than a store of value, allowing prices to reflect collective probability estimates.

2. There may be some selection effects between specific traders and market categories.

The data shows Manifold and Metaculus performing relatively well in science markets despite using play money. Science questions often attract specialized audiences—academics, researchers, and subject matter experts who trade for intellectual engagement rather than profit. These markets have naturally thin liquidity even with real money, as the audience is limited and payoffs can be distant.

In such domains, the marginal benefit of real money incentives diminishes. A physicist forecasting particle physics discoveries is likely motivated more by demonstrating expertise than earning $50. Play money markets can actually attract more of these experts by removing the friction of financial onboarding while preserving the competitive and reputational elements.

3. The number of traders in markets is so low that there isn’t much difference in actual traders.

The government shutdown market has been consistently featured on the homepage of Polymarket as this is the market that people want information on. In total, only around 1,000 unique addresses have traded on this market.

4. You can acquire better information for long-term markets.

There is no opportunity cost or associated platform risk for markets that resolve over 10 years in the future. Play money markets are the only platform where you can feasibly ask questions like:

• What will the human population on Mars be in 2075?

• Will a grandchild of George W. Bush become the US president?

• Will we discover extraterrestrial life by 2050?

Community-driven infrastructure for collective intelligence

Wikipedia pioneered free access to information and proved that you can build one of the most important internet primitives without strong financial incentives. Just 1,300 people create over 75% of the new content posted to Wikipedia every day, demonstrating how small, motivated communities can generate enormous public value without monetary incentives. These core contributors are motivated by intellectual curiosity, reputation within a small community, and the satisfaction of contributing to something greater. These are the same traits as the play money prediction market platform founders and early traders.

Play money market platforms can provide the substrate to experiment with new market design and market structures that real money markets can later adopt. A dedicated core of forecasters, motivated by in-group social status hierarchies and intellectual curiosity, can generate unique insights on specific questions.

Popularity prediction hash: fa1d37d0c53f690dea15407f14d69c038d5d878b6471044cc1f316eb76940c17

Polymarket’s own website states this equivalent as fact:

Prices = Probabilities.

Prices (odds) on Polymarket represent the current probability of an event occurring. For example, in a market predicting whether the Miami Heat will win the 2025 NBA Finals, if YES shares are trading at 18 cents, it indicates a 18% chance of Miami winning.

These odds are determined by what price other Polymarket users are currently willing to buy & sell those shares at. Just how stock exchanges don’t “set” the prices of stocks, Polymarket does not set prices / odds - they’re a function of supply & demand.

The logic mirrors the question posed to market bears who performatively claim asset prices are overvalued: “Are you short?” Here is a liquid venue to express your view and profit if you’re right. When asked, most aren’t actually short.

Last year, Dan Robinson posted a theoretical question about prediction markets in an X poll, asking what the price of a prediction market would trade at, given a known probability of an event occurring:

According to Dan, only 403 out of 2,293 people answered this correctly.

Stop here to think about it before scrolling down for Dan’s solution.

Dan’s solution

If the title didn’t already give it away, it is not the most popular answer: $0.75 (A).

Dan claims that the answer is D: Anything between $0 to $1. His rationale is as follows:

1. Prediction market payouts are denominated in a base currency such as USD. Users use USD to buy contracts and are paid out in USD if the market resolves in their favor.

2. Some prediction markets result in changing the value of the underlying base currency on a per unit basis.

3. Traders engage with prediction markets with full context on the value of the base currency with respect to its ability to purchase goods and services, not in a dollar for dollar vacuum.

4. This means that the different outcomes prices do not necessarily correspond to respective probabilities but rather the average of every trader’s utility function in those two cases.

So a market that asks, “Will the US exist in 5 years?” will not have prices that reflect the actual probability that people believe. Instead, each outcome will trade at the following:

• YES: $1

• NO: $0

NO shares should trade at $0 because if the US does not exist in 5 years the dollar is worthless. So nobody would rationally pay any money for a NO share today because the payout in the case that they are correct would give them no benefit.

Note that this creates some technical edge cases (outside of Dan’s original line of questioning). If the market is on Polymarket, USDC is the numeraire. The market should account for scenarios where the US collapses but USDC retains value through social consensus (low probability, but non-zero). In that case, NO shares might trade slightly above $0.

In the furthest extreme where a question instead asks, “Will humanity exist in 5 years?” Money is completely worthless in the NO outcome, so that market should trade at exactly $1 for YES and $0 for NO, regardless of extinction risk.

Context is that which is scarce, especially in prediction markets.

Back to reality

The natural objection is that this is a clever thought experiment but doesn’t matter in practice. Real prediction markets ask questions where the currency retains value regardless of outcome.

I would then point you to the Jesus Christ market, and, yes, this is a real market:

The value of the dollar in the case of Jesus Christ returning is an exercise left to the reader.

But even for serious markets, there are cases where prices diverge from probabilities.

The US presidential election is one example. “Will Trump win the US presidential election?” (very likely) did not trade at pure probabilities.

If you believe Trump’s policies would weaken the dollar (higher inflation, tariffs, fiscal expansion), then a Trump win means your payout is worth less in real terms. The market price in that scenario would reflect:

Trump winning = X * (purchasing power of $1 under Trump)

Trump losing = (1 - X) * (purchasing power of $1 under Harris)

Just looking at the Polymarket homepage suggests that there are quite a few markets where this logic might apply:

Misc. Thoughts

• Can we quantify the numeraire distortion in real money markets today for a specific market?

• How many markets should one expect this to apply to?

• Play money markets perform extremely well relative to what one might expect given the lack of economic incentives. How much can be attributed to numeraire effects, selection effects, or some other factor?

Popularity prediction hash: 65797d0398c456f539a32c1c156b2d7e69850209f2da544d14c2bb3babae6d6a