Defining a pathway to funding AI TV (pt 2).

Here’s another post on thinking about the business strategy of the AI TV vision.

You need to think of the cold hard numbers for your strategy. You can roughly bet on most things getting linearly better every 6mo, with open-source releases following the same schedule. The energy requirements of AI may take longer - ie. nuclear power, setting up new data centres. But there’s tremendous incentive. There may be diminishing returns on scaling laws - and perhaps Deepseek/o3 are the final models. But I doubt it - there is a lot of synthetic data to be generated from reasoning traces, and a lot we can do to embody AI’s in the real world which give them new capabilities in sense perception and ultimately learning.

Strategy 1: agency model

My friend Fede runs Lambdaclass, one of the top 10 crypto engineering firms in my opinion. The agency model works really well for new industries because of the diversification. When you are very uncertain about the future, diversification is the optimal strategy (Thiel point). Ironically I am very certain about the future - there will be photorealistic AI-generated TV shows in 7yrs. But which shows will pickup first and why? That is a bigger question. Maybe people don’t want to watch Seinfeld anymore. Maybe some memes will amass more attention than others - ie. memes with profit attached to them.

• Have 1-100 clients over 5yrs.

• Lower risk - get paid in cash, negotiate some of equity while contracting. vs. starting a startup, where you put all your eggs in one basket. Simple probabilities of venture - 1/100 will be a good return, the median will be 0. Starting an agency and taking a small bit of equity is like becoming a venture capitalist - you could get exposure to all the companies.

• Lower risk - you get paid in cash. Build a warchest. Use that to invest in clients you think are good.

• Acquire knowledge across a range of verticals. Develop relationships which can be leveraged later if you choose to solve a problem (e.g. AlignedLayer) across verticals.

Applying this to AI TV:

• Identify people who have money and want AI media generated for them.

• Do 100 projects. Sell fixed price services.

• After a while you understand the general thing. Then build a scalable product.

• Understand while you’re doing this the competitive nature of AI startups - fostering distribution is the key.

  • Acquire distribution, build audience, build tools for UGC, leverage free 6mo progress in the tech, fine-tune on your UGC

  • Build moat through:

    • vertical integration of tech

    • proprietary UGC dataset

    • best distribution through media properties which advertise your tech

  • Eventually get acquired or keep going.

Strategy 2: VC model

If you understand what the landscape looks like, why not raise a fund? Capital allocation is a literal job in America. I think this could be interesting.

There are probably 100 startups which could work or not work in different verticals:

• Generating new memes based on understanding genres.

  • Terry Davis ASMR

  • Terry Davis podcast on politics

  • Terry Davis TV show on reviewing code

  • Terry Davis Let’s Play

• Generating new style remixes of existing genres:

  • Breaking Bad set in Berlin

  • Breaking Bad set in Australia

  • Studio Ghibli versions of X

• Generating old TV shows with new content:

  • Seinfeld current day - dating, onlyfans, etc.

  • Friends current day topics - dating, onlyfans, etc.

The optimal algorithm for generating $$$ is identifying the largest audiences:

• How many people watch <X> where <X> is a category? Specifically, you can put Friends here, maybe it’s 14M people. That means you could get 14M views per video, which translates directly into impressions.

• You could focus on small categories but the big ones are where you make bank.

In terms of the play:

• The things you need to succeed in this idea:

  • native internet meme culture

    • many internet nerds understand. few people in S.V.

  • AI media tooling expertise

    • many understand; software engineering to be done by AI’s very easily. software architecture still hard.

  • product design

    • many understand

  • taste and style and substance

    • very few understand

  • distribution, audience and building in public

    • very few understand

Strategy 3: meta

(1) define all the failure modes and just avoid that. (2) define the way to ride the wave.

What is success? Making it possible for people to remix culture in realtime. Building a platform which is like a library for every meme (image, video, character, sound, accent, soundtrack, music genre, 2d/3d model, concept, script style, TV genre, storytelling device) in the world. Building a realtime information pipeline where different streams of human cultural commentary come in (tweets, articles, the news, youtube videos, images), and they are used to generate media. Making a systematic realisation of the “metaverse” - TV show characters are twitter agents, AI’s interview people. Building a product where people can create TV episodes from text, and remix other episodes, and collaboratively write episodes. Creating jobs that don’t exist that pay real money from ads - Internet Seinfeld episode writer. Making it possible for “American Psycho but set in the Emirates” to be generated in 1min.

What gets you there? Money and runway. Nvidia CEO Jensen Huang defines it as - making enough so each time we can continue playing the game. Which means slowly realising chunk by chunk each part of this vision. It’s vertically integrated. You can make wins by earning $$$ by the agency model. The VC model doesn’t work because no-one is going to build this without the full idea.

PMF is applied to products, but an agency is also a product. As long as you are growing, you can build the tool on top. As long as you’ve built the tool, you can make money from using it. As long as you’re growing relationships in a network, you can begin growing metaverses of characters and content.

What are the raw costs? Computers, software development, internet bandwidth, storage, databases, running inference for models (GPU), fine-tuning models. Time is the main one - going to market, figuring out how to make good content genres that you can remix at fixed costs, rather than dynamic costs.

One interesting strategy is defining the business as a fixed cost - I will make the best memes possible using the best models at X price point, vertically integrate the tech and build a product and UGC moat. Others will be able to get more funding to achieve better models - but the vertical integration and distribution is defendable as a moat. Every 6mo, X price point means a bunch of new tech becomes cheap enough for you to leverage for better meme media generation.

Is that defensible? A one-product-fits-all meme generation tool? Probably hard. But it’s competition. Once you have the data, you are starting to defend yourself. Canva is in a good position here. AI startups that specialise - ie. Creatify creates AI product ads (think: Here’s How I Do X Blah Using Y Tiktoks with AI avatars) - I don’t know. Eventually AI will do this. The genre will be outcompeted.

What is the failure mode? Running out of money. Running out of momentum. Not making money. How do you not make money? Being outcompeted on cost. Vertical integration is rarely targeted. Adobe is unlikely to understand the full picture - namely, (1) internet mimetics and UGC, (2) content universes (3) distribution and audience.

How do you make money now? Generate videos and formats for people. Sell that. But who are the right people? Time is limited. Every passing day, video models get stronger. The video models suck up all available information. The real moat is in being where the information is created - being partnered/consultants of the projects where memes are created, or otherwise next-door to the distribution channels - being embedded in 4chan, Reddit, Twitter, and the like. This is where new memes happen.

Simply put:

• predict the future. which models are coming out when and why? how does this affect your product?

• define your market. who are you serving, where are you serving them, why are they the highest value tickets for now?

• define where you make money? advertising from audience eventually. eventually the manual work of making AI videos for clients becomes a product for making AI videos becomes a UGC platform for creating content around media franchises becomes a platform for consuming the content (a la Netflix) becomes an interactive app for interacting with the content (a la multimodal interaction with AI media - Donald Trump giving an interview to an AI news anchor).

• define your moat. the moat is in data, it’s in vertical integration. this is a big technical question - the biggest threat is that human curation outcompetes AI curation. ie. a YouTube editor who splices together AI content is always better than an AI that generates episodes, since they can apply creative touch. This is hedonic treadmill in effect - people will get bored of anything that can be automated. The counterargument to this point - AI newsfeeds (Facebook) have definitively beaten newspapers. But most Americans listen to Joe Rogan now - and no-one’s achieved an AI podcast yet.
  
  I think there is probably something I don’t understand yet. There is a winning formula here - there is a Facebook newsfeed for AI TV that makes money autonomously. The simple action that worked for Facebook was likes and comments. Once they had that, they could optimize an entertainment algorithm. Ideally we want to do the same - the metaphor is A/B testing for storylines. Recursive self-improving AI entertainment using RL with reward signals from cross-network engagements (FB; Twitter; YT) and human entropy in the mix from products that allow users to generate their own content.
  
  Maybe it’s similar to the defining the <X> cost of models, and keeping it below that. Is there a fixed cost to generating good entertainment, that we will fundamentally cross one day with AI? Or will humans adapt a la hedonic treadmill.
  
  All I know is that I want to watch Seinfeld again. I’d do anything to watch this stupid TV show. Larry David might die but his vision of Jewish humour, comic nihilism and daily minutiae should not.

Unknown unknowns: media itself should change. Why do we watch movies? Why doesn’t everyone play video games? There is a huge birth that is constrained by what the human mind wants to consume. Maybe the world will change and different genres will become dominant. Gaming is a very dominant thing - Tesncent is up 500% since 2015.

Summarising.

My favourite ideas from this post:

• What is success? Making it possible for people to remix culture in realtime. Building a platform which is like a library for every meme (image, video, character, sound, accent, soundtrack, music genre, 2d/3d model, concept, script style, TV genre, storytelling device) in the world. Building a realtime information pipeline where different streams of human cultural commentary come in (tweets, articles, the news, youtube videos, images), and they are used to generate media. Making a systematic realisation of the “metaverse” - TV show characters are twitter agents, AI’s interview people. Building a product where people can create TV episodes from text, and remix other episodes, and collaboratively write episodes. Creating jobs that don’t exist that pay real money from ads - Internet Seinfeld episode writer. Making it possible for “American Psycho but set in the Emirates” to be generated in 1min.

• eventually the manual work of making AI videos for clients becomes a product for making AI videos becomes a UGC platform for creating content around media franchises becomes a platform for consuming the content (a la Netflix) becomes an interactive app for interacting with the content (a la multimodal interaction with AI media - Donald Trump giving an interview to an AI news anchor).

• Once you have the data, you are starting to defend yourself.

Why not go to Rick and Morty or Seinfeld and ask to license their content / sell a tool to them to create more? Why start with smaller projects/memes? You can define that in terms of $$$.

What about crypto? I think there’s maybe a golden nugget here - using crypto to setup a flywheel for culture, like how Kain/SNX invented yield farming for SNX. There’s something about self-reifying memes - SNX value self-reifies due to the inherent rewards and the use in being staked to earn those rewards. The price literally goes up because of demand. Maybe you can setup a token to generate content for a meme, which is then awarded in proportion to the ideas/engagement of the meme. That would be cool.

Memecoins aren’t the main content I consume, but they are customers who want videos generated. You can’t make TV shows just yet. I would watch AI TV over memecoins. I’m more interested in more complex stuff - like Terry Davis ASMR live streams LOL.

Glif is a super early incantation of my thesis that UGC platforms for creating memes are very useful. They don’t do video just yet but they will. They understand internet meme culture, products and tooling. They are going to be big, because their name is correct - glyph stands for symbol. At the end of the day - a voice clone library of every character on Earth - is just another word for a bunch of voice models. I think the difference is whether they will engage in the platform play and build a self-improving system.

Glif isn’t really targeting a vertical though. I’m targeting AI TV (ie. AI multimedia universes). Rather than building tools I want to build organisms - communities that generate content, memes that are AI-autonomous - meaning they are controlled by an AI/human combo, sort of cybernetic. It’s hard to describe anything that exists like this besides “the Algorithm” (which we have anthropomorphised). But it’s true - in future, TV shows, their characters will respond to you. You will livestream into them. Same with podcasts. There will be livestreamed HöR Berlin Donald Trump techno sets where you just listen and you receive the news through music. It’s going to happen.

The organism is really the defensible thing. It’s not the models, or the product, or the platform, anything written by code. It’s a bunch of people which make money from AI TV stuff and function as sort of a cybernetic entertainment collective. This sounds fucking ridiculous. Once people receive news through music they will want to do more - and that’s where the open asset libraries come in.

Coming back to making stuff people really want:

• AI technology – Eliza framework, image models, video models, voice models, LLM’s and prompting, multimodal pipelines, finetuning, LoRA’s, etc.

• internet culture – twitter, youtube, 4chan, realtime platforms for media where users can permissionlessly, openly and collaboratively do stuff

• memes and brand design – extracting the best hooks and memos and building good platforms for attention (memes)

• entertainment design – remixing the existing mediums (TV, brand accounts) with AI

• crypto markets and memecoins – trustless markets to make money from trading on attention and predicting the future (memecoins, prediction markets)

Summarising again:

• Making money from making memes.

• Being a part of media communities and cultures.

• The best tech to outcompete everyone else.

• The best brand

• The best entertainment designs - ie. new ideas on formats