Scaling Capitalism in a company
I used to work at Luma Labs. I know many researchers at Black Forest Labs, Runway ML, Google DeepMind, and many other diffusion model pretraining startups. Some of the smartest people I know are distributed in all of these. All are very competant at their job. All of these labs have produced amazing models of their own.
Yet when it comes to the actual product and distribution for image/video generation for movie production, the actual winner of the market currently seems to be Higgsfield. A startup which afaik does not train large scale models of its own. By all means it should be a “wrapper” startup and have been doomed to fail from the startup. After all, any one of the labs filled with geniuses can make the same UI as Higgsfield can, and offer its own models at a much cheaper price than Higgsfield can. Higgsfield is forced to offer models at worse than API prices to make its own profit. All the labs can offer its own models at cheaper than API pricing, and the models of all other labs at worse than API prices. This should strictly be better.
And yet, despite all the advantages of the labs, Higgsfield won out regardless. (PS its not because higgsfield offered all the models whereas the labs only theirs. Many labs’ products have third party models as well.)
Got me thinking on why that would be, and how a company could actually try to avoid such a fate. How can a company actually capture all the value it produces, and avoid losing despite successfully accomplishing the hard research work?
The conclusion of my thinking was that this is largely a failure of the companies to scale capitalism.
Incentive Dilution
In early Communist China, a group of farmers were assigned a large plot of land and required to meet state quotas of crop production. Since no single farmer owned any part of the plot, or any surplus from the production, they had no incentives to work particularly hard to go above and beyond and produce any surplus. Nor did they have any strong incentives to put in more effort than the average farmer in the commune, as that just meant you allowed another to slack off. Thus everyone did the bare minimum to avoid punishments and ostracization. The incentive to work hard was severely diluted. Then two decades later, a group of farmers secretly came together and decided to split up the commune land and give each themselves a section, they would manage their own section, and keep the surplus they might produce. They fixed the incentive dilution they faced. Deng xiaoping saw this and decided to embrace it instead of punishing the farmers. And that began the era of Capitalism in China.
What if I told you the incentive structure of most large companies is actually closer to old communist china than any actual capitalist system?
Google feels like one of the worst victims of this, so lets take it as an example. Lets say you join one of the product teams. You get a fat cash offer along with alot of google stock options. Notice that you get google stock options. Not the options of your product team, but of the company itself. The classical answer of how to incentivize employees to improve the company is to give them equity in the company so the incentives are aligned. But notice that when you get google stocks as the 1000000th employee, you get a tiny tiny slice of the pie. It might be worth alot of money in absolute numbers, but as a percentage its going to be tiny. Sure, you want google to do well now. But this is about the same as being a citizen of old China. Sure you want your country to do well, but no matter how hard you work, the amount you improve your whole country might still end up being a tiny percentange. And all of your improvement will be shared across the entire population of one billion. The incentive is so insanely diluted that its easier to just do the bare minimum and let everyone else do the work.
So will you work? Yes ofcourse, you’ll work to not be punished (getting fired from google or getting sent to prison in china), and you may work to earn a big stake in the large pie itself (promotions, equity refreshes etc), or even for pride in your work. But the incentive to actually make your tiny piece of the pie worth much more is very weak.
When equity is used to align incentives, the purpose is meant to be to incentivize the employee to increase the value of that pie itself. (Theres nothing wrong with someone working hard to earn a larger stake of the pie too if they deserve it, but this is more of a zero sum game)
Exploration Collapse
When a large company launches a new product, competing product teams are generally closed down or merged together. Afterall, doing the same thing again and again would be so wasteful right?
Higgsfield also just had one product team, so its not like Luma/Runway/BFL did anything wrong in this regard!
Except its not just Higgsfield which was trying to make a wrapper of all the diffusion models. Its joined by hundreds of startups which were making their own UIs, running their own marketing, trying to figure out which target audience is most willing to pay for its product. Higgsfield just turned out to be the biggest winner (Krea is also another big winner).
When a large company has just a single product team in a competetive space, they’re kneecapping themselves by not allowing themselves to explore at all. They are forced to follow the singular vision of the chosen product lead. If the product lead focuses on the wrong demographic, doesnt build a good UI/UX, etc., then the large company is doomed to fail regardless of the enormous advantages and resources it has.
Micro Subsidiaries
This is the solution I would propose to solve the above two problems, and to thus efficiently scale capitalism in your company.
If I were to start my own diffusion startup, this is how id align the incentives effectively and attempt to capture all the value produced.
Research
As the research produced by the research teams could be used in multiple product lines, and is the core technology behind all the companies profits. The researchers would get equity in the main company. Think of it like the government producing public infrastructure, they help everyone work better and thus earn more taxes.
Product teams
I would have multiple competing product teams, perhaps about 5 of them, perhaps more (depends on how cheap it is to make the “frontend” ie not just the actual UI but also anything that actually productionizes the research produced by the company).
Each product team would perhaps get some much smaller amount of equity in the overall company, but with the majority of the equity in a microsubsidiary made for its team, which gets access to the “public infrastructure” made by its parent company, ie all research artifacts (eg model checkpoints, software, knowledge, etc). The parent company would own a large amount of equity in the subsidiary (think of it like the parent company funding a startup and also giving it alot of other resources and thus getting a large share).
Benefits
Incentives
Thus the incentives are now aligned. Each product team wants its competing version of the product to win. The company wants one or multiple of the products to win. The researchers want to make research that will help the products.
The product team now isnt just fishing for promotions but now actually needs to make their product profitable. If the product would’ve otherwise won simply because of the research team doing well, now with the majority of the product team’s equity being in the product team itself, they can’t sit back and coast. They must make a product of a high calibre!
The researchers making the public infrastructure are also left with a larger chunk of the pie in the parent company since they dont have to share as much with the product teams
Exploration
Since theres multiple product teams competing internally, alot more product ideas can be explored, multiple types of demographics can be targetted.
Talent
When a large company assigns a product lead to make a product, and pours enormous resources into it, it essentially “king makes” the product lead. The product lead may be totally incompetent but may still end up coming out on top against all of its competetitors because of the massive resources they have access to. You dont want this. You want the best product lead to lead your product!
If the company wouldve instead made the research and other infrastructure accessible to multiple of the product leads, they wouldve competed to make the best product and the actual best product and product lead wouldve won. The other teams could get acquired by the winning teams, or declare bankruptcy.
Recursive Nature
Notice that the parent-subsidiary structure is actually recursive. If my startup wanted to make separate image and video generation teams, it totally could! The researchers are incentivized to make its respective technologies better (and an overall diffusion team could enable information sharing) and to enable its product teams to make better products. Coasting on the work of others could be minimized.
Investing
When a venture capital firm invests in a company, they have rather limited information. A pitch deck, a few hours of conversations, and you might get millions. YCombinator is the extreme end of it. A 15 minute talk and you get $500k.
A company knows its employees orders of magnitude better. And when a founder material employee leaves a company its an enormous loss for the company. Even if offered promotions the employee might not stay, as its not feasible to give them a large enough equity stake in the parent company for them to feel ownership!
VCs already try to provide as much as they can apart from just money, connections, cloud credits, etc. But thats really the most they can do, afterall, theyre just investors!
A company could provide so much more, actual employees which the parent companies pays the salaries for, so much infrastructure, research, credibility, legal support, etc etc. And thats ontop of cash investment. It just makes so much more sense for a company to let a founder build a product with real ownership as a subsidiary. Yet it barely ever happens!
RenDanHeYi
The one exception would be the chinese appliance manufacturing company Haier. It has a structure very close to what I just described, they call it “Micro-Enterprises” instead of “Micro-Subsidiaries” but it sounds pretty similar. Unfortuntely I havent looked into it too much and only have a cursory knowledge of it. I will likely try to find and read books about it, and update this blog as I find out more.