DAOs and Self Driving Companies
If AI is the end of human-led software development, what does it mean for the future of software companies?
The summary of the argument below is as follows:
For decades software generated the majority of returns in the venture investing markets because capturing new opportunities was limited by available developer talent; this netted big developer salaries and created a barrier to entry to newcomers.
Winner-take-all market dynamics prevailed and today there are few viable competitors to established companies like AirBnB, Uber, and Amazon.
These companies have become integral parts of the economic and social landscape and facilitate massive volumes of transactions, yet the underlying software is not particularly innovative: listings, directories, shopping cart checkouts.
As more software development can be automated by AI it becomes possible to operate a company in ‘self-driving mode’ where features are generated automatically on the basis of user feedback and deployed with minimal oversight. Additionally, things like customer QA, dispute resolution, and other customer-facing roles can become increasingly automated as well.
If ownership rights are still only attributable to people, then who owns a self-driving company? The users. They are in position to vote on feature requests, product improvements, service changes, and ultimately pay the costs of delivering the service, which should be 60-80% cheaper than paying for a high gross margin service.
Where is all the money coming from?
For decades venture returns have been driving by the outsized leverage software companies are able bring towards value generation: zero-cost marginal production of goods, geometrically decreasing cost basis, and nearly infinite speed of growth and distribution all enabled by an increasingly mature infrastructure tech stack. In the early days of the dot-com era, companies like Pets.com had to first build and setup their own server racks to handle traffic; today services like AWS have abstracted away most or all of the back-end infrastructure needed to deploy software solutions at scale.
Despite the decreasing barriers to entry software companies still commanded huge premiums and generated massive returns - why? Fundamentally this came down to the arbitrage over developer talent: with a limited pool of skilled software developers labor could both command extremely high premiums and companies that were first to a market could develop a defensible position and earn huge returns. Despite the mature infrastructure stack, talent was still a substantial barrier to entry to deploying and scaling new software solutions. Until now.
LLMs are not perfect coders, but they’re surprisingly good. The common remark among many software developers is that using AI has easily netted a 2-10x productivity improvement in developing code, especially when the features are fairly standard. An LLM isn’t going to come and steal your job as a research engineer developing sophisticated algorithms inside a Wall St trading firm or AI company, but, it sure as hell is going to come for that front-end dev job maintaining the sign-in page and marketing website for a calendar scheduling app or food delivery service. The greater the extent a set of software features has become standardized and replicated many dozens of times over across many companies, the more it has been incorporated into the training corpus of massive LLM models, and therefore the more able they are to replicate this work entirely to the detriment of Jr. software developers.
But developing software is much more than just being a script kitty - you have to think through architecture, logical interdependencies, scalability - in short, you have to use both creative and reasoning abilities. That’s why AI companies who develop AI programmers don’t just train on codebase repositories but also legal documents in order to train these logical and critical thinking abilities. As model architectures, training methods, and suitably chosen training corpus’s develops high quality LLMs from ‘generalized retrieval machines’ to ‘super coders’ the level of engagement of the human user becomes further and further abstracted away from the task of software development and more towards the role of requirements gathering and functional specifications. In short, we all become product managers by the poolside.
DAOs versus Organizations
As LLMs collapse layers of hierarchy starting with the lowest levels of feature implementation and customer support, its reasonable to expect that chains of agents or swarms of agents can start gobbling up successive layers as well: product management, VPs of marketing, directors, etc, considering these roles often consist of translating high level executive decision making by the C-suite into actionable plans, programs, milestones, and deliverables to be brought to fruition by teams of individual contributors. If engineering teams are AI agents, and the entities that manage and direct their workflows are also AI agents, then at some point its agents all the way down. All that’s left is the executive oversight on company direction, high level strategy, and all the things the C-suite does that keeps the evolution of the organization and its products maximally situated to deliver on customer needs and grow into new markets with judicious choices of feature developments, acquisitions, and so on.
Now, consider a situation where the entirety of the organization has been abstracted away except for this critical function: what direction should the company grow into next to best serve its users? Additionally consider there may be a mechanism by which users can directly vote on or advocate for product features and requests to be implemented. A decentralized autonomous organization is exactly such an entity designed to do this - collect votes from members on policy changes, directions, new protocols, and everything in between.
The usual criticism of such a structure is that it misses entirely the ‘founder insight’ that drives the development of an innovative new business. As Henry Ford said, “if I had asked people what they wanted, they would’ve said faster horses.” Therefore such an AI-agent powered DAO company would only really make sense when the specific use-case and market niche for a software service has ossified into a mature and stable form; in other words when the founders initial bold insight has been realized and productized to deliver on that dream to the extent it has become a taken-for-granted background feature of our everyday lives, things like AirBnB, Uber, Amazon, etc. In other words, when the software service has essentially become core infrastructure of modern society.
Software as Infrastructure
Infrastructure is the unsexy, boring, taken-for-granted machinery that makes modern life at scale possible; things like power grids, roads, bridges, airports, train tracks and shipping terminals. Infrastructure services are often operated by, administered or else heavily subsidized and managed by national governments because they often fall into the class of public goods versus private goods, or things best served for the benefit of all citizens by a centralized government because of that governments. Public goods are characterized by being non-excludable, meaning you can’t easily prevent people from using them, and also non-rival, in that one persons use doesn’t directly impede another’s. Roads are a great example in the limit that there’s not too much traffic; private companies that develop roads can and do install toll booths to pay for infrastructure development costs plus their profit margin, but these are often removed after this payback period; or else tolls are maintained to support the ongoing costs of maintenance and repairs due to usage.
Roads financed by toll booths to cover maintenance and repairs are a great example of an infrastructure service where the costs of delivery are directly supported by users in proportion to their use. Making roads extremely profitable does a net disservice to the rest of society because that profit manifests as inflation on the costs of all downstream goods and services, similar to how gas prices increase the cost of groceries at the store. Public goods infrastructure serves the greatest value and greatest number of people when it’s run at or close-to at-cost, and this is only feasible when the managing entity is aligned with the general public interest without a distinct profit motive.
Self Driving Companies and the future of AI DAOs
Where are these trends converging? First, recognize that there are software services today that are starting to approach the nature of ‘infrastructure’ in modern society. Internet communications networks have certainly reached this status as internet is now vitally necessary for massive amounts of economic activity as well as individual lifestyle preferences; social media platforms like Twitter or Facebook might conceivably approach this category; accommodations rental services like AirBnB may as well. But a great example is Uber, the ride-hailing app.
Uber raised quite a large amount of venture capital into order to subsidize early ride prices to capture market share, leaning into the future winds of self-driving cars that would finally turn its economics around and enable it to operate profitably. This hasn’t actually happened; rather Uber product development teams have continued to expand its core offerings into delivery, Uber Eats, Freight, and who knows what else in order to achieve profitability at-scale and with high margins to achieve a positive return to the billions invested in later rounds at high valuations - $13.2 billion total over 27 rounds. Where has that money gone?
Uber has 30,000 employees and about 7 million drivers operating in 70 countries. It’s a massive operation and the number one product Uber provides is the ability to connect with a nearby driver and have them take you someplace. There is plenty of ‘secret sauce’ in the optimization algorithms, routing, and driver and customer rankings no doubt, but the core product is relatively simple and I am always shocked to find out how many employees Uber retains and how large its engineering teams are.
A ride-hailing app is perhaps the ideal example of a company that could feasibly be automated almost entirely in the near-term future and operated as a DAO, operating at break-even without the need to drive high margins because its ultimately owned by the drivers and users in proportion to their engagement and use of the platform. In this situation, drivers and users might earn ‘voting rights’ by engaging and using the service over time; power users have more say over how the product changes or stays the same than occasional users, and the managing entity takes its orders directly through a block chain voting mechanism that then becomes standard operating policy. The software service becomes background infrastructure delivered at-cost to users, which could also lead to higher payments to drivers. Open source Uber could over time eat up market share through lower fees for an equivalent service because it never has to justify a large venture scale return to anyone. The market niche has already been proven out, there is now a large network of drivers willing to join the service, and people have become familiar and comfortable with using ride-hailing apps.
Now, this isn’t to say we could’ve ever gotten to ride-hailing-as-infrastructure without Uber: all the points just mentioned were unproven hypotheses before the long and arduous grind proving that such a concept is feasible, that people would trust strangers to drive them places, that it could achieve scale and adoption in many different markets, and so forth. Given that the hypothesis is now proven its clear what the core product offering needs to look and feel like to be usable and successful. Does it need to deliver food from restaurants, provide luxury tiers of vehicles, deliver packages across town? Almost certainly not. Just like FedEx offers premium services above the more mainstream offerings of the US Post Office, open-source self-driving DAO Uber can just focus on the simplest core offering that captures the largest portion of the market: giving people simple rides in everyday vehicles from point A to B, and do-so in an incredibly unsexy and boring way, but in a way that is far far cheaper and gives users direct control over the product. The only assumption is that AI is good enough to translate a set of voted-upon product features and requests and deliver them into functioning code, something we seem to be on-track for within the next few years.