In 2014, Vitalik defined DAO as follows:
"an entity that lives on the internet and exists autonomously, but also heavily relies on hiring individuals to perform certain tasks that the automaton itself cannot do…"
"Automation at the center, humans at the edges"
The reason “automation at the center” is so important is closely related to Nick Szabo’s concept of social scalability:
Social scalability is the ability of an institution…to overcome shortcomings in human minds and in the motivating or constraining aspects of said institution that limit who or how many can successfully participate.
In 2020-2022, we saw a wave of DAOs that were not this: they were more like co-ops, with humans at the center, and little to no automation. And like most co-ops they failed to coordinate at scale.
Today we are seeing a wave of AI agents doing things onchain, with varying degrees of automation. Many of these agents are tokenized, meaning they have capital at their disposal, and a network of owners—at least some of which are human—at the edges, who have capabilities the models don’t.
The way I’ve been thinking about the emergence of these “agents” onchain is the true beginning of DAOs 2.0, where agents have the potential put the “A” back in DAO.
With AI automating day-to-day objectives, the role of humans can be minimized to focus on providing capital, nudging guidance, long-term planning, and other resources the models do not have at their disposal.
For DAOs 2.0, the question is: what are models uniquely good at today, that can be fully automated, with minimal human intervention?
LLMs are the most obvious example of AI models that are self-sufficient because they can parse human language and take pre-defined actions without any additional human touchpoints. This unlocks text-based, intent marketplaces like Clanker, an agent on Farcaster that uses natural language to enable token launching.
Diffusion models also require very minimal human input. For example, Botto uses a diffusion model to create 20-30k images a week and another model filters out what it knows are “objectively” bad images. This smaller set of “contender” images is presented to the DAO, which chooses the most marketable or “valuable” piece to auction off as NFTs through simple yes/no voting. The role of the humans at the edges is limited to high-value strategic decision-making. To date, Botto has generated $5M in NFT artwork sales.
What’s an example of something I don’t think current AI models can automate well without significant human touchpoints?
Anything requiring long-term planning, like long-term investing. Long-term investing has a lot of unknown external factors and requires fuzzier data that is difficult to capture or requires significant human labor to wrangle into a digestible format. Compare this with an onchain trading bot — it can use simple deterministic algorithms with information that is readily available to it (e.g., price of an asset) to react to information and execute an immediate strategy.
The broader point is that DAOs 2.0 should track the strengths of AI, and as model capabilities grow, the set of objectives they can accomplish grows, and thus, the size of the DAOs economic influence can grow as many originally hoped. I’m excited by that, and eager to see more breakthroughs in economic activity agents can automate onchain.
Thanks to Daniel Barabander, J.Hackworth, and Kyle Samani for discussion/input.