Why would one ever hire an agent (AI or Human)?
There must be some advantage the agent brings. What form could these advantages take?
Some advantages include being:
- Faster
- More careful
- More imaginative
- Better logical reasoning
- Lower value of time
- Mastery of tool(s)
- Expert knowledge in a domain
- Access to information databases
- Connections and relationships
- More efficient operations
- Insured
- Legally allowed to practice
We can classify these advantages as intrinsic or extrinsic to the agent.
For agents, an advantage is intrinsic when the advantage is rooted in agent's own mental and physical capabilities. This means that the benefits come directly from the agent's own speed, wisdom, reasoning, or creativity. In contrast, an advantage is extrinsic when it depends on factors outside the agent, such as societal structures, tools, information networks, or legal permissions.
For both human and AI agents, advantages 1 through 4 are intrinsic and advantages 5 through 12 are extrinsic.
What does this intrinsic/extrinsic distinction give us?
Intrinsic advantages are fueled by new innovations in AI. We have new innovations in computer chip architectures, new model architectures, new datasets, synthetic dataset strategies, and sampling strategies. We have increased scale of datasets and of compute. All of these increase the intrinsic advantages of an AI agent, making them faster, more careful, more imaginative, better logical reasoning, and cheaper to run.
Extrinsic advantages are primarily held by the incumbents. The companies that create computer-aided-design software, the molecular-design software, the graphical drawing programs, all have a natural point to integrate AI into their products. The companies that have indexed databases of the web, patents, videos, and photos all have natural information sources to enhance AI. Accountants, lawyers, and doctors can leverage AI to supercharge their work.
In most application areas, the extrinsic advantages can be quite a firm bottleneck. There are only so many satellite imagery data vendors; the class of lawyers and doctors is fairly fixed; and only so many people who can make an intro to and help comprehend the Department of Defense procurement process. These key processes, people and vendors who hold the extrinsic advantages hold a priviledged position that pure innovation in AI cannot supplant.
When we imagine the state of play for AI adoption, we must imagine all these AI innovators in the arena. Unfolding is a battle of the brains, everyone duking it out with crazy new models and ever more impressive demos. Well, above the arena in the stands is the Emperor holding the crown. The Emperor decides which AI technologies (and companies) they bestow with their extrinsic advantages and let through the bottleneck.
There is literally an "LMSYS Chatbot Arena". Companies package their AI innovations into interoprable APIs which do battle; companies hope to be at the top of the leaderboard and get more customers. It's fashionable to talk about 'being in the arena' but surely being the best gladiator is not a business model.
A pure AI innovator, innovating only on intrinsic advantages, is not in a firm position. The cost to switch them out is fairly low. Sure, it's never fun to fire people and onboard someone new. But, it's hardly infrequently done either.
So, those are the pure AI innovators. What about the fast AI appliers? Fast AI applier companies aim to take AI technologies and apply them quickly to a old domain, gaining the first mover advantage, disrupting the old domain with new technology. Fast AI adopters have a crucial challenge. They need to curry favor with whoever holds the extrinsic advantages in that domain in order to get to the starting line. They need to partner with accounting or law firms, they need to buy data, or they need to hire sherpas to guide them through the old domains. They are reliant on the holders of the extrinsic advantages. They cease to be truly disruptve. By capitulating, by needing the blessing of the holders of the extrinsic advantages, they give up their market power.
What does this all means for innovative AI startups? It means that startups need a plan to disrupt the extrinsic advantages held by incumbents... somehow.
How might startups disrupt extrinsic advantages held by incumbents?
Each of the strategies listed undoubtably has its challenges and these seemingly unsurmountable challenges will be described. However, the breakthrough startups will surmount these challenges (somehow).
Building enough momentum to crash through the wall. Startups might think that they can win by moving fast. Blitz through the market via product-led-growth or viral strategies and capture enough of the market to reshift power from the advantages held by the incumbents to the aggreated audience of the startup. The big problems with this are (1) that markets are often too large, (2) that markets are often too viscous with sales cycles that are too long and adoption rates too slow , and (3) that barriers to entry are too low with too many startup competitors fighting to capture the same market.
Bypassing existing structures. The classic example to go to is Uber flaunting taxi laws, sometimes wiggling through loopholes and mostly just pretending that the rules don't exist. Many startups seem to be going down this route: giving medical advice without being a board certified doctor, giving tax advice without being a certified public accountant, or giving specific legal advicing without being a member of the bar. These structures don't seem as easy to bypass as taxi laws. They are backed by a more powerful lobby, with more history, and a more cogent rationale.
Get yourself acquired. Suppose one has a successful company. They've achieved consumer success, but they also know that they don't have the extrinsic advantages needed to replace the existing oligopoly of incumbents. What is their move? Pit the incumbents against each other. Hold an auction to sell the company. Instead of trying to crash through the wall, the company could sell itself off to the highest bidding gatekeeper. Of course, this further entrenches incumbents.
Go even bigger. Do not accept the extrinsic advantages as given. Internalize them. Build your own version of the Multiple Listing Service. Create or buy your own health insurance company. Building your own shipping company. Make your own chips. This is the strategy of the Biggest Tech companies and is an expensive game to play.
Which domains are allow technological-innovation-driven startups to more easily acquire extrinsic advantages?
We consider startups, specifically, to be technological-innovation-driven startups. Startups whose comparative advantage is their technology.
The newer the domain the better. If the extrinsic advantages have been built up since 1847 (American Medical Association) or if the extrinsic advantages have been built up since 1878 (American Bar Association), good luck. If the domain is more recent like 1931 (American Society of Travel Agents) or 1937 (New York City, Haas Act, Taxi Medallions), there's probably a better shot. This decreases the relative importance of extrinsic advantages related to legal priviledges, financial priviledges, connections and relationships which are difficult for a startup to build. All these legal issues and relationship issues significantly increase the barriers to entry for startups.
The problem AI agent technology startups face versus any other technology startups is that most technology startups offer a product which is completely distinct from the existing paradigms. Startups offer products done via applications on the computer and the internet, incumbents do not.
The problem AI agent startups face is that AI agents are trying to move from software applications back to an interaction pattern that looks like interacting with humans. The problem is that while software applications were completely blue-sky opportunities, there are already existing incumbendents for services that look like interacting with humans.
AI agent startups are pincered on two ends in old domains. In old domains, there are the pre-Internet companies that had previously perfected business with human agents. Then there are the Internet companies that have perfect business with software applications. It seems incredibly difficult to thread that needle and carve out a new market. It might be easier to target a domain where there hasn't been serviced with pre-Internet companies with human agents.
The more abstract the domain the better. Go and chess are completely abstract, so these are the best. Protein crystal structure is a good bet. Specific, well-defined domains of mathematics could be a good bet. Coding could be a good bet.
Be careful, though because often the line between the abstract and the real are quite blurry. In vivo protein interactions are quite tricky, mathematics can be quite tricky if we are suppose to construct a completely new domain of mathematics, and coding can be quite tricky when code is to be deployed directly to production.
In more abstract domains, the extrinsic advantage due to mastery over tools is relatively more important; building an extrinsic advantage over mastery over tools is an easier advantage for a startup to build. In more abstract domains, startups can just be smarter and build better tools. In more concrete domains, startups need to wade through the swamp that is the real world.
The more disaggregated, disorganized, and disconnected the customers the better. This customer profile is a perennial strategy for tech companies: see Google, Facebook, Uber, and Airbnb. Most times the startup targets a small, niche subset of a larger disaggregated, disorganized, and disconnected customer base, which is in fact more connected, and leverages it up to capture the larger market. Notice though, that the heterogenous larger market is necessary for the existence of the niche market: there is no niche market in commodities. Also notice that the existence of the heterogenous larger market is necessary for the startup to have growth: having only a niche market means having a small business.
Reframing the focus in terms of agents, startups would have a comparative advantage in creating agents who are hired by individuals or small businesses — not large companies. Existing incumbents have pre-existing relationships with large companies to continue selling their service or to push AI agents through their existing challenges. Startups will have more advantages in fighting incumbents not where they have these relationships, but where they do not have these relationships. They would be better served to define their market to focus on individuals and small business.
The more meritocratic the domain is the better. Another way to say this is that the less brand matters the better. Closely related is that there should be clear criterias for success. The more meritocratic the domain, the more the market consolidates to a single winner, who wins based on quality. The less meritocratic the domain, the more the market is spread out. So, the more meritocratic the domain, the larger the potential payoff.
The problem with aesthetic domains is that they do not have clear criterias for success. People have individual aesthetic styles and there is room in the market for different styles.
One proposed strategy for startups in aesthetic domains is to be better at a meta-level, e.g. be the best at personalizing aesthetic styles. I think this is horribly misguided. Art has been around for a long time. There has always been a "factory" kind of model for art studios — to create more and more art for more and more patrons. A single studio has never scaled to anywhere near the entire market, and certainly has never held anything near a significant share of the market for a significant amount of time. It is misguided because it treats personalization as some kind of meta-level, which is not part of the artistic process.
When someone thinks, wow, I want 5 pieces of art, that someone is discerning enough to tell the difference between five pieces of art generated by Algorithm 1, 2, 3, 4, 5 from Studio A, versus five pieces of art from Studio A, Studio B, Studio C, Studio D, Studio E.