AI Can Beat Everyone One-on-One, but It May Still Lose to a Team That Knows How to Divide Work

On October 29, the DeepMind Institute published “The Economics of Intelligence,” by Alex Imas, Andrew Koh, and Shane Legg. The article is based on the paper “Organizing Intelligence,” and its main line is direct: focusing on a single agent will lead you to misread AI's economic impact.

Most benchmarks measure models solving problems in a vacuum. Economic transformation happens through organizational competition. A model that beats every person one-on-one can still lose to a complementary human team. DeepMind splits this judgment into two lines: capability and coordination.

They had 100 people take a quiz in six domains: math, verbal reasoning, science, patterns, culture, and economics. Performance in one domain helps little in predicting another. There is no replica of “average intelligence” in the population; rather, there are specialists with different weaknesses. The role of organization is to stitch complementary specialists together and fill each other's gaps; the authors call it mosaic intelligence. Complementary pairing completed tasks 32% better than random pairing.

Teams also have a third role: coordinator, i.e., manager. On benchmarks alone, they are the weakest; their coding and writing are only average, but they make everyone more efficient, so their wages are also high.

Figure 1: Human abilities are jagged; organizations use complementary specialists to piece together a larger capability envelope.

AI and humans differ in two key ways. First, replicability. Once a model reaches a certain capability, it can be rented at the price of compute, and supply is no longer fixed. Once AI matches human capabilities, human wages will be anchored to compute cost. Moreover, AI's weaknesses are similar to one another; copying one only increases capacity, not breadth of capability. Second, coordination. What limits the size of human companies is bandwidth, handoff losses, and incentive misalignment; AI has no such natural limits. Copies of the same model can transfer complete work state. But today's agents do not coordinate well.

Thus there are two thresholds. Capability improvements smoothly change wages, prices, and the human-AI mix, while industry structure stays basically unchanged. Only when coordination crosses the supercooperative threshold will there be abrupt changes in concentration, employment, and wages. The authors divide organizational scale into three segments: when coordination is too poor, working alone is better; in the middle is the familiar world with an upper limit; after crossing the threshold, each additional member makes the team more efficient, and the Coasean ceiling is broken. Coase asked why, if organizing inside a company saves costs, not all production is done by one large company; the answer is that human coordination costs have an upper limit.

Figure 2: Capability determines the level of output; coordination determines how large an organization can grow.

The authors divide the economy into deep markets, niche markets, and relational sectors. Deep markets are standardized large markets such as software, logistics, and finance. The four states can be summarized as follows:

• Minimal AGI: structure barely changes; wages change first. For people whose skills are matched by AI, wages move toward compute cost, while specialists and coordinators actually see raises.
• Supercompetent AI: companies are more efficient but smaller; the authors call it “abundance without concentration”; people shift toward judgment and coordination.
• Supercooperative AI: returns to scale go from concave to convex; a few large AI companies appear in deep markets, and wages plummet. Niche markets remain protected for a time by demand limits, but an influx of workers pushes wages down.
• ASI: both thresholds are crossed, and niche markets no longer protect humans. The authors speculate that wealth will flow to relational sectors, where human participation in fields such as healthcare and education is itself the value, and new jobs will emerge that are unimaginable today.

Figure 3: Deep markets, niche markets, and relational sectors follow different paths under the four capability states.

Real-world evidence matches this. OpenAI's Noam Brown talks about running thousands of agents in parallel on open math problems, saying it is very difficult to get these agents to coordinate effectively, and he would not attribute more than 10% of the credit to multi-agent setups. Labor market signals are similar: leaner startups, flatter organizations, softer entry-level hiring, and stable demand for experienced people.

The article gives two conclusions. First, watch coordination; we need benchmarks for agent organizations, testing whether multiple agents can divide work on unfamiliar problems, avoid duplicated labor, cleanly hand off context, and over long cycles detect rather than amplify errors. Second, AI's impact shows up first in wages, earlier than employment and market structure. Looking only at employment rates cannot prove AI has no impact.

Some online commenters mention that “the single agent is the wrong unit” resonates. They initially tuned one terminal after another, then found that what determines the outcome is organization: who assigns tasks, who accepts them, whom to go to when disagreements arise, what to write in handoffs. No matter how smart a single AI is, without this layer it is chaos.

Models win one-on-one and lose at team assignments; this is nothing new. The key variable is coordination. Continue to look at single-item benchmarks, but also look at whether agents can work together.

Original article: https://institute.deepmind.com/essays/the-economics-of-intelligence/

Paper: https://storage.googleapis.com/deepmind-media/papers/organizing-intelligence.pdf