DeepMind co-founder Shane Legg and others publish "The Economics of Intelligence," projecting organizational evolution from AGI to ASI

0 layoffs, expert salaries cut by 30% first! DeepMind co-founder simulates the first cut of AI New Zhiyuan October 31, 2026 09:20:46 From Beijing

New Zhiyuan report

Just now, Google DeepMind co-founder Shane Legg made his move.

He is DeepMind's Chief AGI Scientist.

This time, he teamed up with AGI Economics Director Alex Imas and Research Scientist Andrew Koh to publish a long article, "The Economics of Intelligence," at the DeepMind Institute, alongside the working paper "Organizing Intelligence."

Imas revealed on X that he joined DeepMind specifically to build an AGI economics research program.

They want to answer a big question: from AGI all the way to ASI, when and in what way will the economy be completely rewritten?

From the human baseline to ASI, how the organizational composition of humans and AI evolves across three types of markets

The paper uses an economic model to calculate an account that sends a chill down everyone's spine:

Instead, the "coordinator," who ranked last on individual tests, pulled off an upset and became the highest-paid of all.

And the even more distant endgame is the "corporate singularity."

Google DeepMind evaluation scholar Sol Messing remarked that this paper has rewritten the direction of the industry's questions.

Your job is still there, but your bargaining power is gone.

While you're still grinding performance reviews with colleagues, your craft has already been repriced by the electricity bill of the data center.

But placed in the real economic world, the splash it makes is pitifully small: employment data barely budges.

To prove this, the team recruited 100 subjects and ran a test across six domains: mathematics, verbal reasoning, science, pattern recognition, culture, and economics.

The result: a person's score in one domain can hardly predict his performance in another.

In reality there is no well-rounded "average smart person" who is balanced across subjects; everyone is a lopsided "jagged specialist."

Radar charts of subject imbalances for 6 of the 100 subjects across six domains, and the mosaic intelligence envelope assembled by complementary teams

It's like putting together a mosaic: every piece is irregular, yet together they fit perfectly.

What companies need to do is piece together lopsided people, dividing labor according to each one's comparative advantage. Strengths cover weaknesses, and the team can assemble a capability map far beyond any individual.

The paper calls this "mosaic intelligence."

Measured data shows that two-person teams paired by complementary abilities performed 32% better than randomly paired teams.

Large models lack exactly this.

Copy the same model a thousand times, and the directions of their lopsidedness and blind spots overlap completely: adding another copy only adds capacity, not coverage of weaknesses.

Without AI, programmers and writers team up 1:1, and specialist wages rise to 4.15, nearly three times that of working alone.

Human experts are valuable because each person has only 24 hours a day and cannot be copied.

The paper lays down an iron law:

As long as a model can match a human capability, that wage will be tethered to compute costs.

The great role reversal plays out.

Programmers and writers lose their price advantage in the large market and are squeezed into a narrow niche market: wages fall from 2.90 all the way to 0.39, more than 90% below their peak.

Instead, those 2 bottom-ranked coordinators are eagerly sought after by AI companies.

Employment in entry-level positions in affected occupations falls by 7.7% to 19%, and starting salaries drop by 4.2%.

Since internal division of labor can save costs, why doesn't the whole world merge into one giant company?

For nearly a century afterward, the answers economists gave all pointed to human limits.

Once a team expands to dozens or hundreds of people, attention runs short.

Word passes through several hands, and by the end it's all distorted. Besides, everyone is calculating their own interests.

With more internal friction, the efficiency saved by division of labor is eaten up, and the company can no longer grow.

In mainstream workflow benchmarks, multi-agent systems misalign goals, lose context at handoffs, and break verification, with task failure rates as high as 41% to 87%.

In sequential planning tasks that require continuity, adding a few more agents makes performance plummet by 39% to 70%.

A swarm of agents lacking coordination can also amplify small upstream errors by up to 17.2 times.

OpenAI researcher Noam Brown also admitted that getting a large number of agents to tackle math problems in parallel makes it hard for them to cooperate effectively.

As for the credit due to multi-agent collaboration, he thinks it doesn't even amount to 10%.

Programmers and writers complementing each other to raise the output level (left) and the three-stage curve of the organizational scale law (right)

To quantify this shackle, the team, imitating the Scaling Law of large models, put forward the "organizational scale law" for firms.

In one sentence: individual capability determines how much each person can do, and collaboration capability determines how large an organization can grow.

When collaboration capability is below 1, the larger the team, the more per-capita output first rises and then falls.

In the paper's model, the optimal size of a pure AI company is only about 7 copies.

Once collaboration capability crosses 1, each additional member makes per-capita output higher.

The Coasean ceiling is thus broken.

In reality, the exponent by which AI team output grows with headcount is only 0.48 to 0.68.

It is still far from 1.

Sol Messing points out that the weights of frontier models are frozen and cannot keep learning on the job.

The paper uses two unemployment curves to show two completely different modes of replacement.

The first is gradual replacement within firms.

With AI collaboration capability fixed at 0.96, individual capability slowly rises from 3 to 3.4;

The moment the collaboration score climbs to 0.8521, the cost of a pure AI company breaks through the floor of a human company.

This time, even the coordinators are not spared.

In the super-cooperation scenario, about half of the coordinators lose their jobs. At the ASI stage, all 2 coordinators lose their jobs.

You think working harder than the colleague at the next desk will let you stay.

Of course, the authors also leave some room.

"Super-cooperation" is a highly simplified mathematical property.

The evolution of organizational composition across the four major intervals—Minimal AGI, Supercompetent, Supercooperative, and ASI—in two types of markets.

When AI crosses both the "supercompetent" and "supercooperative" thresholds at the same time, it enters the ASI interval, that is, artificial superintelligence.

Most goods and services will become extremely cheap.

In reproducible production links, human wages will be anchored to compute costs.

Where will the money flow?

One possibility the authors offer is "relational industries."

Medical care, education, offline experiences—human participation itself is part of the value.

Just as in 1950, no one imagined that so many people would later make their living as "software engineers."

In his tweet, Imas condensed the entire paper into two sentences.

First, to see clearly how AI will rewrite the economy, you must keep a close eye on the collaboration capability among agents.

Only when thousands upon thousands of agents learn to divide labor and hand off seamlessly will the corporate boundaries of the old world truly dissolve.