Why We Keep Looking for a Single Genius

Every field develops a pantheon. In artificial intelligence it currently consists of a dozen or so names, and the coverage of the field is largely organised around them.

The biographies are genuinely interesting, and the individuals have genuinely mattered. But the pantheon does explanatory work it cannot support.

Pipelines, annotators and the funding nobody praises

• The role of funding agencies and universities that paid for decades of unfashionable research.
• The engineers who built the data pipelines, the chip designs and the training infrastructure.
• The annotators and domain experts whose labour produced the training data.
• The organisational decisions that determined what got built and what did not.

Recruiting, fundraising, hearings

It is not just a media preference. Companies need figureheads for recruiting, fundraising and lobbying. Investors need attributes to evaluate. Governments need people to summon to hearings. A story with a protagonist is easier to sell than a story about a supply chain.

The forecast that keeps being wrong

The cost of the myth is bad forecasting. Someone who believes progress is driven by individual brilliance will expect a breakthrough whenever a talented person changes employer, and will expect a plateau when the talent consolidates. Neither prediction has much support.

The history of the field, read carefully, is a story about infrastructure that took decades to build and institutions that were willing to fund work with no near-term payoff. That story is less quotable. It is more likely to be true.

Work that funding bodies sometimes explicitly rejected

The techniques behind modern deep learning were developed over decades by researchers whose work was frequently unfashionable and sometimes explicitly rejected by funding bodies. The institutions that kept supporting it were, in several cases, public agencies and universities.

That history is awkward for a narrative built around a handful of individuals, because it suggests the decisive input was tolerance for long-term work with no commercial application rather than any particular insight.

Basic research, academic compute, university departments

If progress depends on infrastructure and institutions, then the policy questions worth asking concern funding for basic research, the availability of compute to academics, and the health of university departments. Those receive a fraction of the attention given to personnel changes at large labs.

The coverage follows the narrative rather than the evidence, and the narrative keeps being wrong in the same direction.

A related problem is that the field’s self-image rests on insight rather than accumulation. Attribution to a named individual is satisfying because it implies that the next advance depends on finding the right person. Attribution to institutions is unsatisfying because it implies that the next advance depends on funding decisions nobody wants to make.