The Specific Event

Tech journalist Kevin Roose has published a new book documenting the internal history of OpenAI and Anthropic, including what he describes as the "funniest, most bizarre, and just plain weird moments" that shaped the current AI industry. The coverage emphasizes the organizational chaos, interpersonal conflict, and institutional improvisation that characterized both companies during their most consequential periods. This is not a story about technology. It is a story about what happens when organizations try to coordinate around capabilities that nobody fully understands yet, under conditions where the normal mechanisms of competence verification have broken down entirely.

Why Organizational Chaos at AI Labs Is Not Accidental

Roose's account, based on reporting from inside both organizations, describes decisions made under extreme uncertainty, internal factions operating with incompatible beliefs about risk and mission, and leadership structures that shifted rapidly as the stakes escalated. Readers tend to frame this as a story about eccentric personalities. That framing misses the structural explanation. OpenAI and Anthropic were not disorganized because of who was in the room. They were disorganized because they were doing something that classical coordination theory has no framework for handling.

Classical coordination mechanisms, whether markets, hierarchies, or professional networks, all assume what economists call ex-ante competence. The organization knows what it needs, it knows what expertise looks like, and it selects accordingly. Platform environments and AI development laboratories both violate this assumption in the same fundamental way. Nobody inside OpenAI in 2020 could verify with confidence who actually understood what they were building, because the benchmarks for that understanding did not yet exist. What Roose documents as "bizarre moments" are actually the predictable outputs of coordination failure in an environment without established competence signals.

The Awareness-Capability Gap Inside the Lab

One pattern that appears to run through Roose's reporting is the gap between people who were aware that something extraordinary was happening and people who could actually respond to it effectively. This distinction is central to algorithmic literacy research. Kellogg, Valentine, and Christin (2020) documented the same gap in platform labor contexts: workers develop accurate awareness of the systems governing their outcomes without developing the adaptive capacity to navigate those systems well. The gap is not about information. It is about whether awareness translates into actionable schema.

Inside OpenAI and Anthropic, the awareness-capability gap had a different shape, but the same structure. Researchers and executives could articulate what was at stake. What they could not do was coordinate around shared responses, because their mental models of the problem were fundamentally incompatible. Hatano and Inagaki (1986) draw a distinction between routine expertise, which is procedural and context-bound, and adaptive expertise, which operates from principle and transfers across novel situations. The internal conflicts Roose describes look less like personality clashes and more like collisions between people operating with routine expertise in domains that were changing faster than any procedure could track.

What This Tells Us About Institutional Design for Novel Technologies

The Roose book arrives at a moment when both companies are significantly larger and more structurally formalized than they were during the periods he documents. That formalization is often read as maturity. Organizational theory suggests a more cautious interpretation. Rahman (2021) argues that institutional structures around novel technologies tend to encode the assumptions of their founding moment, making the organization brittle in exactly the ways that matter most later. The "invisible cage" Rahman describes is not a physical constraint but a cognitive one: the organization learns to coordinate around its early folk theories, and those theories persist even after the environment has changed.

If Roose's account is accurate, the folk theories that governed early OpenAI, about safety, about scaling, about what the product actually was, were contested from the beginning. The question organizational theory would ask is not whether those conflicts were resolved, but what coordination mechanisms crystallized around them. Structures built on unresolved foundational disagreements tend to produce predictable pathologies downstream: duplicated governance functions, ambiguous decision rights, and what Schor et al. (2020) describe as dependence without integration.

The Larger Pattern

Roose's book is being read as entertainment, which it may well be. But the organizational record it documents is genuinely useful for understanding how institutions form around technologies they do not yet understand. The "bizarre moments" he catalogs are not anomalies. They are data points in a much larger pattern, one that organizational theory has the vocabulary to describe even if the technology industry does not yet have the vocabulary to recognize it. The transfer problem is not just a challenge for platform workers learning new systems. It is a challenge for organizations trying to govern capabilities that outpace the schemas available to interpret them.

References

Hatano, G., and Inagaki, K. (1986). Two courses of expertise. In H. Stevenson, H. Azuma, and K. Hakuta (Eds.), Child development and education in Japan (pp. 262-272). W. H. Freeman.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to soft control. Administrative Science Quarterly, 66(4), 1017-1057.

Schor, J. B., Attwood-Charles, W., Cansoy, M., Ladegaard, I., and Wengronowitz, R. (2020). Dependence and precarity in the platform economy. Theory and Society, 49(5), 833-861.