OpenAI Co-Founder on Scaling AGI, Team Loyalty, and Crisis

In the summer of 2015, the founding team faced the daunting task of establishing a credible independent research lab in a field dominated by massive tech giants. To break the symmetry of hesitation and commit to a shared direction, the group gathered for an offsite where they drafted a technical plan that would guide their efforts for the next decade. This blueprint rested on three progressive steps: first, solving reinforcement learning; second, solving unsupervised learning; and third, gradually learning more complicated tasks. This early formulation shifted the lab from an abstract mission to a concrete engineering pathway.

By 2017, the organization encountered a fundamental bottleneck in the physical reality of building artificial general intelligence. Mathematical projections of required compute power made it clear that a non-profit fundraising model, which inherently caps capital generation, could not support the acquisition of specialized hardware and massive data centers. The transition to a capped for-profit entity was born out of this physical and mathematical necessity. The decision was driven by the realization that securing exclusive access to advanced computing infrastructure was the only viable path to achieve the core mission.

A pivotal shift in understanding occurred during the training of an unsupervised model on a simple next-character prediction objective. Rather than merely learning grammatical rules and syntax, the network spontaneously developed an internal sentiment neuron that understood whether text was positive or negative. This breakthrough demonstrated that training on a basic observational next-step task can force a neural network to develop deep semantic representations of the world. It proved that comprehension and meaning can emerge as natural side effects of optimization for predictive accuracy.

The development of agentic systems for complex, real-time multiplayer video games challenged prevailing academic assumptions about planning. Traditional reinforcement learning algorithms struggled with a lack of hierarchical planning, suggesting they would quickly hit a wall. However, by continually pushing simple baselines to their limits with massive compute, the systems surpassed top human experts. This revealed that scale paired with simple algorithms is far more potent in practice than complex, hand-engineered architectures. It demonstrated that human-like intuition could be modeled using a network with computational capacity comparable to an insect brain when scaled appropriately.

The apparent dichotomy between predicting the next word and executing first-principles reasoning is resolved through the lens of intelligence as compression. Predicting the next word of an exceptionally complex thinker requires possessing an intellectual depth equivalent to that thinker. True predictive intelligence does not merely replicate known patterns but applies compressed understanding to entirely novel, unseen scenarios. Unsupervised training establishes this deep foundational model of the world, while reinforcement learning refines how the system acts on that model.

The modern architecture of reasoning models relies on a distinct two-stage pipeline using the same underlying predictive technology but with different data structures. The first stage, unsupervised learning, is observational and static, exposing the model to vast amounts of historical data to build background knowledge. The second stage, reinforcement learning, allows the model to learn from its own generated data through interactive trial and error. By evaluating the outcomes of its own actions, the model reinforces successful behaviors, shifting from passive observation to active, agentic problem-solving.

Deploying powerful models involves a choice between closed, secretive development and progressive public releases. The strategy of iterative deployment rejects the idea of a single, high-stakes launch in favor of continuous real-world interaction. By releasing intermediate systems, developers allow society to gradually adapt and build defensive infrastructures. This approach also exposes unpredictable failure modes that theoretical safety testing cannot foresee. For instance, early deployments revealed that the primary misuse of language models was medical spam, an operational reality that would have remained invisible in a laboratory setting.

In an environment where participants genuinely believe they are building technology that will redefine humanity, ordinary organizational dynamics undergo a profound transformation. Normal corporate activities, such as office politics and debates over project credit, cease to be mundane disputes. Instead, they take on an existential weight, as team members view every decision through the lens of societal safety and historical legacy. This intense pressure can either forge high-trust, cohesive teams or splinter them into fragments, reflecting the extreme psychological demands of frontier technology development.

The resilience of a highly technical team is tested when sudden structural shocks threaten the stability of the organization. During a dramatic leadership dispute where the board removed key founders, the subsequent cohesion of the workforce surprised even its leaders. The refusal of the entire staff to accept lucrative competing offers highlighted a cultural focus on peer loyalty rather than financial gain. This collective rebellion demonstrated that technical teams are sustained by shared mission-driven bonds and in-the-trenches leadership rather than traditional corporate structures.

As the global landscape of artificial intelligence intensifies, organizations face choices regarding what details of their models to make public. Hiding the intermediate chain of thought in reasoning models serves both competitive and technical purposes. Economically, it prevents competitors from distilling the model's intelligence by copying its structured thinking. Technically, exposing the chain of thought encourages the model to generate steps that look favorable or desirable to human readers, which degrades the faithfulness of its actual reasoning process. Keeping intermediate steps hidden preserves the integrity of the model's inner logic.

The relationship between humanity and computers is shifting from direct tool usage to agentic collaboration. Historically, humans have contorted their bodies and workflows to fit the constraints of the machine. The upcoming economic phase will reverse this dynamic, as computers begin performing complex multi-step work on behalf of humans. Rather than acting as static software, advanced networks will function as personal agents that understand individual context, negotiate, purchase, and solve long-term goals. This democratizes the role of the builder, enabling anyone with a clear vision to direct an autonomous technical workforce.

True progress in frontier fields requires an explicit willingness to encounter hard truths and accept personal discomfort. Avoiding difficult decisions or dragging feet on organizational misalignment ultimately stalls technical momentum. Leadership in highly uncertain domains requires moving past consensus-driven paralysis to make decisive, sometimes painful choices. This philosophy views suffering not as an unnecessary obstacle, but as an inevitable and productive element of building lasting value.