> GPT-2 — "Too Dangerous"_

In February 2019, OpenAI announced GPT-2, a language model with 1.5 billion parameters, ten times the size of its predecessor, and made an unprecedented decision: it would not release the full model. The reason, OpenAI stated, was that GPT-2 was too dangerous. The model could generate remarkably coherent multi-paragraph text on any topic, and OpenAI feared it could be used to mass-produce disinformation, spam, or propaganda. Instead, the organization adopted a "staged release" strategy, first sharing only a small 124-million-parameter version and gradually releasing larger versions over the following months as they studied potential misuse.

The decision ignited a fierce debate that previewed the capability-versus-safety tensions that would come to define the AI field. Critics accused OpenAI of hyping its own technology through fear, calling the decision a publicity stunt. Others argued that the organization was setting a responsible precedent for how powerful AI systems should be released. Some researchers pointed out that any well-funded actor could replicate GPT-2 with publicly available techniques, making the restriction performative rather than protective. The controversy foreshadowed the much larger debates that would erupt around GPT-3 and GPT-4 about who should control access to increasingly powerful AI systems.

Meanwhile, 2019 also saw remarkable demonstrations of AI in competitive gaming. In April, OpenAI Five, a team of five neural networks, defeated the reigning world champions in Dota 2, a complex real-time strategy game with a vast action space, imperfect information, and the need for long-term strategic planning and team coordination. In October, DeepMind's AlphaStar achieved Grandmaster-level play in StarCraft II, ranking above 99.8% of human players. These victories demonstrated that deep reinforcement learning could master domains far more complex than board games, handling real-time decisions, partial observability, and multi-agent coordination.

The year 2019 was a turning point in the social contract between AI developers and the public. The staged release of GPT-2, whatever its motivations, established the principle that AI organizations had some responsibility for anticipating and mitigating the harms their creations might cause. The competitive gaming achievements showed that AI capabilities were advancing on multiple fronts simultaneously. And the growing scale of models, from 117 million to 1.5 billion parameters in just one year, hinted at the explosive scaling that was about to come. The question was no longer whether AI could generate convincing text or master complex games but what would happen when these capabilities grew by another order of magnitude.