In late May 2020, OpenAI published a paper about a new language model called GPT-3. The paper itself was not the kind of thing most people would read, but the demonstrations that followed in the weeks after release made the case more visibly than the academic benchmarks did.
GPT-3 was a much larger version of GPT-2, which OpenAI had released in 2019. The architectural changes were not radical. The training was on more data, the model had more parameters, and the engineering effort to train it at that scale was significant. What made GPT-3 different from previous language models was less about any single breakthrough and more about a quantitative change that produced what felt like a qualitative shift.
Text generated by GPT-3 in 2020 was not just better than what came before. It crossed a threshold where, for many short and medium length tasks, it became hard to immediately tell whether something had been written by a human. Not always. Not for every task. But often enough that the conversations about what could be automated, what counted as creative work, and how to verify the source of written content all had to be revisited.
The release method was unusual. Rather than open source the model, OpenAI made it available through an API with a controlled waitlist. The argument was about safety. GPT-3 could clearly produce convincing misinformation, spam, and impersonations at scale. The waitlist was a way to maintain some control over how the early uses played out.
The waitlist filled fast. Within weeks, examples of what people were doing with the API were appearing across the internet. Code generation from natural language descriptions. Plausible product copy. Conversational interfaces that handled context better than anything that had come before.
What I remember most clearly from that period was the shift in how people who were paying attention talked about language models. The conversation changed from impressive but limited to genuinely uncertain about where the limits actually were. That uncertainty was the most consequential change. The technical capability had existed at smaller scale before. The widespread realisation that scale was producing capability jumps that had not been predicted was new.