GPT-3 demonstrates in-context learning at 175B parameters
OpenAI publishes GPT-3 research showing that a 175B-parameter model can switch tasks from instructions and a few examples in the prompt without updating its weights.
What happened
Event details
OpenAI published Language Models are Few-Shot Learners, introducing GPT-3 at up to 175B parameters. In zero-shot, one-shot and few-shot settings, the model reads text instructions or examples without gradient updates, then attempts translation, question answering, cloze tasks, arithmetic and other ad hoc tasks. This made prompting and in-context learning central interfaces for large language models. The paper did not claim consistent success across tasks: GPT-3 still failed clearly on some datasets and faced limitations involving training-data contamination, web-corpus bias, factual reliability and evaluation design. This event records publication of the research, not product availability. OpenAI did not release the weights; controlled access began with an API private beta two weeks later.
Assessment
Why it matters
GPT-3 combined scale, prompting and in-context learning into a general model paradigm, seeded a major API ecosystem and made reliability and bias central long-term issues.
Availability
Access notes
The paper was public, but model weights were not released; an API private beta began on June 11, 2020.