moinsen / Software historyDE ↗
← 2020

Event

GPT-3 und Aufgaben mit wenigen Beispielen

Context and research status
  • Kennung: 2020-gpt3-forschung
  • Zeitraum: 2020
  • Datumsgenauigkeit: Jahr / year
  • Art: Forschungsarbeit / research paper
  • Kategorien: Sprachen und Techniken, Ereignisse, Zahlen und Fakten
  • Akteure und Rollen: Tom B. Brown und Forschungsteam: Autoren; OpenAI: Forschungsorganisation
  • Region: International

The 2020 research paper “Language Models are Few-Shot Learners” presented GPT-3, an autoregressive language model with 175 billion parameters. Among other things, it studied performance on tasks with a few examples in the input; the authors also described limitations. Source

GPT-2, introduced in 2019, is an earlier model series. The API launched as a private beta in 2020; the research publication and product access are distinct events.

In those experiments, task examples were placed in the input text; the researchers did not update model weights through task-specific fine-tuning. The paper also identifies tasks with weak results and potential methodological problems associated with web training data. Original paper

Sources

Offen

Forschungsarbeit, Zugang und spätere ChatGPT-Veröffentlichung getrennt betrachten; Parameter sind kein Genauigkeitsmaß. / Distinguish the research paper, access and later ChatGPT release; parameters are not an accuracy metric.