FrontierThe story, in brief
WebGPT: Improving the factual accuracy of language models through web browsing
OpenAI just solved one of GPT-3's biggest problems: making stuff up. Here's how.
The KeyNews take
Why it matters
WebGPT demonstrates a critical capability advancement—fine-tuning GPT-3 to reduce hallucinations through web access—that shifts how enterprises think about LLM reliability for factual tasks. This was a watershed moment for grounding language models in real-time information.
The key facts
5 to knowFine-tuned GPT-3 with text-based web browser integration
Focus on open-ended question answering with improved factual accuracy
Addresses hallucination problem in language models
Published December 16, 2021
Foundational work on retrieval-augmented generation (RAG) before it became industry standard
Go to the source
OpenAI Blogopenai.com
Publisher excerpt: We’ve fine-tuned GPT-3 to more accurately answer open-ended questions using a text-based web browser.