OpenAI Replaces GPT-3 With InstructGPT to Better Follow User Instructions
OpenAI launches InstructGPT as the new default GPT-3 model, using human feedback to reduce toxic and nonsensical responses. The smaller, cheaper model outperforms its massive predecessor by actually understanding and executing user prompts.
OpenAI introduces InstructGPT as the new default GPT-3 model, designed to align artificial intelligence responses directly with human intent. The startup creates this updated model to solve the previous version's habit of answering simple questions with bizarre, tangential, or toxic outputs. By using reinforcement learning from human feedback, InstructGPT understands when a user asks a question or gives a command instead of just treating the input as a text prompt to continue.
The practical difference shows up clearly in everyday tasks, such as asking the AI to write a poem about a wise frog. While the older GPT-3 model responds by asking its own writing questions, InstructGPT actually generates the requested poem. Furthermore, this improved accuracy comes at a lower cost, as InstructGPT operates with a smaller set of parameters while still matching or exceeding the capabilities of the much larger 175-billion-parameter GPT-3 model on academic evaluations.
This release builds on the fine-tuning features OpenAI unveils late last year and arrives as the company completely removes its GPT-3 API waitlist. Wider access to the technology already sparks major integrations across the tech industry, with Microsoft leveraging the model for GitHub Copilot and Azure OpenAI Service. Developers now immediately embed these safer, more helpful InstructGPT models into their own applications to create more reliable conversational agents.