LLMs have evolved from simple next-word predictors into systems that follow complex instructions and deliver genuinely useful responses. Powered by instruction-tuning and RLHF, today’s models can write, summarize, and problem-solve far beyond what GPT-3 could do. This blog explains how these advances happened and how you can fine-tune such models yourself.

OpenAI released davinci-003 on November 28th, using reinforcement learning with human feedback (RLHF) for more human-aligned text generation. Unlike davinci-002, it optimizes responses through human-rated rewards. Scale partnered with OpenAI to provide this feedback.