ChatGPT is Fun, but the Future is Fully Autonomous AI for Code

ChatGPT and other AI tools based on Large Language Models (LLMs) have grabbed the headlines for their ability to write poems, short stories and other kinds of content – including code. The intuitive interactive interface makes them easy to use and they can be a real time-saver. LLM-based coding tools have begun to proliferate, with multiple options - like Copilot, Ghostwriter, and Codewhisperer - now available. But they all come with the same major drawback: the code they suggest is also often wrong in ways that can be hard to spot – so they require constant human supervision. But LLMs are not the only way to do AI for code: fully autonomous code-writing is possible by using reinforcement learning to write code that is guaranteed to compile, run and be correct. In this session we'll look at how LLMs and reinforcement learning approaches work, their pros and cons, what they are good at doing – and where they struggle, regardless of how much compute power you throw at them.