Most of today‘s AI coding assistance is based on the principles of Large Language Models (LLMs) – learning common token sequences from millions of annotated code examples, and then adapting them into the desired context. However, this “observation-only” approach has two problems: First, the reservoir of available code to learn from is limited; second, reasoning from abstract code to concrete executions is not a task LLMs have shown to excel at. In this talk, I sketch how future AI systems will be able to massively experiment with programs, their inputs, and their code to automatically learn how these programs behave, to predict the effects of their inputs and code changes, and in return predict and suggest actions on how to achieve arbitrary effects. Such AI systems will act as artificial program experts, tirelessly accumulating knowledge about the code and its environment, and – in contrast to current AI coders – be perceived as “super coders” that may become way more competent than the most experienced programmers: “Is there an input that bypasses authorization, and which is it?”

Andreas Zeller is faculty at the CISPA Helmholtz Center for Information Security, and professor for Software Engineering at Saarland University. His research on testing and analyzing software has proven highly influential. Andreas is one of the few researchers to have received two ERC Advanced Grants, most recently for his S3 project. He is an ACM Fellow, an IEEE Fellow, an IFIP Fellow, holds an ACM SIGSOFT Outstanding Research Award and an IEEE Harlan D. Mills Award. For details, see Andreas’ Curriculum Vitae.

You can find Andreas

• via his home page at andreas-zeller.info,
• on LinkedIn as andreaszeller,
• on Bluesky as @andreaszeller.bsky.social, and
• on Mastodon as @AndreasZeller@mastodon.social.