Updated
Updated · MIT News · Nov 17
MIT's Jacob Andreas Builds AI Models That Explain Themselves, Targeting 3- to 5-Year Shifts
Updated
Updated · MIT News · Nov 17

MIT's Jacob Andreas Builds AI Models That Explain Themselves, Targeting 3- to 5-Year Shifts

2 articles · Updated · MIT News · Nov 17

Summary

  • Newly tenured MIT associate professor Jacob Andreas is developing methods that let AI models describe what they learned in human language, aiming to make deep-learning systems easier to understand in high-stakes use.
  • His lab studies both how large language models process information and how training data shapes behavior, seeking to predict real-world performance without relying only on external interpretability tools.
  • One technique trains language models to explain other neural networks, assigning natural-language labels to individual processing nodes—for example, a node that detects "patches of grass near animals."
  • Andreas says that work should grow more important as AI agents and scientific-discovery systems become more capable, especially when models are trained on large bodies of raw, poorly understood data.
  • The push comes as AI changes every 3 to 5 years, forcing researchers and educators like Andreas to keep redesigning both tools and courses to match the field's rapid evolution.

Insights

Can an AI truly translate its complex, high-dimensional thoughts into human language without hiding the truth behind fluent lies?
If human minds often fail to explain their own subconscious choices, is demanding perfect internal transparency from AI an impossible standard?