Interview
Bjarne Stroustrup: Deep Learning, Software 2.0, and Fuzzy Programming
- Machine Learning (ML) is characterized as "fuzzy" programming that relies on empirical performance metrics (e.g., 84%–99.8% accuracy) rather than the deterministic reliability, efficiency, and testability inherent to traditional languages like C++.
- The speaker argues against universal accessibility in programming, asserting that critical safety systems (e.g., aircraft and vehicle controls) require the precision of engineered tools and trained professionals rather than general-purpose, "messy" biological or ML-based approaches.
- While ML error rates (approx. 16% false positives) may be acceptable for pre-screening tasks, the speaker deems them insufficient for life-threatening scenarios where 70% or lower accuracy is often the baseline.
- A significant concern regarding ML integration in autonomous vehicles is the hazardous "human-in-the-loop" fallback mechanism, where an unprepared human must resolve AI failures within a critical time window (3–30 seconds).
- The preferred engineering solution for high-stakes environments (autonomous vehicles, nuclear reactors) is systems designed to operate autonomously without human intervention, necessitating the use of reliable languages like C++.
- Current major systems utilize a hybrid architecture where ML and neural networks operate atop foundational C++ frameworks, requiring distinct tools and principles for different application layers.
- The speaker maintains a monitoring stance on ML developments but explicitly refuses to pursue expertise in the field, citing a fundamental incompatibility with their professional focus on precision engineering.