Interview
Bjarne Stroustrup: Deep Learning, Software 2.0, and Fuzzy Programming
- Machine learning systems are characterized as inherently messy, fuzzy, and unreliable, with current performance typically hovering around 84% accuracy rather than higher benchmarks.
- Future progress requires significant effort to achieve 92.5% accuracy, yet empirical measurement will likely remain limited to approximate success rates such as 99.8%.
- The technology remains insufficient for life-threatening scenarios where absolute reliability is required, though it is deemed acceptable and cost-effective in areas where some fuzziness is tolerable.
- Critical control systems for aviation, autonomous vehicles, and nuclear reactors will remain the domain of specifically trained experts rather than being accessible to the general public for programming.
- Autonomous vehicle operations present a significant risk regarding the cutaway interaction, where human operators may face only three to thirty seconds to diagnose AI failures, a transition described as extremely difficult and frightening.
- Personal strategy involves designing C++ based systems for critical applications that operate without requiring human intervention, maintaining these as the preferred standard over AI-dependent controls.
- C++ is expected to continue serving as the foundational layer for all AI applications, with the speaker intending to monitor the field without pursuing personal expertise in machine learning.
- Major systems will continue to be constructed from multiple interacting languages and principles rather than relying on a single programming language.