The Sequence Opinion #504: Does AI Need New Programming Languages?
And some old computer science theories that can become sexy again in the era of AI-first programming languages.
Artificial intelligence (AI) has pushed modern programming languages beyond their original design constraints. Most AI research relies on Python for ease of use, complemented by low-level languages like C++ or CUDA for performance. This dual-language paradigm is a compromise: Python is slow and lacks native parallelism, while C++ offers speed but at the cost of usability. As AI models become more complex and safety-critical, the question arises—are existing languages adequate, or do we need AI-specific programming languages?
This essay explores the limitations of current programming languages in AI development, the potential benefits of AI-first languages with built-in support for differentiable programming, neural networks, and probabilistic constructs, and the importance of formal verification and advanced type systems. The discussion extends to theoretical frameworks such as category theory and dependent types and evaluates emerging AI-focused languages like Julia and Mojo. Finally, we consider how new languages can enhance AI’s scalability, reliability, and interpretability.