AI Programming Languages: From Lisp to Python and C++

Lisp in 1960, Python's library route into AI, and the C++ and CUDA layers underneath, with 2026 survey and 2025 GitHub data.

Guide: AI BasicsLast verified: October 20266 min read

Python is the usual language for writing AI code, but the heavy arithmetic underneath typically runs in C++ or CUDA. In the 2026 annual developer survey, published October 6, 2026, 58.0 percent of respondents had used Python, level with the 57.9 percent of the 2025 survey. GitHub's Octoverse 2025, still the latest edition as of October 27, 2026, ranks Python first among languages in AI-tagged repositories. The path to that position runs from Lisp in 1960 through numerical libraries to deep learning frameworks.

Key takeaways

  • John McCarthy's paper on Lisp appeared in Communications of the ACM in April 1960 and tied the language to symbolic AI research.
  • Guido van Rossum posted Python to USENET in February 1991 as a scripting language, and it gained its AI role through libraries such as NumPy.
  • Deep learning frameworks such as PyTorch expose a Python interface while moving performance-critical parts to C++.
  • JavaScript, SQL, and TypeScript all rank close to Python in the 2026 survey, so AI projects often involve several languages at once.
  • No single language is required for AI. The choice follows libraries, hardware, and team skills.

Lisp and the symbolic era

Early AI research favored languages that manipulate symbols and lists rather than numbers. McCarthy's paper, "Recursive Functions of Symbolic Expressions and Their Computation by Machine," appeared in Communications of the ACM in April 1960 and is the original description of Lisp. The copy hosted at Stanford shows that the idea was to build programs from functions applied to symbolic expressions. For decades afterward, systems that reasoned with rules and logic were commonly written in Lisp or in related languages built for the same purpose. That period is separate from today's statistical methods, which depend on arithmetic over large arrays of numbers.

How Python became the front door

Guido van Rossum began Python during the Christmas holidays of 1989 while working at CWI in the Netherlands, and posted it to USENET in February 1991. The official Python FAQ says the aim was a better scripting tool than C or the Bourne shell, with ideas drawn from the ABC language. Nothing in that history mentions AI.

The connection came through libraries. A 2020 paper in Nature by Charles Harris, K. Jarrod Millman, and 24 co-authors describes NumPy as the main array programming library for Python and calls it the foundation of the scientific Python ecosystem. Deep learning frameworks were then built on the same idea of arrays, now moved onto graphics processors. The PyTorch paper of 2019 argues that a framework can be both easy to use and fast, with every part of a model expressed as an ordinary Python program.

What actually runs where

A Python line such as a matrix multiplication rarely does the arithmetic itself. It calls compiled code. The PyTorch design documentation notes that the project first rewrote its autograd engine in C++, then most operator definitions, and later added TorchScript and a C++ frontend. NVIDIA's CUDA platform, used to program its graphics processors, lists C++, Python, and Fortran among the languages developers can use.

The table below is a simplified picture of one common stack. It is an illustration, since projects differ.

LayerTypical languageExample task
Experiment and glue codePythonLoading data, defining a model, plotting results
Array and framework coreC++ and compiled extensionsGradient computation and operator kernels
Accelerator codeCUDA and similarMatrix operations on a graphics processor
Data accessSQLPulling training records from a database
Product interfaceJavaScript or TypeScriptA web app that calls a model

What the 2025 and 2026 numbers say

The 2026 annual developer survey reports language use for 14,099 respondents. The table sets those figures beside the 2025 survey, where 31,771 people answered the same question.

Language2026 survey2025 survey
JavaScript62.0 percent66.0 percent
SQL58.4 percent58.6 percent
Python58.0 percent57.9 percent
TypeScript43.8 percent43.6 percent
C++24.3 percent23.5 percent
Rust18.0 percent14.8 percent

Python's seven-point jump in the 2025 survey did not repeat, and it did not reverse either. Because the number of respondents fell by more than half between the two editions, small differences in the table should be read as noise. Respondents are self-selected developers, so the figures describe that community and not all developers.

GitHub's Octoverse 2025 reports that TypeScript passed Python in August 2025 as the language with the most monthly contributors, about 2.64 million against roughly 2.6 million. In AI-tagged repositories Python still led with 582,196, up 50.7 percent on the prior year. The report notes that other industry indexes use different methods and may rank languages differently. No Octoverse 2026 had been published as of October 27, 2026, so the 2025 figures remain the newest from GitHub. A language that leads on one measure says little about which is best for a given task.

Languages built to close the gap

Some projects try to remove the split between an easy language and a fast one. On February 14, 2012, Jeff Bezanson, Stefan Karpinski, Viral Shah, and Alan Edelman published a blog post on the reasons for creating Julia, asking for a language with the speed of C and the dynamism of a scripting language, usable interactively and also compiled. The sources reviewed here give no basis for ranking Julia against Python in AI work, and the survey figures above do not break it out.

Choosing a language for a first project

  • Check which libraries the target task already has. A model trained in an existing framework needs that framework's language.
  • Look at where the code will run. Phones, browsers, and servers each favor different tooling.
  • Consider the team. A language the team already reads is cheaper to maintain.
  • Read the license of any library before shipping a product that depends on it.

FAQ

Is Python the only language used for AI?

No. Python leads in AI-tagged GitHub repositories, but C++, CUDA, SQL, JavaScript, and others appear in the same systems, often in different layers.

Was Lisp the first AI language?

McCarthy's 1960 paper is the original Lisp description and the language became widely associated with early symbolic AI. Whether it was the first language used for AI work is not something the sources checked here establish.

Do beginners need to learn C++ to work in AI?

Not to start. Most introductory work uses a Python framework, and C++ matters mainly for people writing framework internals or performance-critical code.

Why did TypeScript pass Python on GitHub?

GitHub's report measures monthly contributors across all projects, and TypeScript grew faster in 2025. The report does not mean that fewer people use Python for AI.

Last verified: October 2026

Sources

  • McCarthy J., Recursive Functions of Symbolic Expressions and Their Computation by Machine, Communications of the ACM, 1960
  • General Python FAQ, Python Software Foundation
  • Harris C. R. et al., Array programming with NumPy, Nature, 2020
  • Paszke A. et al., PyTorch: An Imperative Style, High-Performance Deep Learning Library, NeurIPS, 2019
  • PyTorch design philosophy, PyTorch documentation, 2026
  • About CUDA, NVIDIA Developer, 2026
  • 2026 Developer Survey, Technology, October 2026
  • 2025 Developer Survey, Technology, 2025
  • Octoverse 2025, GitHub, 2025
  • Blog post on creating Julia, Bezanson et al., February 14, 2012

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics