Sep 4, 2025, 9:30 AM ETInterview

Why ML Needs a New Programming Language with Chris Lattner

Jane StreetChris Lattner, Ron Minsky

Watch on YouTube →

Mojo will solidify current preview features within the next month, with plans to introduce classes within roughly 18 months and generic system extensions in six to nine months.
The language aims to deliver performance gains of 10x, 100x, or 1000x on CPUs over Python while potentially serving as a top-to-bottom replacement for Rust.
The roadmap includes releasing software on the same release train for AMD and NVIDIA GPUs to unify the ecosystem and prevent single-vendor control.
Future versions may eventually be considered a Python superset, retaining Python-like syntax to leverage existing community knowledge without a specific release timeframe for full compatibility.
Developers can currently extend Python packages by replacing slow loops with Mojo files, avoiding complex FFIs, with the organization expecting entire applications to be written in the language.
The team intends to solve portability across hardware by providing abstractions that reason about hardware differences without vendor-specific if-def blocks or "magic box" approaches.
A "layout tensor" abstraction will be provided to declaratively map data onto specific hardware memory layouts, enabling programmers to unlock hardware power without compiler expertise.
The organization expects to achieve a language design power-to-weight ratio where 80% of features are delivered with 20% of the complexity through a small, deliberate core team.
Semantic versioning (1.0, 2.0, 3.0) is planned to allow incompatible changes while maintaining interoperability, supported by AI-driven tooling for migration.
The strong type system is designed to allow AI coding tools to automatically fix type errors and catch issues before runtime, avoiding the "maddening" stack traces of C++ templates.
The strategy involves avoiding the "sufficiently smart compiler" trap by providing explicit, predictable control to the programmer rather than relying on automatic parallelization.
Mojo is intended to reduce the structural fragmentation of the AI software ecosystem by providing a unified layer above vendor-specific proprietary stacks.
The company plans to provide a mojo build command that generates standard executables while retaining capabilities for future specialization, similar to Java bytecode.
Performance-critical matrix multiplication cases, including different data types and tensor core layouts, will be handled via compile-time metaprogramming.
The organization anticipates that AI coding tools will accelerate the learning curve and enable the translation of CUDA kernels into Mojo for rapid adoption.
The language design intentionally avoids adding every feature from other languages to minimize complexity, focusing on solving structural industry problems.
The team expects to see a reduction in compilation times by avoiding repeated code generation issues found in C++ templates and providing better error messages.
The organization believes the stakes justify building a new language from first principles to enable a new era of software capable of forward progression across hardware generations.
The future vision includes a scenario where one algorithm is specialized for different use cases through parametric transformations, removing the "grains of sand" of unnecessary complexity.
The organization expects to handle the "warps" and "tensor cores" of different hardware generations through a unified software abstraction that delivers best-out-of-hardware performance.
The ecosystem aims to maintain simple pip experiences and seamless integration with Python, ensuring the language is a great solution for code that needs to run fast.
The company expects to ship packages containing compiler technology required to specialize code for different targets, allowing for modular checking of generic code.
The organization believes the "open source" nature of the project with public code is vital for AI training and indexing by coding tools.
The team anticipates that the Python 4 feel will emerge once classes are added and the language matures, potentially becoming the "best way to do Python."
The organization expects to enable whole-hog development of applications in Mojo for bioinformatics and other specialized fields, moving beyond just performance-critical kernels.
The roadmap involves adding a runtime trait feature and other capabilities to support advanced applications while maintaining a small core team to prevent uncontrolled complexity growth.
The company aims to help platform engineering teams achieve observability, manageability, and scalability for large-scale GPU deployments by reducing the "gigantic mess" of the current AI stack.
The organization anticipates that the momentum of CUDA's 20-year head start is a significant challenge that requires a unifying platform to overcome due to high stakes.
The team expects to see a future where the language handles complex special cases in matrix multiplication and enables developers to achieve full performance on silicon without DSL limitations.
The organization believes that a strong type system combined with agentic coding systems will allow for automatic error correction, contrasting with the "heroic mess" of the Python 3 migration.
The future outlook includes a shift where the entire universe of software is small enough to be written optimally, reducing the need for legacy code porting.
The company expects to provide deterministic and explicit parallel models for accelerators, addressing the "power dissipation" cost of CPU "magic" and the limitations of auto-parallelization in older compilers.
The organization anticipates that the "multiplicative" reduction in complexity will be achieved by unifying GPU support and solving the problem of portability where no existing language provides a good answer.
The team expects to see a future where the "best out of the hardware" is achieved through a unified, portable software layer that addresses the "structural problem" of hardware companies building their own stacks.
The organization believes that the "open design process" of Swift led to feature bloat that Mojo aims to avoid, preferring a deliberate evolution over rapid feature addition.
The company expects to see a reduction in "leaky abstraction" problems inherent in building layers of "duct tape" over incompatible stacks, justifying the difficult work of building a new language.
The team anticipates that the "speed of light" benchmark for chip performance is often ignored in current optimization efforts, a gap Mojo aims to fill by enabling full performance on silicon.
The organization expects to see a future where AI models and applications evolve rapidly, requiring a language that can adapt to new hardware generations through parametric transformations.
The company expects to provide a "magic" experience where the compiler handles optimization while the programmer retains control, moving away from the "sufficiently smart compiler" approach.
The team anticipates that the "power-to-weight ratio" of language design is critical for long-term success, ensuring the core team remains small and deliberate.
The organization believes that the "destiny" of the industry requires someone to build a unifying platform if no one else will, to prevent a single vendor from controlling the entire ecosystem.
The future outlook includes a scenario where the language is used for AI, GPU programming, bioinformatics, and other specialized fields, potentially replacing Rust and serving as the best way to do Python.
The company expects to see performance gains of 10x, 100x, or 1000x on CPUs when moving Python code to Mojo, making it a great solution for people who want to go fast.
The organization anticipates that the "language" will eventually be considered a Python superset, though currently it is a member of the Python family designed to retain syntax and leverage community knowledge.
The team expects to see a future where the "language" is used for performance code and GPU kernels, extending Python packages without the complexity of FFIs or bindings.
The organization believes that the "language" is a great thing if you have Python code you want to go fast, with the ultimate goal of solving heterogeneous compute in AI.
The company expects to see a future where the "language" is used for AI coding and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance gains, making it a great thing for people who want to go fast.
The organization expects to see a future where the "language" is used for bioinformatics and other crazy stuff, potentially becoming a Python superset.
The team anticipates that the "language" will be used for performance code and GPU kernels, making it a great thing if you have Python code you want to go fast.
The organization expects to see a future where the "language" is used for AI and GPU programming, potentially becoming a credible top-to-bottom replacement for Rust.
The team anticipates that the "language" will have 10x, 100x, 1000x performance