Today on Elixir Wizards, hosts Sundi Myint and Charles Suggs catch up with Sean Moriarity, co-creator of the Nx project and author of Machine Learning in Elixir. Sean reflects on his transition from the military to a civilian job building large language models (LLMs) for software. He explains how the Elixir ML landscape has evolved since the rise of ChatGPT, shifting from building native model implementations toward orchestrating best-in-class tools.

We discuss the pragmatics of adding ML to Elixir apps: when to start with out-of-the-box LLMs vs. rolling your own, how to hook into Python-based libraries, and how to tap Elixir’s distributed computing for scalable workloads. Sean closes with advice for developers embarking on Elixir ML projects, from picking motivating use cases to experimenting with domain-specific languages for AI-driven workflows.

Key topics discussed in this episode:

The evolution of the Nx (Numerical Elixir) project and what's new with ML in Elixir
Treating Elixir as an orchestration layer for external ML tools
When to rely on off-the-shelf LLMs vs. custom models
Strategies for integrating Elixir with Python-based ML libraries
Leveraging Elixir’s distributed computing strengths for ML tasks
Starting ML projects with existing data considerations
Synthetic data generation using large language models
Exploring DSLs to streamline AI-powered business logic
Balancing custom frameworks and service-based approaches in production
Pragmatic advice for getting started with ML in Elixir

Links mentioned:

https://hexdocs.pm/nx/intro-to-nx.html
https://pragprog.com/titles/smelixir/machine-learning-in-elixir/
https://magic.dev/
https://smartlogic.io/podcast/elixir-wizards/s10-e10-sean-moriarity-machine-learning-elixir/
Pragmatic Bookshelf: https://pragprog.com/
ONNX Runtime Bindings for Elixir: https://github.com/elixir-nx/ortex
https://github.com/elixir-nx/bumblebee
Silero Voice Activity Detector: https://github.com/snakers4/silero-vad
Paulo Valente Graph Splitting Article: https://dockyard.com/blog/2024/11/06/2024/nx-sharding-update-part-1
Thomas Millar's Twitter https://x.com/thmsmlr
https://github.com/thmsmlr/instructor_ex
https://phoenix.new/
https://tidewave.ai/
https://en.wikipedia.org/wiki/BERT_(language_model)
Talk: PyTorch: Fast Differentiable Dynamic Graphs in Python by Soumith Chintala
https://hexdocs.pm/axon/Axon.html
https://hexdocs.pm/exla/EXLA.html
VLM (Vision Language Models Explained): https://huggingface.co/blog/vlms
https://github.com/ggml-org/llama.cpp
Vector Search in Elixir: https://github.com/elixir-nx/hnswlib
https://www.amplified.ai/
Llama 4 https://mistral.ai/

Mistral Open-Source LLMs: https://mistral.ai/
https://github.com/openai/whisper
Elixir Wizards Season 5: Adopting Elixir https://smartlogic.io/podcast/elixir-wizards/season-five
https://docs.ray.io/en/latest/ray-overview/index.html
https://hexdocs.pm/flame/FLAME.html
https://firecracker-microvm.github.io/
https://fly.io/
https://kubernetes.io/
WireGuard VPNs https://www.wireguard.com/
https://hexdocs.pm/phoenixpubsub/Phoenix.PubSub.html
https://www.manning.com/books/deep-learning-with-python
Code BEAM 2025 Keynote: Designing LLM Native Systems - Sean Moriarity
Ash Framework https://ash-hq.org/
Sean’s Twitter: https://x.com/seanmoriarity
Sean’s Personal Blog: https://seanmoriarity.com/

Erlang Ecosystems Foundation Slack: https://erlef.org/slack-invite/erlef
Elixir Forum https://elixirforum.com/
Sean’s LinkedIn: https://www.linkedin.com/in/sean-m-ba231a149/

Special Guest: Sean Moriarity.

Michael Johnson | Consistency is the key to excellence and Olympic medals