Nix and AI: Are we there yet?

AI stacks look like the perfect use-case for Nix: Massively multiplying dependency matrices, double-digit-GB OCI images, tedious, Sisyphean build → push → pull → test loops. In ML/AI dev, "It runs" literally means "It runs … on this machine." So … why aren't more ML teams using Nix? This talk is a field guide to the logistics and sociotechnics of what it takes to make Nix happen in AI. Its point of departure is the following question: "Why do we ship what we ship the way we ship it? Either Nix fits the conveyor belt people already ship on, or ML teams learn a new way to build → ship → deploy software. So what will it take to fit Nix to this conveyor belt?

Transcript

1. PLANETNIX 2026 Nix and AI Are We There Yet? The Agony and the Empathy Flox Nix AI @stahnma 1 / 55
2. OPENING AI has every problem Nix was built for Dependency chaos. @stahnma 1 / 55
3. OPENING AI has every problem Nix was built for Lack of reproducability. @stahnma 1 / 55
4. OPENING AI has every problem Nix was built for Huge artifacts and complex setup. @stahnma 1 / 55
5. VISUAL LANGUAGE Perspective key ML engineer Me @stahnma 1 / 55
6. @stahnma 1 / 55
7. OPENING · PARADOX So why aren’t ML teams already on Nix? @stahnma 1 / 55
8. A DAY IN THE LIFE From the ML engineer’s chair One tutorial says Conda. @stahnma 1 / 55
9. A DAY IN THE LIFE From the ML engineer’s chair Another says pip. @stahnma 1 / 55
10. A DAY IN THE LIFE From the ML engineer’s chair Another says build from source. @stahnma 1 / 55
11. A DAY IN THE LIFE · CONTINUED Conflicting setup is normal Linux and Windows flags diverge. @stahnma 1 / 55
12. A DAY IN THE LIFE · CONTINUED Conflicting setup is normal Teammates have different setups. @stahnma 1 / 55
13. A DAY IN THE LIFE · CONTINUED Conflicting setup is normal Both “work,” until handoff time. @stahnma 1 / 55
14. A DAY IN THE LIFE Not a skill issue These are brilliant people. @stahnma 1 / 55
15. A DAY IN THE LIFE Not a skill issue Their domain is models, not plumbing. @stahnma 1 / 55
16. A DAY IN THE LIFE Not a skill issue Tooling sprawl is a response to incentives. @stahnma 1 / 55
17. A DAY IN THE LIFE · CONTINUED I got it working. I’m not touching it. This is rational behavior when your KPI is model quality, not environment architecture. @stahnma 1 / 55
18. THE HANDOFF Every SDLC boundary creates rework Operational teams reverse-engineer what “working” means. @stahnma 1 / 55
19. THE HANDOFF Every SDLC boundary creates rework Two scientists can mean two productionization projects. @stahnma 1 / 55
20. THE HANDOFF Every SDLC boundary creates rework The conveyor belt is usually duct tape. @stahnma 1 / 55
21. THE HANDOFF This is where Nix should shine Lock it. Reproduce it. Remove reverse-engineering from the release path. @stahnma 1 / 55
22. JUST USE NIX Barrier set #1 Where do I start? @stahnma 1 / 55
23. JUST USE NIX Barrier set #1 What is a derivation? @stahnma 1 / 55
24. JUST USE NIX Barrier set #1 Which command do I want: `nix develop` or `nix shell`? @stahnma 1 / 55
25. JUST USE NIX Barrier set #2 Shell/path behavior feels unfamiliar. @stahnma 1 / 55
26. JUST USE NIX Barrier set #2 Tutorials fail in subtle ways. @stahnma 1 / 55
27. JUST USE NIX Barrier set #2 The feedback loop is expensive. @stahnma 1 / 55
28. JUST USE NIX Barrier set #3 Needed package missing in nixpkgs. @stahnma 1 / 55
29. “JUST USE NIX” Barrier set #3 Then pip-in-Nix warnings and “broken guarantees” debates. @stahnma 1 / 55
30. JUST USE NIX Barrier set #3 Builders/substitutes setup appears before model work starts. @stahnma 1 / 55
31. JUST USE NIX This is a UX problem The barrier is not intelligence. The barrier is onboarding and workflow ergonomics. @stahnma 1 / 55
32. GOOD NEWS CUDA GPU setup used to be a major pain point. @stahnma 1 / 55
33. GOOD NEWS CUDA Flox + NVIDIA + NixOS Foundation changed the baseline. @stahnma 1 / 55
34. GOOD NEWS CUDA It proves hard barriers can move. @stahnma 1 / 55
35. CPU TWIST Production inference is often CPU Teams train on GPUs. @stahnma 1 / 55
36. CPU TWIST Production inference is often CPU They frequently scale inference on CPUs for cost. @stahnma 1 / 55
37. CPU TWIST Production inference is often CPU Environment strategy must cleanly cover both targets. @stahnma 1 / 55
38. WHAT WILL IT TAKE? The seam: where Nix ends and Python begins Not “nixify everything.” Not “just Docker.” Use the right layer boundary. @stahnma 1 / 55
39. LAYER OWNERSHIP Nix/Flox owns system and native deps CUDA libraries, PyTorch binaries, C extensions. @stahnma 1 / 55
40. LAYER OWNERSHIP Nix/Flox owns system and native deps This is where reproducibility pain is highest. @stahnma 1 / 55
41. LAYER OWNERSHIP · CONTINUED UV owns pure Python app dependencies Strong lockfiles and traceability for app-level Python. @stahnma 1 / 55
42. LAYER OWNERSHIP · CONTINUED UV owns pure Python app dependencies Together with Nix: provenance plus practical workflow. @stahnma 1 / 55
43. WHY THIS MATTERS Artifact size is not academic Generic PyTorch wheel: ~800 MB. Nix-built target-specific package: ~55 MB. @stahnma 1 / 55
44. PRIORITY 1 Package freshness AI/ML/Python packages must stay current continuously. @stahnma 1 / 55
45. PRIORITY 1 Package freshness Freshness is highly variable @stahnma 1 / 55
46. PRIORITY 2 Better Python-to-Nix bridges uv2nix and dream2nix are promising. (but have been in that state for a while) @stahnma 1 / 55
47. PRIORITY 3 Friendlier onboarding Meet ML teams where they are. @stahnma 1 / 55
48. PRIORITY 3 Friendlier onboarding Reduce shell/path/syntax friction. @stahnma 1 / 55
49. PRIORITY 3 Friendlier onboarding Build paved roads. Improve UX. @stahnma 1 / 55
50. AI CASE STUDY AI may unlock Nix adoption for AI teams 2.5 years ago: 31 iterations for a working Nix package. Yesterday: one prompt, clean package. @stahnma 1 / 55
51. SUMMARY Top 3 improvement recommendations Keep AI/ML/Python packages continuously fresh. Improve Python-to-Nix bridge reliability (uv2nix/dream2nix). Focus on onboarding as a workflow. Meet your users where they are. @stahnma 1 / 55
52. CLOSE Are we there yet? Not yet. @stahnma 1 / 55
53. CLOSE Are we there yet? But the shape of the answer is clear. @stahnma 1 / 55
54. CLOSE Are we there yet? The distance is closing from both directions. Thank you Q&A @stahnma 1 / 55
55. flox.dev @stahnma 1 / 55