Apple Just Killed Docker Desktop on Mac (Open Source Containers) — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Switch to Apple's Apache 2.0 Swift container runtime on Apple Silicon if you want full OCI compatibility without Docker Desktop's battery drain or commercial seat pricing.
Evaluate Modal's managed private LLM endpoints if you need auditable inference — customers retain access to the underlying code, not just a black-box API.
Self-host Latitude for observability if you want repeated failures collapsed into real issues and plain-English search over production conversations.
Skip LLMs for multi-GPU kernel work — Together AI's parallel kernel bench shows the best zero-shot score was 28/87, and models plateau on rank coordination even with iterative debugging.
Read CMU's free *Modern GPU Programming for ML Systems* book before touching a multi-GPU cluster — it covers data layout swizzling, 3D TMA, and Blackwell programming.
Test your own LLM workflows against Megatron-LM and DeepSpeed workloads from parallel kernel bench before trusting model-generated GPU code in production.
Audit any AI-as-a-service vendor for code and weight visibility before signing a contract, given the video's emphasis on the 'own your stack' shift away from opaque APIs.

Video summary

Companion notes

Apple shipped an Apache 2.0 Linux container runtime in Swift with full OCI compatibility, letting Mac developers drop Docker Desktop and its commercial seat pricing.

Apple's Container Project

Apple's new runtime runs actual Linux containers on Apple Silicon, implemented in Swift and licensed under Apache 2.0. You get the container runtime "without Docker", without the battery drain, and without the commercial seat pricing. The framing in the video: "why rent your local dev environment when you can actually own one."

The 'Own Your Stack' Meta

The creator ties the release to a broader pattern running through the week: developers "clawing back control from black box SaaS tools and commercial licenses" across local containers, private inference, and observability. Modal's new managed private LLM endpoints are cited as another instance — customers "have access to the underlying code," can modify it, and audit it. Latitude is held up as an open-source, self-hostable observability tool that collapses repeated failures into real issues and lets you "search production conversations in plain English."

Where the Wave Stops

The hard limit hits at the lowest level. Together AI's parallel kernel bench (released the same day as the video) tested frontier LLMs against real multi-GPU workloads including Megatron-LM and DeepSpeed. Best zero-shot score: 28 out of 87 correct. Even with iterative debugging loops, models "plateau very quickly" — they can fix syntax errors and basic debug loops, but "reasoning about rank coordination and communication mechanisms" is where they fail. The takeaway: AI can orchestrate containers and endpoints, but kernel-level performance on multi-GPU clusters is still human territory.

What To Read Next

CMU released *Modern GPU Programming for ML Systems* as a free online book. Topics: data layout swizzling, 3D TMA, and Blackwell programming. The creator calls it "the dividing line between people who call APIs … and people who actually build the underlying systems."

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=vR8TMbsv8Qc