Can you revive a dead software project using only AI agents, without writing code yourself?
I tried it on @ApacheMXNet — a deep learning framework retired to the Apache Attic in 2023, last release 1.9, the promised 2.0 that never shipped. There are quite a few reasons to run it in 2026, such as rebuilding the Dive into Deep Learning notebooks. Unfortunately it builds on neither modern GPUs nor Apple Silicon.
No Digg Deeper questions have been answered for this story yet.
The honest answer: yes, mostly. Agents now work independently for hours, even days. Claude's permission system is genuinely excellent. But the judgement still needs supervision — Codex once spent a whole night quietly switching dtypes to fp64 and tanking performance, and Claude once filed one glorious progress report after another, like a struggling engineering manager without ever fixing the bug (Codex solved the problem in 10 minutes).
The full write-up, including why a (software) corpse is the perfect experimental subject and why you shouldn't write new code using MXNet:
👉 https://alex.smola.org/posts/46-mxnet-zombie/
Can you revive a dead software project using only AI agents, without writing code yourself?
I tried it on @ApacheMXNet — a deep learning framework retired to the Apache Attic in 2023, last release 1.9, the promised 2.0 that never shipped. There are quite a few reasons to run it in 2026, such as rebuilding the Dive into Deep Learning notebooks. Unfortunately it builds on neither modern GPUs nor Apple Silicon.