Compare LLM APIs, GPU clouds, vector databases and LLMOps tools - normalised and re-checked every… - Compare LLM APIs, GPU clouds, vector databases and...

One statically typed source compiles to readable CPython, a standalone `set -euo pipefail` bash script, or a source-less bytecode artifact. The compiler is written mostly in itself, bootstrapped from a 6.4 KB Python stub. Today Externum is launching (Show HN + DevHunt launch week). This post is the honest version of that story: what it does, what it's bad at, and the numbers I'd want to see before trusting any "multi-target compiler" pitch. ## The three backends * **Python target** — emits plain CPython source. Typed source in, debuggable Python out; you can review the emitted file in a normal PR. * **Bash target** — emits a standalone `set -euo pipefail` script. Functions become bash functions with `local` params, recursion works via `$(...)` capture. Where bash has no equivalent (lists/dicts/classes), the compiler warns loudly instead of silently emitting junk. The first version silently produced empty files — which is how I learned that "empty output, exit 0" is the worst failure mode a compiler can have. * **Bytecode target** — a `.exbc` artifact run by a small VM. For "distribute the tool, not the code". A pipeline operator `x |> f(a, b)` desugars to `f(x, a, b)` in every backend, including the VM. ## The honest numbers My first benchmark claim was wrong: I measured the CLI's code-generation mode instead of execution and briefly believed the Python path was 4.5× faster than CPython. Real numbers after fixing the harness: * transpiled-Python path: **~1.6× slower** than hand-written CPython (the cost of the typed multi-target story) * **~11× faster than bash** on the same arithmetic loop * VM: **~30× slower** — it exists for source-less distribution, not speed If a benchmark in a language launch looks too good, assume the harness is measuring the wrong thing until proven otherwise. Mine was. ## Engineering facts * ~380 tests, differential harness in CI: every backend runs the same suite and the outputs must agree — that harness found 3 real VM bugs that unit tests missed * self-hosted: the compiler is written in Externum, bootstrapped from a minimal Python stub * MIT, self-hosted, browser playground, `pip install externum` ## Launch week The code is MIT and free forever. If you want the curated path: the **Pro Pack** (70+ exercises with solutions + a 10-page production guide on choosing between targets) is **$29 instead of $50 this week** with code `LAUNCHWEEK` — link in the repo README. Ask me anything here or on the HN thread — especially about the parts you think are wrong.

Audio variants, addressed by URL

One YAML file. One endpoint. Every MCP server plus the skills and agents you author alongside them

https://wed3.base44.<

Last week, I launched my open source project, Wunjo Community Edition CE, on DevHunt. I’m...

Sync databases and files from your server to your Mac with just two clicks

A launchpad for dev tools, built by developers for developers, open source, and fair.

A launchpad for dev tools, built by developers for developers, open source, and fair.

Keep Your Site Up, Fast, & Secure—24/7