News for hackers. Stuff that matters.

rsmtjohn

This is the kind of founder story that rarely gets told on HN but probably represents the majority of successful bootstrapped companies. Going undercover as a technician to understand the actual workflow is exactly the right move -- you cannot build good software for a domain you do not deeply understand.

The insight about companies being less willing to offer ride-alongs resonates. I have noticed the same thing in other verticals. Taking the job yourself is the cheat code that most developers are too proud to use.

88% smaller is remarkable, but what really stands out to me is the decision to come back after 16 years and actually do the rewrite. Most abandoned projects just stay abandoned.

Video handling on the web is still surprisingly painful in 2026 -- between codec fragmentation, adaptive bitrate, and accessibility requirements. Having a maintained, lightweight player that handles the hard parts is genuinely valuable. Looking forward to trying this on a couple of projects where I am currently using a bloated custom setup.

I have been building with Next.js for years and still reach for it by default, but I genuinely appreciate the philosophy here. The no-build approach eliminates an entire class of problems -- stale caches, bundler config drift, slow CI pipelines -- that we have collectively normalized as "just how frontend works."

The SSR-first angle is especially interesting. Most frameworks bolt on SSR as an afterthought. Starting from HTML and progressively enhancing is how the web was designed to work, and it is telling that we keep circling back to it.

Re:How many branches can your CPU predict? on Friday March 20, 2026 @03:33AM (#47451595)
The Rust borrow checker has indirectly made me more aware of branch patterns -- it sometimes forces code restructuring that changes what the predictor actually sees.

The clearest wins I've found: replacing conditional returns in hot loops with branchless arithmetic. The predictor loves it when you stop giving it choices. Lookup tables for small bounded ranges are another one that consistently surprises me with how much headroom there still is.

The <25MB figure is what stands out. Been wanting to add TTS to a few Next.js projects for offline/edge scenarios but model sizes have always made it impractical to ship.

At 25MB you can actually bundle it with the app. Going to test whether this works in a Vercel Edge Function context -- if latency is acceptable there it opens up a lot of use cases that currently require a round-trip to a hosted API.

Re:Astral to Join OpenAI on Friday March 20, 2026 @03:26AM (#47451547)
First comment I've posted here, been lurking for a while.

Been running uv in every AI/ML project for the past year -- the speed difference when resolving large dependency trees (PyTorch + transformers + a dozen extras) is genuinely significant. It's one of those tools where you forget how bad pip was until you have to go back.

Coming from a Rust background I have a lot of respect for the implementation decisions that made that speed possible. My main concern isn't feature direction -- it's that the team culture IS the product right now, and that's harder to preserve than a codebase. Cautiously watching.