Ask HN: What are you working on? (February 2025)

Reworking my CSS library for turning semantic HTML into looking like authentic RFC documents. Current version (https://vladde.net/blog/rfc-css) is not quite there yet.

I'm interested in uniform approximation with generalized polynomials -- these are linear combinations from families of parametrized continuous functions over some domain that satisfy some technical conditions, but its also fine to think of them as sums of regular monomials like 1, x, x^2, ..., x^N. This problem has been well understood for real intervals (classical case) for a long time, but I'm interested in this problem where we're approximating functions over complex domains.

There is a theoretically stable algorithm for the classical problem called the Remez exchange algorithm, and an extension to complex domains due to P.T.P. Tang in his 1987 PhD thesis at Berkeley. Theoretically Remez and its complex extension are very stable, but unfortunately implementations my advisor and I are aware of seem to struggle with large degree polynomials, where large is bigger than say n=45 -- errors begin to explode.

In any case, independently of this I've been learning more of the nitty gritty details of deep learning for a project at work (I'm a SWE in my day job, the math is more moonlighting), so to ground my efforts there I've been exploring deep learning approaches to this problem of complex uniform approximation, implementing results from various papers and tweaking things for my use case, and coming up with questions. That's much of what I'm thinking about this week!

Also, I'll be having a half-day long ADHD evaluation session on Friday -- so a bit apprehensive about that.

i think part of the problem is 30mph/48kmph doesnt feel fast - and to get people to drive slower manufacturers need to design cars so 20mph/32kmph feels faster

It is the "systemic/constant/permanent" recording, record-keeping. etc. a.k.a. "processing" (GDPR "processing" means "if it exists and you touch it, your are processing it").

Back in 2005 I remember working with some physical sec company that were setting up cameras in a factory, and they wanted the cameras to 'not record traffic, be activated on if THIS part of the screen has motion')(sidewalk vs sidewalk-right-on-our-doorbell vs road). Also, sudden changes in lighting would trigger it :)

Then you need to have retention period (good luck). Most people use those door-cams are violating GDPR. UNLESS when people complain and take you to court (very very very rare), you can prove that "I auto-delete records after 24h when there is no incident", "I have proximity scanner so it is only 0m-2m from my front door", etc.) (violating GDPR because "hey you pervert why do you record my kids EVERY DAY going in and out")

Privacy and Data protection is very very very difficult with GDPR (and thank you Facebook for messing up back in 2015ish!!!)

You can set up your cam but have the "AI" automatically pixelating all license-plates, and the video recording (if any) should be post-pixelating, and not the original feed. How about you put something with a speed-measuring-sensor (that is NOT a camera), so you only get 'anonymised' data, i.e. "20 moving items", and their speeds. But you will not be able to tell if the 300km/h was done by a bicycle or a Hayabusa ;)

A tool For logistics and container planning and movers:

LLM driven 3d packing written in F#

https://3dpack.ing