Ask HN: What are you working on? (October 2024)

We're working on a specialised graph compiler for speeding up simulations and computing derivatives automatically (backpropagation/adjoint differentiation). It now supports C++, Python, and C#; AVX2 and AVX512 instruction sets, multithreading; Windows, and Linux.

Essentially, it allows model developers (such as quants in finance, engineers, and ML specialists) to code in Python without needing to think about the performance of repetitive calculations. As we all know, Python is one of the least efficient languages when it comes to complex calculations/simulations - and we help to resolve it. Long story short, with very few tweaks to the code certain types of calculations (such as pricing of derivatives, curve building, computing financial risks, or "small" NNs) can be accelerated by 100+ in Python and x20+ in C++/C#.

We're now looking to add support for Java (but it doesn't have Operator Overloading, so it's tricky), and some customers are asking to support GPU - which is a bit tricky because it's got a closed instruction set.

If there's one thing I've learn over the last few years is that it's incredibly difficult working on a project when you don't have peace of mind. It can be difficult working on something or staying motivated when you're battling depression, or if you're sad, don't have friends, or if anything is troubling your mind.

I've been there and I know what it feels like. I would try to perhaps solve the mental health problem first perhaps before tackling big projects. Maybe try gym, hiking, somehow finding friends, talk to more people, solve your sleep, eat better - something. Because again, if you don't have peace of mind, working on a project is difficult (at least for me it is).

Hope that helps.

I've been working on some projects in Rust relating to image processing and rendering. I'm between a few projects at the moment though but the biggest one is an image processing application I've been working on for quite some time. A lot of stuff I've programmed and learned about over the last 3 years has been leading up to the goal of making something like this haha. I wanted to leverage OpenCL for compute but I had a lot of trouble getting OpenGL OpenCL interoperability to work.

A big motivation for such a project is my passion for photography. I've taken many thousands of photos over just the last 2 years alone. A lot of them are digital, and so far a few dozen rolls of film. A big challenge for me is that I'm not satisfied with the tools available to develop the raw files that are free or open source. Either they're quite finnicky, or they have noticeable issues with color transformations.

I've done a lot of rendering projects over the last few years relating to color that have been focused on getting a better understanding of working with color spaces. Lots of 2D and 3D fractals haha.

Unfortunately I've had quite a turbulent life the last few years so development is very off/on. Every autumn for me seems to be a period of change, this one no different as I'm moving and I'm a bit uncertain of things. However, a side project to all of this has been an OpenGL project where I'm working on things related to voxels! I did a lot of research on data structures like interval trees, octrees, segment trees, etc. It does seem that a lot of people jump for the octree approach, however I've been able to render a lot of voxels with just a hashmap of chunks and a 3d array haha (albeit, mostly initial implementation of chunk generation, single threaded at that!). With this I'm hoping to explore OpenGL compute as I intend of generating world geometry in compute shaders :D

I havent published a project in a while and I'm hoping to get back to putting things out there, so hopefully some of the stuff I've been working on goes well and I can put it up on GitHub or something

An entity database (rest api on postgres so dB is a stretch) with a crud frontend for platform engineering. Adding a similar control loop to kubernetes and bash script (or executables) as the extension language.

This way you can write all of your glue code into one platform and create an IDP at the same time.

This came from a lot of situations where I'd need like one random thing to be stored in a database, but adding that complexity is a bit of a hump for just one thing.

Also crossplane is a really cool idea but starting a kubernetes cluster so that you can manage infra sounds insane to me, even as an avid kubernetes fanboy.

I'm working on ways to allow developers and deployers of LLMs to express how and why their overall system is compliant and adversarially robust, and what to do when that's not the case.

Specifically, my team and I are making assurance cases and ontologies that can seamlessly integrate with the system and its guardrails. For example, if you want to deploy some mix of filters underneath a user-facing LLM app, you would able to: 1) express the logic of how they should be deployed and why (e.g., if X=1, then Y, else Z); 2) see how they perform over time and evaluate alternatives; 3) investigate what happened when an attack succeeds; 4) prove to the auditors that you're taking all measures necessary to be robust and compliant with the EU AI Act.

It started as an informal collab early this year, but we have since published a few workshop papers on this concept [1,2]. We're building a Python demo that would show how it all fits together.

[1] https://arxiv.org/abs/2410.09078 [2] https://arxiv.org/abs/2410.05304