Ask HN: What are you working on? (January 2026)

Not particularly inspired by the array languages in general. I do like the idea of a specifics DSL or library that is built from the ground up for ‘vector’ operations (Futhark would certainly be an inspiration here). I do think an (syntax divergences aside) implementation of what is essentially CUDA-like language/library is very possible.

I don’t really have any lofty goals, the language exists to let me have my personal ideal programming language. One of the things I would like to see adopted, by any language, is that the language is able to encode and present provable guarantees about code. So, memory safety, freedom from UB, lack of overflow (integer or stack) is provable in some subset of the language and the proof is done in a logically consistent, verifiable manner.

Of course most of those proofs are not anywhere near feasible for general case code, but the language can restrict allowable constructs in a computation, function, process, or module to a set that makes the desired properties provable by construction.

The language should be capable of unifying and abstracting over SIMD (or other hardware implementations), GPUs, and OS or Userspace multithreading at levels of abstraction from assembly to Haskell or other high level languages.

ScreenRecord.in: https://screenrecord.in I have worked remotely for most of my career and have found screen recording a useful tool to share ideas or ask questions. I liked the UX of Loom but always wondered if one could do it without installing any app. Turns out you can! Given today's browsers especially on desktops one can do a lot of things. ScreenRecord.in can record screen, webcam, mic and system audio if the browser supports them. Chrome has the best compatibility. All recording are stored in local storage I am figuring out what features to add.

I'm working on a continuous chain-of-thought reasoning architecture for generative AI models.

It is similar to Meta's COCONUT. However, instead of the training forward and backwards passes of the reasoning tokens being done serially (slow), they are done in parallel (fast). At inference the reasoning tokens are still decoded serially. The trick is that even though training was done in parallel, the reasoning tokens are still causal and you still get the benefits of increased computational circuit depth.

The kicker is that the architecture is modality agnostic (it doesn't rely on language for its chains of thought), and I want to use it to bring COT reasoning to protein and anti-body generation. Basically, I hope for it to be the OpenAI o1 or DeepSeek R1 for domain-specialized scientific AI.

Doing Math Academy