Ifixit is such a good company. I've now fixed two devices using ifixit tools, replacing parts from ifixit kits. Much $$$ saved. Easy guides and parts. They're well known in geek circles but if you're even remotely handy, and you break your hardware, you really should try a repair

If the lower bound is L, and length is K, and V is an unconstrained simplex L + V*(1 - K*L) Gives a lower bounded simplex. Figured that out today when trying to implement a constraint in a differentiable way in torch.

This is it, what I've trained for

I got @OpenWebUI working with Microsoft's taskweaver using the new open-webui pipelines tooling. Super cool.

I now am running an ollama powered discord bot, ellama for emacs, open-webui, and firefox page assist. Ollama really made making your own personal suite of llm tools so easy to use.

New paper: Adding bounds or known values to Cholesky factors of correlation matrices. Code in @mcmc_stan. It's a short paper but allows some cool stuff 1/

https://t.co/XTX8RY5f9V Linux-friendly configuration tool for the Akai LPD8 Mk2 midi controller. This was a fun 3-evening hackathon. All this, because of I wanted to configure my midi pads as foot pedals for my multi-effect raspberry pi project, and I don't like windows.

Linux driver working. I feel amazing.

TIL: Python 3.10+ has pattern matching ala Haskell/Scala/{other functional languages}.

There we go.

Using Wireshark to listen to Sysex midi signals sent by the official windows/osx editor, so I can implement it for Linux. Run experiments and see what hex code means what.

Haven't done this since I had a Gameboy GameShark. Learning hex codes so I can write my own editor for the Akai LPD8 mk2.

Not enough people sanity check their data, and it shows. It should be second nature for people who have scientific backgrounds, but it's not, and that puzzles me. How can you reason about it without being confident in the data's integrity and adherence to some assumed process?

I've been called sir twice in two days. Shit.

It'd be pretty cool if the admin would take all historical loan payments, and just cancel the interest altogether, retrospectively applying the payment toward the principle.

Scala + Metals-lsp + Emacs = Awesome.

Confluence's search sucks. So I dumped our space out, converted to markdown with pandoc, and got RAG working for them with ollama. Much better.

I just start yelling "agent" until it gives up and connects me to a person. But even today, it connected me to the wrong person (three times), and they had no central operator team to connect me to the correct person at the correct office.

I swear, the worst application of technology still in use today is automated phone directories. There is no quicker way to drain my patience and be annoyed at your entire company than dealing with 15 menus and answer 10 questions narrated by a bot who can't understand me.

Ollama is awesome. Got it integrated into a web UI, phone app, a personal discord bot, and my editor. Even have it talking to me with stt and tts. Getting close to having it call functions and write itself code if it can't answer itself. Local LLMs are actually great tools.