Do you use vitest in ui mode?

npx vitest --ui

#vitest #nextjs

Now that it does, I'm going to implement the actual caching part. And then... figure out how and what to measure to see how it's behaving.

I added Multi-head Latent Attention, one of Deepseek's optimizations, to Nanochat, the tiny educational LLM project by Karpathy. I did a couple of runs with and without MLA just to see if the loss behaves somewhat reasonable.

I wonder if I should record a video for non-Rails developers about why the experience with the ORM, database, and tests in Rails is so good.

I want my editor to behave in a specific non consistent way:

When I open files, I want two files, side by side.

When I open diffs, I want them to take the whole space.

Anyone else?

Instead, we should be teaching how to manage deb, how to manage and tracking spending money that you don't have, how to calculate when it is and isn't a good idea.

Because there will be new products that allow people to go into massive debts.

The problem with teaching "credit cards are dangerous" is that then a new product appears, like buy-now-pay-later and catches a whole generation by surprise, with many people going into debt.

I'm playing with Prisma and it's yet another ORM that pretty much ignores the need for testing. I wish more ORMs learned from ActiveRecord and the Rails testing framework about what a good devex is in this category.

GPT-5.2 seems to be the only that implemented the compression of queries.

When it comes to the actual neural network, GPT is the hardest to follow, which may be in part because it implemented more, but also it used short style variable names and very few comments.

I'm playing around with implementing Multi-head Latent Attention after reading the DeepSeek-V2 paper. My least favorite implementation is Gemini's where the size of the qk head is the addition of the rope part and the nope part _sigh_.

#deeplearning #llm #ai

Juggling is not like riding a bicycle. You can actually forget it. TIL

Did you know that the engine behind Chrome came from a browser that almost nobody knows about, Konqueror, which was part of KDE, a desktop environment for Linux.

I feel that those developers rarely get the credit they deserve.

I find variable naming in the world of machine learning to be very hard to follow. It's like they are all named nk_nmr_bdrfoo. And you have to remember this is the same as nmr_bdr_n in another part of the program, completely arbitrarily.

Anyone else?

#deeplearning #ml

My plan for this weekend is to read and understand all the code for nanochat, the small LLM that Andrej Karpathy created for educational purposes. Wish me luck! #AI

API keys, in most cases, should belong to organizations, not users.

It's cold... do I turn on the heating?

Nah... I start another LLM training run. That'll warm up my office.

How to scare a Cursor user: You are projected to reach your usage limit...

#vibecoding #ai #llm

The feature that I want the most from Slack is being able to move a message to be in the thread started by the message above. Am I the only one?

Make it possible and easy for engineers to deploy micro-tools complying with your security.

At Google this was the reason we had an explosion of internal tools, and some of those were good and became very impactful. But if it's hard, engineers will just not do it.

Come on local models! Get good enough to give me an excuse to buy a 5090!

#ai #llm