Found a bug of the "I have no idea how that could have possibly worked" kind:

class arena {
  std::vector<std::unique_ptr<intptr_t[]>&gt; allocs_;
public:
  intptr_t* allocate(size_t n) {
    allocs_.emplace_back(new intptr_t[n]);
    return allocs_.end()-&gt;get();
  }
};

</std::unique_ptr<intptr_t[]>

Notice that instead of returning the last element, the code returns the end() iterator which points to after the last element. To random memory. Oops?..

It's really annoying how many coding-oriented LLM say nothing about the licensing of the datasets they have been trained on. Every now and then I research if I can use fancy AI completion for my open source work and the answer is the same: yolo. The tooling is great - you can spin up a model in like 5 minutes plus download time. The legality of it continues to be ignored my just about everyone.

Pleeease give me a model registry and an option to filter on compatibility with the license of my project. Is that too much to ask?

You know what's the absolutely guaranteed way to make sure I never respond to your job opening email? Name your company "Powerpoint Engineering Ltd."

I had enough slides for a lifetime.

Knowing a bit about how sausages are make (not at Apple specifically), I'm willing to bet that this isn't as much "withholding features" as "getting on the hype train asap, but doing it in a DMA-compliant way would delay the launch by far too long".

Also, it should give you a pause to think what is it about their intelligence features that is not DMA-compliant and why would the rest of the world want that stuff in their pocket.

https://indieweb.social/@stevestreza/112656039226675473

Hmm, do you need to put up a cookie warning if you are using only LocalStorage?

Is it okay to have more dependabot commits in your repo than your own?

I am back at it. Today we will pat that cat.

continuous here: https://m.nevkontakte.com/o/fd3fd5a9c91947c1843aa7cbbcca5dc2

Today's objective is to code something for fun, in a day. It's been far too long I've done something with no purpose other than enjoying the process. An idea has been kicking around my head since (checks timestamps) mid 2020... Time flies

I want to put a cat on the internet. He chills on his page and you can pet him. Or not. He likes when he's getting pets That's it. I have more ideas, but minimalism is the name for the game for today.

Technically, I want to keep things simple and lightweight. It should be clean. It should load fast. Interactions should be simple, no manual required. Giving a pat should leave a footprint of sorts.

To make things a bit more fun, this will be a live development thread

Been thinking about moving off microblog.pub to something else, primarily because it's getting far too slow for what it is. Unfortunately things don't look promising, it seems like there's no good way to migrate existing content to the new account, and replacing the engine in-place is also a source of endless issues. https://voidfox.com/blog/the_sisyphean_effort_of_activitypub_migration/ is a good writeup from about two years ago that covers some of the issues.

This is very annoying. I guess I could pour a bunch of effort into optimizing microblog.pub, but I so don't want to get involved with python.

https://webllm.mlc.ai/ is amazing and terrifying at the same time.

Amazing because you don't need fancy GPU servers to run an LLM, it can just run in a browser. From purely technical point of view this is mind blowing.

Terrifying because before long every other site will download not just 50 MiB of JavaScript and CSS, but also 6 GiB of neural network so they can grace you with another useless AI assistant ready to sell you whatever garbage they are peddling.