Pavel AI

The example is extremely impressive. But I should ask why on Earth it took so long? Developers have been trying to customize default controls for two decades...

Anyway, nice demo @nerdy.dev

Bridges are the reason why I want #atproto to have ability to block accounts by domain name, not by their handle

That's a piece of coding art!

Need more arrows! #js #dev

Github?

One more custom <select> demo, with a bunch of CSS transforms and animations.
I can't get enough of this. So cool.

#AI #dev Structured outputs are waste of context space, if you're using LLM with OpenAI's Harmony output

💡 Just use noop tools to retrieve information. Harmony's tool description requires less tokens, than JSONSchema

cookbook.openai.com/articles/ope...

Reverse Engineering a $500M Mystery: From HashHop to Memory-Augmented Language Models A Blog post by Asankhaya Sharma on Hugging Face

This context retrieval trick addresses context related problems and could boost models productivity. It's being tested on a relatively small 1.8B QwenCoder 2.5, and looks very promising to me

huggingface.co/blog/codelio...

I have several thoughts on this. But it would be nice to:

* See how many documents were searched at all
* See the whole row's data at once, like a card view
* Make title columns to look like a link or a button
* The search field should be closer to the output, just right above the data-view

Burn logo with signature "Accelerate Progress with Burn"

Can Rust beat Python and become the main language of ML as it has became the second main language for the web?

Burn (Deep Learning Framework written in Rust): github.com/tracel-ai/burn

Doesn't it mean you need to install go first?

There would be more features. But all modern programming languages would become fundamentally obsolete. Rust seems to be the most prominent one of golden era of programming, but it's time to move to the next frontier

Well done 👍

On-Policy Distillation On-policy, dense supervision is a useful tool for distillation

And if the learning process would be enhanced with on-policy distillation the overall progress would be significant

thinkingmachines.ai/blog/on-poli...

@hardmaru I believe your work shows something bigger. If the model would be taught with special learning-time data-markers (e.g. tokens) describing the data, it would speed up the learning and make the outcome model stronger

I *need* 20 🤷‍♂️

The demo requires better UX, but the speed surprised me (and probably you!

#ai

Does Mistral have Bsky/AtProto account?

Depends on what usefulness means for you. I use it mostly for learning Rust faster or to write some SQL, JSONSchemas, etc. so to me it works pretty well. I don't overload context with complex tasks, keeping things simple. Roughly the success rate varying from 80% to 100% usually

Here is the research details:
bsky.app/profile/hard...

Well, it was published the day before, the comment itself is about their research where agents compete by writing code, but this is about same sort of things

bsky.app/profile/hard...

Ollama Get up and running with large language models.

It's super easy to start with Ollama.com, it let's everyone try many of open-weight models locally (or from the cloud). This is like Docker but for local models

While I appreciate their work, it should be done with a single specification file and a browser engineers work. This problem doesn't worth to build a full-fledged business around it

I believe in 2025 Tailwind should not exist in the way it is. And what's really sad is that we are stuck with outdated technological stack, which requires a sophisticated solutions just to use CSS without headache. On demand classes solve most of problems which Tailwind is solving now

Yes, it's for websites, but I think, it doesn't make much of a difference for repositories. My suggestion was mostly a proposal, maybe it worth to be a separate initiative to make llm-related in-repo data standard

Isn't it what llms.txt file by llmstxt.org is made for? IMO It should have some neutral non-proprietary name and be stored in a subdirectory like .llm/readme.md where you can put more llm-related things

I hope they will change their mind. Because even with the support of Gumroad and other big companies they didn't solve the core issue

And they could become a very good AI enhanced/accompanying company, like Vercel or Bun. But they seem to be against it and they refuse to move the progress, benefitting from the obsoleteness of the technology

So the core of their business is built around obsolete solution. What they (Tailwind team) should had do with all their reputation and knowledge is to push CSS WG to implement on-demand classes. It would be an end of Tailwind toolset, but a big step forward for the Web

Their business was very good protected, because people who work in browser development are somewhat against on-demand classes yet. Example

bsky.app/profile/bram...