Marco

Your LLM does not care about MCP...

It only sees a list of tool definitions, it doesn’t know or care what’s happening behind the scenes.

The benefit of using MCP here isn’t for the LLM, it’s for you as the developer.

Read my latest blog ⤵️

Every time you share the “3 Bs in blueberry thing” and no one explains its relation to tokenizers, you help everyone get just a tiny bit less informed

"Coral" was probably a bad name, in the context of searching Google for "coral loss" haha.

Oh interesting, this is totally different than regression + bucketing and softmax. I'll take a look, thanks!

On a first pass, it doesn't seem like they directly consider the bucketing case here as an alternative. I wonder why? It seems to work well on my toy problem (even without equal-width).

It is a bounded range of ordinals -- think like rating a movie with 1 to 5 stars, as this is both ordinal (but finite) and categorical.

So if you just do regular regression, your output range is (-∞, ∞) and you can predict things way outside the desired range.

A presentation of our first plenary speaker at JuliaCon Local Paris 2025: Laura Grigori, Full Professor of Applied and Computational Mathematics at EPFL. She will be giving a keynote entitled "Randomization for solving high-dimensional problems: algorithms and software". The conference will take place on October 2nd and 3rd at Conservatoire national des arts et métiers, see the website https://juliacon.org/local/paris2025/ for more information!

Want more open-source software and scientific computing in your life? Get your ticket for JuliaCon Local Paris 2025 before the end of the early bird pricing!
We have an impressive lineup of keynotes, starting with Professor Laura Grigori (EPFL) and not stopping there!
juliacon.org/local/paris2...

Also, this problem is a pedagogical example for someone I'm working with, so I'm purposefully trying to show alternatives/pitfalls/etc.

That is my parallel option. But, since ratings are ordinal, ratings 3 and 4 are more related than 1 and 4, which isn't captured by naive softmax.

I'm also doing label smoothed softmax (i.e., assign nearly 1 to the correct rating, but some mass to it's neighboring buckets).

Surely this has been written about somewhere. What is it called? Claude suggests "equal width ordinal binning" but I don't get many results with that.

Any analysis on how much it helps/hurts?

So anything from (0.5, 1.5] gets mapped to 1. This has width 1, which is 25% of the range of sigmoid. But ratings 0 and 4 are assigned only 12.5% each (like [0, 0.5)).

So instead I could extend the range to like -0.5 to 4.5 so each rating has the same ratio of the sigmoid range.

Suppose I'm predicting integer labels (e.g., ratings) so they are in some sense categorical and ordinal, but I want to do it via regression (predict a scalar, then bucket to a rating).

Say I have ratings 0-4 and I use a NN w/ sigmoid at the end and then multiply by 4 and round to the nearest int.

slash pages A guide to common pages you can add to your website

Oh this really is great: slashpages.net

Is this the official The Golden Ticket manga adaptation?

llama.cpp supports the new gpt-oss model in native MXFP4 format · ggml-org llama.cpp · Discussion #15095 The new gpt-oss model is fully supported in native MXFP4 format across all major ggml backends, including CUDA, Vulkan, Metal and CPU at exceptional performance. This virtually brings the unprecede...

Wow, that was fast.

I guess it says "Recommended" and not "Related" but I still am not quite sure how decided those should be recommended. After expanding the list, the next 3 were all at least from the same journal.

Cannot say I understand the connection.

Caw-league

My compiler, when I debug via

```
cout << "HERE YOYOYOYOYO (1)\n";

cout << "HERE YOYOYOYOYO (2)\n";

...

cout << "HERE YOYOYOYOYO (14)\n";
```

I feel like this joke is a good candidate to be the Ea-nāṣir meme in the year 3025.

Also, I can't tell what the lines are on the woman's face.

Is one her lips and one a crease in the chin? Or maybe it's the top and bottom of her lips? Or maybe her mouth is open (like gasping) but they just didn't draw teeth? Or her lips and a pencil moustache?

My compiler, when I debug via

```
cout << "HERE YOYOYOYOYO (1)\n";

cout << "HERE YOYOYOYOYO (2)\n";

...

cout << "HERE YOYOYOYOYO (14)\n";
```

I pass this sign daily, and I just really like the faces.

NB: It's much more innocent than it seems. It's about an app that helps people find nice spots/things to do/consultation/etc in their neighborhood.

I got absolutely one-shotted by Wingspan like right when I turned 30 and now I have an ever-growing "Birds of X" photo album.

🐦‍⬛🐦‍⬛🐦‍⬛🐦‍⬛🐦‍⬛🐦‍⬛

I ran into exactly this yesterday. Even worse, it isn't just multiple spaces, it's multiple sequential whitespace characters, which can be mixed. In mine, the example was " \r" and those got removed together (but, as you said, not if you specify " " as the delimiter).

Right at the top!

Julia is well represented on that one.

Oh, I hadn't considered that aspect of it. It's a bad situation all around :/

This is my "hiking and beer outfit".

It's also just my hiking outfit, but mountain top beers hit different.

Yeah, this was a very weird decision imo. I guess it was decided a while ago, but surely even then they knew about the NAACL name change debate and the push to go outside the US/similarly well-represented countries (even before the current environment).