Just wrote: "Am I scared for my job?"

blog.separateconcerns.com/2026-04-17-a...

Continuous Adversarial Flow Models We propose continuous adversarial flow models, a type of continuous-time flow model trained with an adversarial objective. Unlike flow matching, which uses a fixed mean-squared-error criterion, our ap...

I am an absolute fan of Peter (Shanchuan) Lin papers. Always super interesting and straight to the point.

This one just came out: arxiv.org/abs/2604.11521

R.I.P.

Séverine et vous avez vraiment fait un travail d'intérêt public pour l'information sur #lezecrans. Ses proches peuvent être fiers.

Bonjour à tous.tes
Nous avons l'indicible douleur de vous faire part du décès de Séverine Erhel, survenu ce lundi 13 avril 2026 à Rennes.
(information diffusée à la demande / avec l'autorisation de la famille)

Albert API donne accès à une large gamme de modèles, des plus légers aux plus puissants, tous hébergés sur des infrastructures françaises : Modèles de chat : Mistral-Small, Mistral-8B, Qwen3-Coder (pour le code).

🤦

Just blogged: "When we have droids"

blog.separateconcerns.com/2026-04-06-d...

Les types ralentissent au début des projets mais sur une grosse base de code ça s'inverse. Comme avec les humains dans un IDE, avoir les types permet à l'agent d'éviter d'aller lire toute la codebase. Surtout si on lui donne l'outil LSP.

If you use the API you can reroll cleanly by changing your user ID. If you use oauth you need to overwrite a constant in the code and it will not survive updates.

Just posted Thoughts & Links 15.

blog.separateconcerns.com/2026-03-28-t...

Controversial opinion: with how good AI has gotten, as an industry it's probably time to stop focusing on Web / application accessibility on the provider end (Sisyphean task) and switch to building great assistive tech on the client end.

Agreed, they're under-estimated. That's why I wrote this one a long time ago: github.com/catwell/fake... - and some people ended up using it in production which resulted in 150k downloads (luarocks.org/modules/catw...) but that's another story 😅

👀 @gawen.me

apenwarr.ca/log/20260316

This is why good teams right now are a few real seniors augmented with AI. Because a defining trait of seniority is knowing exactly where and when you can cut corners and when you need extra trust.

Of course this is not very sustainable...

Il suffit que Musk et SpaceX se penchent dessus pour que tous ces industriels qui refusent de faire du réutilisable se retrouvent à poil !

Jeff Kaplan: World of Warcraft, Overwatch, Blizzard, and Future of Gaming | Lex Fridman Podcast #493 YouTube video by Lex Fridman

@salayanara.bsky.social elle est pour toi celle-là :) 5h10...

www.youtube.com/watch?v=H9rF...

The Flow Is Gone — jd:/dev/blog I used to hold entire systems in my head. Now I hold seven terminals. The trade was worth it, but something real got lost.

Everyone is seeing the same thing at the same time... Here @juldanjou.bsky.social and Stanislas Polu.

julien.danjou.info/blog/the-flo...

spolu.sh/notes/202603...

Benjamin Bayart raconte les débuts d’Internet… en cuisinant un bœuf bourguignon | LBCN - Partie 1 YouTube video by Quentin Adam

Tout ça va beaucoup trop loin à mon avis 😅

www.youtube.com/watch?v=b-YM...

announcing our €3,8M seed round and more on what's next

Well deserved! blog.tangled.org/seed

The future of documentation: redis.antirez.com

Open Source After the Extraction — jd:/dev/blog The old open source deal is dead. What replaces it isn't a fix, it's a transformation. Open source stops being a community and becomes a supply chain.

"Open source isn’t dying. It’s being industrialized." (by @juldanjou.bsky.social)

julien.danjou.info/blog/open-so...

Just blogged: Thoughts & Links 14

Almost 100% agentic coding-related. As it should.

blog.separateconcerns.com/2026-02-24-t...

Yea, GLM 5 is probably better. OpenCode Zen often has very strong models for free, currently GLM 5, Kimi K2 and MiniMax M2.5.

Current free agentic coding setup: pi + GLM-5 on OpenCode Zen. I still use Amp too but the free credits don't last long so I keep it for the harder parts.

What percent wrong is the output of your colleagues? Do you check it more or less thoroughly?

The AI Vampire This was an unusually hard post to write, because it flies in the face of everything else going on.

For those who are already using coding agents 100% of the time or close, this is important: steve-yegge.medium.com/the-ai-vampi...

The "Claude as a pair" way to work is by far my favorite but it is not sustainable because of this. As models and agents improve we'll need to learn to really delegate.

Moi ce que je veux surtout c'est que Anne Le Hénanff retourne dans sa Bretagne pour qu'on puisse enfin avoir Laurent Chemla ministre.

GitHub - catwell/multimodal: A CLI tool to display the output of several Modal jobs. A CLI tool to display the output of several Modal jobs. - catwell/multimodal

Vibe-coded project of the day. Watch logs from modal runs & kill them.

Made with Teal this time, because agents can use languages they don't know much too. I have a real use-case for this, but mostly it was to test terminal.lua.

github.com/catwell/mult...

🐇

C'est la même maison, mais Shiki faisait surtout des plats cuits (très bons :p). Ça fait un moment que je n'y suis pas allé, il faudrait que j'y retourne aussi !

@keltounet.keltia.net Je viens de voir ton update sur blog.keltia.net/la-maison-sa...

Ayant longtemps habité pas loin (rue du Dôme) je suis deçu aussi. J'ai immédiatement regardé ce qu'il advenait de Shiki et je suis rassuré : les cuisiniers de Sanki les ont rejoints shiki-boulogne-billancourt.com