Blog
DeepSeek shipped V4.1 Flash a few weeks after GLM 5.3 Flash and Qwen 3.8 Flash Next. Three labs, three small-and-fast models, one shared promise: near-frontier behaviour at a fraction of the compute.
Read article →
GLM 5.3 full scores 92.8 locally against 93.8 through the API. One point, below judge noise. The real finding is elsewhere: the reasoning effort of a cloud run is not data, it is an assumption.
Read article →
GLM 5.3 Flash shipped open-source on Friday night. A weekend of work later it runs on one machine, holds within striking distance of the cloud, and never stops thinking.
Read article →
An anonymous model landed second on our board and refuses to say where it comes from. Three failed probes later, its token counter gave up its lineage.
Read article →
Qwen shipped the weights for 3.8. The full model fits on none of our servers. A tour of the Qwen family running locally, with the bench figures.
Read article →
Three days chasing DeepSeek V4 became a week down a rabbit hole: six confident wrong verdicts, and a bug hiding one layer below all of them.
Read article →
The same benchmark batch, half the cost overnight, no code changed. How a routing table of 51 measured models delivers frontier level at 10% of the price.
Read article →
The best open model in the world ran complete on five Macs, for half a day. What it took to get there, the three walls it hit, and why it was gone before dinner.
Read article →
I wanted a local, European, French-fluent LLM for EU regulatory work. I tested every Mistral model — flagship to small — on a 1.28TB Apple Silicon cluster. They all collapsed on the same test. The model I chose is Chinese. Here's why.
Read article →
A 3-billion-parameter active model scores 49 out of 50 on production Python generation. The same model scores 265 out of 500 on long-form prose. A 397-billion-parameter model in BF16 reaches 460 out o
Read article →
Non, un fichier safetensors ne peut pas vous pirater. Mais les risques réels sont ailleurs.
Lire l'article →
Nouvelle de science-fiction générée par Qwen3.5-397B en 13 minutes. 10 000 mots, zéro retouche.
Lire l'article →
Un benchmark d'écriture créative pour LLM : 6 critères, 500 points, une seule nouvelle de 5 000 mots.
Lire l'article →
The question is no longer if. It's who gets there first — and what they do with it.
Read article →
When you send a prompt to GPT-4o, Claude, Gemini, or DeepSeek — what precision is the model running at?
Read article →
Same document. Same prompt. Three models.
Read article →
February 19, 2026 — Sophie, The Monocle Bear
Read article →
Quand cinq IA « de pointe » échouent sur un problème qu'un humain résout en dix minutes
Lire l'article →
Sam Altman called advertising in AI "uniquely unsettling" in May 2024. A "last resort," he said.
Read article →
Lire l'article →
When people talk about artificial intelligence, the conversation often swings between two extremes. On one side, blind optimism: machines will soon outshine us, creativity will be automated, and the f
Read article →
Quand on parle d’intelligence artificielle, le débat se réduit souvent à deux caricatures. D’un côté, l’enthousiasme aveugle : l’IA va tout remplacer, automatiser la créativité, reléguer l’humain au s
Lire l'article →
Depuis deux ans, le “prompt engineering” est devenu une discipline à part entière. Les utilisateurs de modèles comme GPT-3.5 ou GPT-4 ont rapidement découvert quelques techniques miracles pour “forcer
Lire l'article →
In the past two years, “prompt engineering” has grown into a discipline of its own. Users of GPT-3.5 or GPT-4 quickly discovered some “magic tricks” to make AI reason more reliably. Two of the most fa
Read article →
One of the lingering habits of early podcast distribution is the idea that you can “publish” a podcast on YouTube simply by pairing the audio with a static background and a moving spectrogram. Technic
Read article →