Blog
Browse thoughts, experiments, and tips.
2026-07-21
Fugu-Cyber: Sakana AI Points Its Orchestrator at Security Work
Sakana AI adapts its Fugu multi-model orchestrator into a gated tool for vulnerability verification and threat-intel detection — with a human in the loop before anything gets called real.
2026-07-21
Microsoft and Mistral Go Deeper: Sovereignty as a Feature
The expanded Microsoft-Mistral deal isn't just distribution — it's a bet that 'runs fully offline, on your own infrastructure' is what wins regulated enterprises.
2026-07-21
Why Every Lab Suddenly Has a 'Cyber' Model
Gemini 3.5 Flash Cyber lands the same day as Sakana's Fugu-Cyber, two and a half months after OpenAI's GPT-5.5-Cyber. That's not a coincidence — it's a pattern.
2026-07-19
Qwen3.8-Max-Preview: Alibaba's Answer to Kimi K3
Two days after Kimi K3, Alibaba previews a 2.4-trillion-parameter multimodal flagship it calls 'second only to Fable 5.'
2026-07-17
Understanding AI Model Quantization: From Bits to GGUF
Learning notes on how AI models get compressed from 16-bit weights down to 4-bit integers — bits, scale factors, PTQ vs QAT, and what GGUF names like Q4_K_M actually mean.
2026-07-16
Kimi K3: The Largest Open-Weight Model Yet
Moonshot AI's Kimi K3 is a 2.8-trillion-parameter open MoE that trades blows with the closed frontier — and undercuts it heavily on price.
2026-07-15
Soofi S: Europe's Sovereign Open-Weight Foundation Model
A look at Soofi S, the German consortium's sovereign hybrid Mamba-2/MoE model — trained entirely on EU infrastructure, with full data transparency.
2026-07-15
Inkling: Thinking Machines Lab's First Open-Weight Model
Mira Murati's Thinking Machines Lab ships Inkling, a 975B/41B-active MoE foundation model — and is refreshingly honest about where it doesn't lead.
2026-06-10
DiffusionGemma: When Text Generation Stops Being Sequential
Google's DiffusionGemma swaps token-by-token autoregression for parallel denoising — trading a bit of accuracy for a very different kind of speed.