Data science, career and other topics
21 September 2026
DeepSeek-V4.1-Flash is a 552B multimodal MoE model that attacks KV cache cost along all three of its multiplicative dimensions at once, then redesigns prefill and cache persistence around the result. It activates 8B parameters per token during prefill and holds 890 bytes of global cache per token.
10 September 2026
A write-up of the ROGII Wellbore Geology Prediction competition: how I adopted an AI driven approach and got a silver medal.
08 September 2026
Qwen-Drive-1.0 turns Qwen3.5-4B into a unified autonomous-driving model for 3D perception, scene understanding, reasoning, and trajectory planning while largely preserving its general vision-language capabilities.
28 August 2026
OpenAI's post-mortem of the Hugging Face incident got me thinking about what it would take for an AI system to run without anyone launching or controlling it.
24 August 2026
A review of Build a DeepSeek Model (From Scratch), a hands-on book that rebuilds the DeepSeek innovations one at a time, from KV caching and Multi-Head Latent Attention through DeepSeek-MoE, multi-token prediction, FP8 training, GRPO, and distillation, at a scale that runs on a laptop.
03 August 2026
Kimi K3 is a 2.8-trillion-parameter open-weight multimodal MoE model that combines hybrid linear–global attention, selective residual connections across depth, and large-scale agentic reinforcement learning.
