Jakub Strawa·Jul 12Native Multimodality — Beyond Language-Centric Multimodal ModelsI gave a talk recently about native multimodality, and enough people came up to me afterwards asking for a write-up that I figured I should…
Jakub Strawa·Nov 11, 2025Policy Descent Mirror Optimization Is The New Golden Standard In RL?There have been reports and leaks, that the frontier labs with leading models like Kimi-2 Thinking are using a new KL objective for…
Jakub Strawa·Sep 23, 2025DeepEyes & Mini-o3 — Thinking with Images through Reinforcement Learning.At the end of May this year, a team from Xiaohongshu (RedNote) and students from Xi’an Jiaotong University released an interesting paper…
Jakub Strawa·Aug 29, 2025Cascade Reinforcement Learning, MPO, GSPO — Multimodal ReasoningShanghai AI Laboratory has just unveiled InternVL3.5, a major leap forward in multimodal reasoning models. In this post, I’ll walk through…
Jakub Strawa·Apr 23, 2025Diffusion Models from Scratch in PyTorch: A Step-by-Step GuideOver the past few years, diffusion models have taken the generative modeling world by storm — powering tools like DALL·E 2 and Stable…A response icon1A response icon1
Jakub Strawa·Apr 7, 2025Why Cross-Attention is the Secret Sauce of Multimodal ModelsEver wonder how models like GPT-4V, Flamingo, or Stable Diffusion can look at an image and describe it, read text and refer back to the…A response icon1A response icon1
Jakub Strawa·Mar 14, 2025Faster Transformers? Flash Attention(s)The road to fastest transformers model!
Jakub Strawa·Mar 2, 2025Building Agents with Model Context Protocol (MCP)A blog post about how to build Agents with MCP.A response icon1A response icon1
Jakub Strawa·Jan 13, 2025DPO & ORPO — Overview of Preference Alignment algorithms for LLM finetuning.What’s the best way to align LLM with human preferences?
Jakub Strawa·Dec 29, 2024Fine Tuning Florence-2 VLM for invoice extractionExtract key aspects from your invoices!