Had tons of fun last week at the @SakanaAILabs symposium 🗼 Blog: I first met Jürgen in Vienna at ICML 2024. We had just finished working on DiscoPOP and were pushing hard towards the AI Scientist v1. Fast forward to October 2026, and we are moving faster and faster towards a world of autonomous discovery that is solely limited by the resource constraints of the physical world. Very excited to have Jürgen join us on our path and looking forward to continuing the discussions on credit assignment in (autonomous) research communities, Free Willy, and benchpress PBs.

Early preview release of Fern 0.1 by @mirendil. Built on Mirendil-Base, the 0.1–0.5T coding model one researcher pretrained from scratch in 3.5 months for under $10M, with agents writing over 99% of the code. Target use: high-volume work for people building AI, like auditing training data, debugging runs from logs and reading agent traces end to end. Main feature: 10M-token context, about 23 MB of code or ~500K lines, roughly a Megatron-LM-sized codebase. - Next-token loss keeps going down out to ~10M tokens on their own repos. - RULER benchmark at 10M: 100% single-needle, 90% multi-key retrieval, 45% multi-hop QA. - Agentic, run by Mirendil with OpenCode, 1 attempt: 73.4% SWE-bench Verified, 53.2% SWE Atlas QnA, 44.2% DeepSWE, 6.1% Terminal-Bench 4.0. Larger versions are still in training: Base-Large (0.5–2T) is 80%+ through pretraining, XL (2–10T) 20%+.

Native Python Object-Oriented Agents: Reusing native Python abstractions for agent development eliminates the learning curve of new frameworks and leverages LLMs' existing Python knowledge, achieving great results on ARC-AGI-3. 1️⃣ Agent-as-Python-object programming model: methods = actions, fields = state, docstrings = prompts, type annotations = contracts 2️⃣ NOOA combines typed i/o, code as action (models write Python instead of tool calls), loop engineering 3️⃣ Pass-by-reference with bounded previews renders args as compact previews in prompts while models operate on full live objects in code execution. 📖: 📝: 🧑‍💻: