autonomous coding + long-term coherence

Z / Runner 🏃🏻‍♂️

🏃🏻‍♂️ Scaling long-running autonomous coding: Running for weeks To test this system, we pointed it at an ambitious goal: building a web browser from scratch. The agents ran for close to a week, writing over 1 million lines of code across 1,000 files. You can explore the source code on GitHub.

🏃🏻‍♂️ implementing AlphaGo from scratch (repo will be open sourced soon) to catch up on foundational deep learning techniques, and also to re-learn how to program with the full power of modern coding agents.

• >/experiment Apply maximal update parameterization to find the best hyperparameters as I scale up. Start with GoResNet-100M as base. Use d-muP for depth-wise stability. Train 1 epoch on dev-train-100k. Submit up to 4 parallel Ray jobs.
• >/experiment Study scaling laws. Propose 5 model sizes from 1M to 1B params. Train each on dev-train-100k with bfloat16. Fit power law L(C) = a*C^b. Generate Chinchilla-style compute-optimal plots.
• >/experiment Play the NN checkpoint against gnugo10. Max 81 moves per game. Evaluate 20 games. Visualize the policy softmax heatmap on each board state.
• >Implement MCTS in src/alpha_go/mcts.py using AlphaGo paper nomenclature. Node with N (visit count), Q (value), logP (log-probs). Add PUCT selection, expansion, backpropagation. Test with toy environment.

🏃🏻‍♂️ "Coding assistants will soon become so good that they can conjure any digital system in an effortless way, like having a wish-granting genie for the price of $20 a month. Soon, an engineer can point their AI of choice at the website of any SaaS business and say, "re-implement that, frontend, backend, API endpoints, spin up all the services, I want it all".

Stephen / Basketball 🏀

🏀 Claude Opus 4.6 performs on a variety of benchmarks that assess its software engineering skills, multilingual coding ability, long-term coherence, cybersecurity capabilities, and its life sciences knowledge.