The goal is not to manufacture certainty. It is to turn broad curiosity into questions that can be examined, challenged, and revised. Each path begins with a concrete problem and keeps a record of what changed along the way.
Study hard what interests you the most in the most undisciplined, irreverent and original manner possible.
Study hard what interests you the most in the most undisciplined, irreverent and original manner possible.
Which model learns better with the same resources, and which delivers better quality for the same cost? NanoTLM compares nanochat and Extropic's Z1T under shared parameter, data, and GPU-time budgets.
t3rl combines T3 Code's programming tools, GitHub workflows, and coding agents with reinforcement learning experiments and LLM post-training research.
NanoTLM is a fast, reproducible laboratory for testing thermodynamic language-model architectures, comparing them with traditional baselines such as nanochat, and letting AI optimize candidates through autoresearch.
Microsoft’s MDASH reached 88.45% on CyberGym. The next eleven weeks showed why cybersecurity’s durable AI advantage will live in open agents, tools, benchmarks, and proof, not in any single model.
The next entries will continue the Bayesian agent experiment, develop the post–von Neumann reading series, and turn promising questions into prototype write-ups. Each update will link methods, limitations, and artifacts when they are ready, while older notes remain open to correction.
Browse every research note