Pinned
TL;DR: There is a new kind of company shape emerging. It does not feel like a robot CEO. It feels like a ticket queue that learned to hire temporary workers. In human companies, scaling means hiring. In AI companies, scaling means routing: sending each task through the right agents, tools, permissions, budgets, evals, and human approvals without losing the plot.
Recent AI agent progress made me start thinking about what comes next. My answer is AI organization.
Read more
Pinned
TL;DR: Software engineering is moving through three stages: manual coding, copilot-style autocomplete, and now agentic engineering where AI agents execute plans autonomously. 55% of developers already use AI agents regularly, 27% of production code is AI-authored, and agent session durations nearly doubled in three months. The transition is not coming -- it is underway.
Karpathy called it vibe coding in February 2025. Fifteen months later, at Sequoia AI Ascent 2026, he proposed agentic engineering as its professional successor. But framing this as vibe coding versus agentic engineering misses the larger picture. What we are actually watching is the third phase of a longer transition in how software gets built.
Read more
Pinned
TL;DR: OMC wraps AI agents in the most corporate language possible — talent markets, PIPs, HR, org charts — as a deliberate self-mockery. The point: organizational structures outlast the technology underneath them. We built the dragon so someone can break out of it.
OneManCompany (OMC) is an open-source agentic operating system where you play CEO and AI agents are your employees. From the very beginning, it was intended as a self-mockery of real-world capitalist companies. We deliberately wrapped a bunch of heterogeneous AI agents in the most “corporate” and cringe enterprise language we could find: talent markets, employee lifecycles, HR, PIP reviews, KPIs,...
Read more
TL;DR: Python remains central to training, fine-tuning, data science, evals, and research prototypes. In the application layer around agents, TypeScript keeps appearing because agent products inherit many constraints from web apps: schema validation, streaming UI, tool calls, user state, auth, and deployment. This is not a replacement story. It is a split between model work and product work.
TL;DR: Python remains central to training, fine-tuning, data science, evals, and research prototypes. In the application layer around agents, TypeScript keeps appearing because agent products inherit many constraints from web apps: schema validation, streaming UI, tool calls, user state, auth, and deployment. This is not a replacement story. It is a split between model work and product work.
Read more
TL;DR: An all-AI company is no longer science fiction. When AI agents can handle not just individual tasks but the full spectrum of organizational functions, a single founder might run a company that rivals a traditional team. This changes the economics, speed, and very definition of what a company is.
The Thought Experiment
Read more
TL;DR: What we call 'creativity' may be nothing more than the recombination of historical experience — individual and evolutionary — perturbed by noise and filtered by selection.
Throughout history, creativity has been regarded as one of humanity’s most defining traits. We build economies around it, award Nobel Prizes for it, and tell our children it is what separates us from machines. The “creative genius” — da Vinci sketching flying machines, Einstein reimagining spacetime, Jobs unveiling the iPhone — sits at the apex of our cultural mythology.
Read more
Drafts
Draft
Jul 28, 2025
PaperReading
TL;DR: This blog is a paper reading summary on meta reinforcement learning, and its application in LLM training procedure.
Key Takeaways
Preview Draft
Draft
Jul 28, 2025
PaperReading
TL;DR: This blog is a paper reading summary on LLMs, Reinforcement Learning, and In-Context Learning.
Introduction
Preview Draft
