Keep pulling the thread on Andrej Karpathy.
The role of a programmer is shifting from being a code writer to an orchestrator of AI agents.
Some applications, like the MenuGen web app, should cease to exist as traditional software stacks and instead be replaced by direct transformations performed by multimodal models.
LLMs and reinforcement learning automate tasks that can be verified, whereas traditional software automates tasks that can be specified.
A frontier model's capability spikes are a function of a task's verifiability, the training attention from labs, data coverage, and its economic value.
A key startup opportunity is to create reinforcement learning environments for valuable, verifiable domains that are currently undertrained by major AI labs.
Software products must be built with 'agent-native' surfaces like CLIs and APIs because the primary user will increasingly be an AI agent, not a human.
The capability of AI agents like Claude Code, Codex, and Cursor experienced a step change around December 2025, with generated code becoming larger, more coherent, and more reliable.
In the "Software 3.0" paradigm, the context window is the primary lever for programming an LLM, which acts as an interpreter over that context.
The installation process for OpenClaw, which uses a block of text instructions for an agent, is more powerful and adaptive than a traditional brittle shell script.
The significant improvement in GPT-4's chess-playing ability was likely due to the inclusion of a large amount of chess data in its pretraining set, not just from an increase in general intelligence.
State-of-the-art models exhibit 'jagged intelligence,' capable of complex tasks like finding zero-day vulnerabilities while failing at simple common-sense reasoning like whether to walk or drive 50 meters to a car wash.
The performance multiplier for engineers who master agentic workflows will be significantly greater than the traditional '10x engineer' concept.