A reading note on AI application startups: begin with outsourced, rule-bound work and sell the outcome, not merely a smarter tool.
-
What We Talk About When We Talk About AI-Native Games
From conversational characters and world-aware NPCs to worlds that continue evolving after the player leaves, AI is becoming part of gameplay itself. This essay explores what AI-native games can offer players, and how memory, emergence, and generative freedom must remain grounded in game design.
-
When Code Becomes Cheap
Coding agents make code and prototypes cheap, shifting the scarce resources of software development toward judgment, verification, coordination, learning, and responsibility.
-
Demystifying evals for AI agents (Anthropic)
The capabilities that make agents useful also make them difficult to evaluate. The strategies that work across deployments combine techniques to match the complexity of the systems they measure.
-
What Should World Models Be, and How Should We Use Them?
World models matter less as beautiful video generators than as internal simulators for agents. This post compares the main routes, explains why JEPA is appealing, and revisits Critique of World Model and PAN.
-
Make OCR Great Again
A retrieval-oriented view of production OCR: what a parser should deliver for text, sections, formulas, images, and tables; how to route documents across Docling and MinerU; why DoclingDocument makes a good unified representation; and why the representation layer is not the query layer.
-
Synthetic Data as Engineering: Starting from Terminal-Corpus
Notes on SFT synthetic data engineering, starting from NVIDIA’s Terminal-Corpus and connecting it to classic and recent work on instruction synthesis, taxonomy-driven generation, executable tasks, and curation pipelines.
-
Werner Vogels' Last Lesson: The Renaissance Developer in the AI Era
Werner Vogels’ final re:Invent keynote framed the AI-era developer as a Renaissance Developer: curious, systemic, communicative, responsible, and polymathic.
-
From Designer to Agent Builder: See, Make, Think
Designer training also works for AI Agent builders: see more real systems, solve real problems, and keep asking why.
-
JoyAI-VL-Interaction: From Chat Back to Continuous Interaction
Chat is one way we package generative models, not the natural shape of the model itself. Interaction models return models to continuous tokens, perception, and action, with JoyAI-VL-Interaction as a concrete vision-language example.