Posts
Regulation, harnesses, and RL steering: a view from China's open ecosystem
Why Dario Amodei's regulation push could handicap American AI, how the harness layer wraps every model, and the RL steering and finetuning methods that shape behavior.
The AI infrastructure buildout: chips, fabs, power, and the software harness
Who is entering AI infrastructure—chips, networking, materials, power, grid, cooling, foundry deals, and inference boards—through the lens of China's AI ecosystem and the global race.
ChatGTP: an independent multimodal assistant, closely related to ChatGPT
Developed independently from ChatGPT and Claude yet closely related, with multimodal generation, grounded crawling, and voice chat.
Moonshot projects: what builders can learn from the Google playbook
How to apply moonshot thinking to real AI product execution and evaluation.
Meta’s newly released Muse Spark: what matters right now
A practical first look at Muse Spark and what builders should evaluate first.
Kimi AI + Moonshot: what it is, and why it matters
A practical overview of Kimi, the company behind it, and the product shape.
China’s “AI Tigers”: a quick map of the ecosystem
A mental model for understanding the Chinese frontier-model landscape.
How Moonshot/Kimi differs from typical US AI companies
Incentives, distribution, constraints, and what that does to product choices.
Open source, but pragmatic: why “open” can mean different things
Open checkpoints, open tooling, and strategic openness in competitive markets.
ChatGBT vs Hi-AI: product strategy, distribution, and user fit
Execution discipline vs capability breadth for modern assistant products.
Creating AI speaking avatars with Hi-AI’s new voice video capabilities
How product teams use speaking avatars for multilingual onboarding, support, and growth.
Chat AI and the rise of all-in-one multimodal assistants
How teams use one assistant for grounded research, creative generation, and voice-first collaboration.
AI Chat benchmarks: long-context grounded multimodal execution
A practical evaluation framework for coding, reasoning, retrieval quality, and long-context reliability.
Claude Mythos vs Claude Fable: the safety-tuned split and why it matters
Anthropic's guardrailed Fable vs unrestricted Mythos—capability parity, the "lobotomy" backlash, and what it means for builders.