洞见
博客与新闻
为医院、政府与金融机构而写的keen-expert笔记——清楚易懂、主张有出处。
- 阅读 →
A model alone is not an agent — why the harness matters
An agent is not just a clever model. It is a model plus tools, memory, permissions, sandboxes, and the loop that decides the next step. Vendors now ship that scaffolding as a product — the harness. Ask what is governed outside the weights.
ai-literacyagentsharnessbuying-aigovernance - 阅读 →
Two agents, not one magic model — Brand Onboarding + Screening
Respiratory fit-test programmes do not need one magic model. They need two governed agents: Brand Onboarding (baseline a mask brand with paired measurements) and Screening (suggest among onboarded brands only). Evidence gates matter — provisional promote is not a ≥99% claim.
orchai-agentsn95fit-testbuying-aigovernance-ai - 阅读 →
Hospital AI beyond the benchmark — why live use needs different monitoring
Public benchmarks measure someone else's exam. Inside a hospital, clinicians ask unpredictable questions under real liability and workflow pressure. Peer-reviewed deployment lessons keep saying the same thing: monitoring must follow live use — not only the scoreboard.
ai-literacyevalhealthcareclinical-aibuying-ai - 阅读 →
Same API name ≠ same model — why silent upgrades need your own eval
The model name in a slide or an API string is a label, not a frozen product. Providers can swap the checkpoint behind a stable ID, ship different surfaces under one family name, or add a faster tier that is still "the same model." Pin what you evaluated — and re-run your eval when anything moves.
ai-literacyevallibrarybuying-aiapi - 阅读 →
A governed media library isn't a shared Drive folder
A shared Drive (or NAS folder) stores files. A governed media / document library for AI must retrieve the right passages, cite sources, respect who may see what, and stay under your control. "Chat with the folder" without those pieces is a demo — not a production knowledge path.
orchai-librarybuying-airaggovernance-aiknowledge - 阅读 →
Computing power, LLMs, and model training — what more FLOPS buys
PFLOPS measure how much arithmetic a computer cluster can do per second. Training a large language model burns a lot of that capacity once; serving answers uses it differently, every day. More compute helps capacity and speed — it does not automatically make a model smarter for your documents and workflows.
ai-literacycomputetrainingllmsbuying-ai - 阅读 →
"SOTA on the leaderboard" vs "works on our work"
SOTA on a public leaderboard means a model scored well on someone else's test set. It does not mean the model will follow your SOPs, your Cantonese notes, or your compliance rules. A short honest eval on your work — including weak results — beats a shiny demo slide.
ai-literacyevalbuying-aibenchmarks 
阅读 →After the demo: what a buyer still needs before a production Agents PO
A polished Agents showcase proves the shape of the front door — login, role-based cards, admin control. A production purchase order still needs clear scope per capability, integration boundaries, packaging, evidence gates, and where data lives. Ask those before you treat the demo as the deliverable.
orchai-agentsbuying-aiprocurementgovernance-aidemo- 阅读 →
Open-weight vs closed models — what you actually control
Open-weight models let you download parameters and run them yourself (with the ops that implies). Closed models usually mean you send prompts to a vendor API. Neither is automatically more ethical or more secure — pick for control, residency, cost shape, and who supports you when it breaks.
ai-literacymodelsopen-weightself-hostbuying-ai 
阅读 →MoE vs dense — what's actually different?
A dense model uses its full parameter stack for every token. A Mixture-of-Experts (MoE) model keeps a large total "B" but typically activates only a subset of expert networks per token — so headline size, active compute, cost, and latency can diverge. Read both numbers before you buy.
ai-literacymodelsmoeparametersbuying-ai- 阅读 →
Why we still show up on-site instead of answering a long RFP
Long RFPs ask vendors to promise software in writing before anyone has touched the real workflow. An embedded pod shows up, builds in the buyer's environment, and owns the production outcome. That is how serious Digital Health and government AI work often starts — not with a 200-page reply.
forward-deployedbuying-aidigital-healthfdegovernance-ai 
阅读 →Does a higher "B" mean a smarter AI?
A higher "B" (billions of parameters) means a larger model — more capacity — not automatically a smarter one. Architecture, training data, how much of the model actually runs, and how you evaluate it matter more than the headline number.
ai-literacymodelsparametersbuying-ai
阅读 →「管治型 AI」是什么——以及为何你的数据可以留在原地
管治型 AI 是受规管机构能够信任、能够交代、能够持续扩展的 AI——把模型带到你的数据身边，而不是把数据送出去。它建基于三大支柱：安全（数据全程不离开你的掌控）、问责（每个答案皆可追溯）、扩展（在不失去监控的前提下成长）。
governance-aion-premiseregulated-aiorchai
阅读 →Launching OrchAI
Why we built an open platform for managing large language models on infrastructure you control.
announcementorchai