提问
从产品尚无法回答的问题出发。
Archymedes 实验室
一个讲究测量的实验室。
实验室研究产品背后的问题：AI 何时该行动，它所知道的在实地是否成立，以及如何让它便宜到随处可用。我们公开方法、代码和局限。
方法
实验室如何工作
测量
以真实情况为准，而不是看起来是否可信。
舍弃
不合格的就放弃，无论多方便。
发布
公开论文、代码、数据以及明确的局限。
研究方向
三条研究线
不确定性下的决策
对于无人尝试过的事，AI 何时该行动？
Generative models are good at proposing options and bad at choosing between them. This line of work keeps the two apart: models judge, calibrated maths decides, and an uncertain option is checked against a fallback instead of being vetoed.
成果
- Unknown Is Not Risky ↗Paper, October 2026
- cdm ↗Open-source library and data
扎根物理世界
它在实地成立吗？
Rebuilding real places as verified, lightweight 3D scenes from free public data. Every layer is registered against the best measurement available and its error is reported; a source that fails against the ground is dropped, however convenient.
成果
- geo3dReference build: a mountain valley
- Gisenyi, Lake KivuA second scene, in Rwanda
高效的智能
它能在普通硬件上运行吗？
Making capable models cheap enough to run where connections are slow or absent: compression and quantisation for small language models, and tooling that tunes GPU code without trading away correctness.
成果
- Small modelsCompression and quantisation research
- WarpLabGPU code tuner, early beta
论文
Unknown Is Not Risky
Unknown Is Not RiskyCreative Decision Making with Calibrated Decision Models
AI 何时该采纳一个无人尝试过的想法？不要直接否决，而是在保留退路的前提下先检验，并为检验可能漏掉的风险定价。
- 01
One check, one number
For a single decision, the value of a check depends on the belief about harm only through its mean. Ignorance and a known coin flip warrant the same action; the difference matters for learning across decisions.
- 02
Checking beats vetoing
On held-out synthetic problems, checking uncertain options halves the regret of vetoing them (0.045 vs. 0.091) at about 1% undetected harm, against 17–23% for acting on them unchecked.
- 03
Tested on a real judge
Over 237 TypeSafe Jev judgments with known ground truth, Jev separates explicit harm from explicit safety almost perfectly (AUROC 0.9998), and cdm has the lowest regret and harm of the policies compared.
- 04
Stated limits
The gain comes from checking, not from the CVaR reliability gate; the Jev benchmark is small and templated; none of this yet shows better decisions on real outcomes.
