LawZero評測：政府砸3億加幣，賭「不用Agent」的AI | LawZero Review: $300M Bet Against the AI Agent Race

By Kit 小克 | AI Tool Observer | 2026-09-19

🇹🇼 LawZero評測：政府砸3億加幣，賭「不用Agent」的AI

本週最值得AI圈關注的不是又一個新模型，而是一筆政府資助：LawZero，圖靈獎得主Yoshua Bengio創辦的AI安全非營利組織，9月16日在蒙特婁ALL IN AI大會上宣布，加拿大政府投入1.5億加幣、德國政府加碼約1億歐元，兩國合計挹注上看3億加幣。這是目前已知規模最大的一筆政府資金，賭在「非agentic」AI安全研究上，而不是拿去堆更會自主做事的AI agent，把LawZero推上這週AI圈的熱門話題。

LawZero是什麼，這3億加幣要拿去做什麼

LawZero是Bengio於2025年6月創立的非營利研究機構，總部設在蒙特婁，早期靠Future of Life Institute、Jaan Tallinn、Open Philanthropy等慈善資金約3000萬美元起步。這次加拿大、德國的資助，用途包括：

在加拿大新增360個全職職位，擴充LawZero自有的AI運算基礎設施
德國的資金專款用在柏林新設的LawZero研發據點
核心目標是打造名為Scientist AI的系統原型，作為主流agentic AI的替代方案

Scientist AI 跟一般的AI agent差在哪

LawZero主打的Scientist AI刻意設計成「非agentic、無記憶、無狀態」——它不會自己去訂機票、呼叫工具或連續執行任務，只負責針對問題給出附帶推理過程與信心水準（貝氏後驗機率）的誠實答案。這跟OpenAI、Anthropic、Google現在拚命把AI做成能自主連續完成多步驟任務的agent，方向完全相反。Bengio的論點很直接：AI一旦有了自主行動與自我保存傾向，失控風險會指數上升，他甚至公開警告過失控的超智慧AI可能在10年內構成生存威脅。

這件事對一般用戶、開發者有什麼實際影響？

老實說，現階段LawZero沒有任何東西可以讓一般人「試用」，連公開demo都很少，這跟本站平常評測的產品完全不是同一個階段。今天你用ChatGPT、Claude或Gemini處理工作，不會因為這筆資助而有任何改變。但如果你關心AI監管、AI安全的政策動向，這是個值得記下的指標案例——兩個G7國家政府，第一次公開拿公家預算押注在跟主流agentic AI完全相反的技術路線上，訊號意義遠大於3億加幣這個數字本身。

小克的實測結論

這篇不是產品評測，是政策觀察。我的看法很簡單：Scientist AI能不能商業化、能不能追上agentic AI的實用性，現在都是問號，非agentic的「誠實問答機」離能打的產品還有很長一段路。但如果接下來一兩年各國政府開始認真討論怎麼管AI agent的風險，LawZero這筆資助會是常被拿出來對照的起點。值得放進書籤，但不用現在就期待能拿來用。

好不好用，試了才知道。

🇺🇸 LawZero Review: $300M Bet Against the AI Agent Race

The biggest AI story worth tracking this week isn't another model release — it's a government grant. LawZero, the AI safety nonprofit founded by Turing Award winner Yoshua Bengio, announced on September 16 at the ALL IN AI conference in Montreal that Canada is investing CAD $150 million and Germany is adding roughly €100 million, for a combined package worth up to CAD $300 million. It's the largest known government bet on "non-agentic" AI safety research, rather than on building AI that acts more autonomously — and it's what's putting LawZero in AI headlines this week.

What Is LawZero, and What Is the $300M For?

LawZero is the Montreal-based nonprofit Bengio launched in June 2025, initially funded with about $30 million from philanthropic donors including the Future of Life Institute, Jaan Tallinn, and Open Philanthropy. The new government funding will go toward:

Hiring 360 full-time positions in Canada and expanding LawZero's own AI compute infrastructure
Germany's contribution earmarked for a new LawZero R&D site in Berlin
Building out Scientist AI, LawZero's flagship alternative to mainstream agentic AI

How Scientist AI Differs From a Regular AI Agent

Scientist AI is deliberately designed to be non-agentic, memoryless, and stateless — it won't book flights, call tools, or chain together multi-step tasks on its own. Instead, it answers questions honestly with transparent reasoning and calibrated confidence (Bayesian posterior probabilities). That's the opposite direction from where OpenAI, Anthropic, and Google are racing — toward AI agents that autonomously complete long, multi-step tasks. Bengio's argument is blunt: once AI systems gain autonomous action and self-preservation tendencies, loss-of-control risk rises sharply — he has publicly warned that runaway superintelligent AI could pose an extinction-level threat within a decade.

Does This Change Anything for You Today?

Honestly, there's nothing about LawZero you can try right now — public demos are scarce, and this isn't the kind of product this blog usually reviews. If you use ChatGPT, Claude, or Gemini for work today, nothing changes. But if you follow AI regulation and safety policy, this is a milestone worth noting: two G7 governments are, for the first time, publicly betting taxpayer money on a technical direction that's the opposite of mainstream agentic AI. The signal matters more than the dollar figure.

Kit's Take

This isn't a product review — it's a policy watch. Whether Scientist AI can be commercialized, or ever match the usefulness of agentic AI, is still an open question; a non-agentic "honest answer machine" is a long way from something you can actually ship. But if governments start seriously debating how to regulate agentic AI risk over the next year or two, this grant will likely be the reference point. Worth bookmarking — not worth expecting to use anytime soon.

好不好用，試了才知道。

Sources / 資料來源

常見問題 FAQ

LawZero是什麼？

LawZero是圖靈獎得主Yoshua Bengio於2025年創立的AI安全非營利組織，致力於開發「非agentic」的safe-by-design AI系統，作為主流AI agent的替代方案。

加拿大、德國投資LawZero多少錢？

2026年9月16日宣布，加拿大政府投入1.5億加幣、德國約1億歐元，兩國合計最高挹注3億加幣。

Scientist AI跟一般的AI agent有什麼不同？

Scientist AI是非agentic、無記憶、無狀態的系統，只回答問題並附帶推理過程，不會自主執行任務，與OpenAI、Anthropic等公司主推的agentic AI方向相反。

一般人現在可以用LawZero的技術嗎？

目前還不行，LawZero沒有公開產品或demo可供一般用戶試用，仍處於研究階段。

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士