Mirendil評測：讓AI做研究的新創，估值3個月翻5倍 | Mirendil Review: The Startup Building AI That Does AI Research

By Kit 小克 | AI Tool Observer | 2026-09-24

🇹🇼 Mirendil評測：讓AI做研究的新創，估值3個月翻5倍

Mirendil是什麼？為什麼估值3個月漲5倍？

Mirendil是一家2026年才從隱身模式出來的AI新創，主打打造「會自己做研究的AI」——不是聊天機器人，而是能設計實驗、搜尋超參數、評估模型結果、再跑下一輪訓練的自動化研究系統。今年6月，Mirendil以10億美元估值募得2億美元種子輪；才過三個月，9月傳出正在洽談新一輪最高10億美元的募資，估值直接跳到50億美元。這種漲幅在AI新創圈也算誇張，成為這幾天矽谷圈子裡討論度最高的話題之一。

Mirendil的創辦團隊背景硬在哪？

公司由Behnam Neyshabur（CEO）與Harsh Mehta（CTO）共同創辦，兩人都是2025年底才離開Anthropic的資深研究員。Neyshabur是知名優化演算法SAM的共同發明人，Mehta則在Anthropic內部推動過「用AI工具自動化研究流程」的專案。團隊目前約20人，來自Anthropic、xAI、DeepMind、OpenAI等一線實驗室，種子輪由a16z與Kleiner Perkins聯合領投，NVIDIA也參與其中，新一輪則傳出Kleiner Perkins主導、a16z洽談跟投。

Mirendil實際要解決什麼問題？

Mirendil的目標是把AI研究員的工作流程自動化：設計實驗、挑選訓練設定、評估模型表現、決定下一步怎麼調整，再自動跑下一輪。公司規劃讓化學、醫療、機器人等領域的科學家，不需要養一整支AI工程團隊，也能用這套系統訓練、迭代自己的專屬模型。這個方向跟近期業界討論的「AI自動化AI研發」趨勢一致——只是Mirendil想把它做成一個對外開放的平台，而不是留在單一實驗室內部用。

誠實地說，這波估值該怎麼看？

目前Mirendil沒有公開產品、沒有公開benchmark，外界完全看不到這套「自動化研究系統」實際跑起來的樣子。50億美元估值，賭的其實是團隊履歷和「自我改進AI」的敘事，而不是已驗證的成果——這跟過去幾年不少AI新創「先拿到夢幻估值，產品再慢慢補上」的劇本很像。值得留意的地方在於：如果Mirendil真的做出能穩定自動化實驗迭代的系統，對中小型研究團隊會是真實的效率提升；但如果只是包裝過的自動化管線，那這波募資更多是反映投資人對「遞歸自我改進」故事的追捧，而非技術本身的突破。在看到實際產品、客戶名單或第三方測試之前，建議把它當成一個要持續追蹤的觀察對象，而不是急著下定論。

好不好用，試了才知道。

🇺🇸 Mirendil Review: The Startup Building AI That Does AI Research

What Is Mirendil, and Why Did Its Valuation 5x in 3 Months?

Mirendil is an AI startup that came out of stealth in mid-2026 with a simple pitch: build AI that does the job of an AI researcher — not a chatbot, but a system that designs experiments, searches hyperparameters, evaluates model results, and kicks off the next training run on its own. In June, Mirendil raised a $200 million seed round at a $1 billion valuation. Three months later, in September, it is reportedly in talks for up to $1 billion more at a $5 billion valuation — a jump that has turned heads even by AI startup standards and become one of the more talked-about stories in Silicon Valley this week.

Who Is Behind Mirendil?

The company was co-founded by Behnam Neyshabur (CEO) and Harsh Mehta (CTO), both senior researchers who left Anthropic in late 2025. Neyshabur co-invented SAM, a widely used optimization algorithm; Mehta previously helped build internal tooling at Anthropic to automate parts of its research pipeline. The team has grown to roughly 20 people pulled from Anthropic, xAI, DeepMind, and OpenAI. The seed round was co-led by a16z and Kleiner Perkins with NVIDIA participating; the new round is reportedly led by Kleiner Perkins, with a16z in talks to join again.

What Problem Is Mirendil Actually Trying to Solve?

Mirendil wants to automate the AI researcher workflow end to end: design an experiment, pick training settings, evaluate results, decide what to adjust, and launch the next iteration — with minimal human intervention. The plan is to let scientists in chemistry, medicine, and robotics train and refine their own specialized models without needing a full in-house AI engineering team. It is the same "AI automating AI R&D" trend labs have been discussing internally, except Mirendil wants to package it as a platform other researchers can actually use, rather than keeping it locked inside one lab.

Honestly, How Should You Read This Valuation?

Right now, Mirendil has no public product and no public benchmarks — there is no way to independently verify what this "automated research system" actually does in practice. A $5 billion valuation is a bet on founder pedigree and the "self-improving AI" narrative, not on proven results, which echoes a familiar AI-startup script: raise on the story first, ship the product later. The upside worth watching is real — if Mirendil delivers a system that reliably automates experiment iteration, that is a genuine efficiency gain for smaller research teams who cannot afford a large AI engineering org. But if it turns out to be a well-packaged automation pipeline dressed up in "recursive self-improvement" language, this round says more about investor appetite for that narrative than about any technical breakthrough. Until there is a real product, customer list, or third-party evaluation to look at, this is one worth tracking, not one worth concluding on yet.

You won't know until you try it.

Sources / 資料來源

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