從 Vertical AI 到 Physical AI：台灣最強的 vertical，會不會在硬體？

LLM 從 70 分走到 85 分

過去這幾年，在網路的世界裡，大型語言模型扮演了很重要的角色。從多數人認知的 ChatGPT 出現以來，大概是 2022 年，那是讓一般使用者可以用更親近、親民的方式接觸到 AI 的突破口。後續幾家大型模型業者也各自演化，不管是 Anthropic 或 Gemini。最近 Anthropic 有 IPO 的消息、公佈了 S-1，有機會我們再來分享 S-1 的內容。〔S-1 連結〕

大致上可以看到，大型模型的初衷就是收集各式各樣公開的 public data，拿來訓練自己的模型，然後持續精進。這幾年精進的結果，不管是 AI 的反應或者思考模型，大概都到了大家覺得非常可用的程度。只是在特定的 vertical，它的表現在該領域專家眼裡，不一定有那麼好。

用白話文講，一般的大型語言模型，從一開始大家覺得「哇，已經有 70 分這麼好」，後來提升到 75、80，甚至 85 分。對一般人而言，在你不熟的領域跟這些 AI 對談，能拿到一個 85 分的答案，基本上已經綽綽有餘，足夠讓你初步認識這個領域的知識。

Vertical AI 的 moat：引擎是別人的，原料是你的

但這幾年下來，多數的投資人和創業者也慢慢領悟到 vertical AI 的重要性。因為你可以把你獨特領域的 proprietary data，或者這個領域的 know-how，寫進你的 AI agent 裡。外部的大型語言模型，不管開源或閉源，只是一個推動你的引擎。只要餵進去的原料是你獨有的，引擎只是讓你跑得更快，你就不會在短期之內被大型語言模型取代。這就是一種 vertical 的 moat。

當然，每個 vertical 大家都可以用這個故事來講自己的 moat。至於其他延伸性的 moat 是什麼，就得另外在個別案子裡探討。

台灣做 Vertical AI，最強的強項是什麼？

從台灣的角度出發，我也曾經想：所謂的 vertical AI，我們到底可不可以透過現在既有的優勢，在軟體業做出一些好的成果，接軌上這一波大型語言模型的典範轉移？

我認為這些典範轉移的成果，都是在大市場才享受得到。所以你也看到近年有很多團隊往美國市場去，希望拿到更多 traction，甚至拿到當地投資人的投資，建立起一家世界級的軟體公司。

回到台灣的觀點：台灣如果要用某個 vertical 來做 AI 應用，最強的強項會是什麼？在我看來，目前比較明顯的答案，就是硬體相關的 AI。這裡指的不是「生產硬體」那一類題材，像是讓 AI 有算力、更節能、散熱這些；而是基於硬體或工業應用這個 vertical，我們會不會有什麼樣的軟體 AI 應用？

回到 LLM 的起點，再看 Physical AI

回到網路世界大型語言模型的思考。當時能夠起來，也是因為大量資料收集之後，訓練出很強的模型，才有前面比喻的從 70 分變 75、80，再到 85。ChatGPT 出來的那一刻，有點像所謂的 ChatGPT moment，大家一開始很驚豔，久了就覺得「它還蠻常出錯的」。但時至今日，多數內容出錯的比例已經降低很多，AI 的幻覺在你懂得用其他技術控制之下，也不會這麼明顯，實用度大大提升。

但如果從 Physical AI 或 Robotic AI 的角度來看，這場仗才剛開始。

如果我們說 LLM 是軟體的大腦，你可以想像，這顆軟體的大腦有沒有機會裝到機器人裡，變成機器人的大腦？或者變成我們工業應用場景裡的那顆大腦？

我們分幾個階段來講。如果是把現在的大型語言模型套進 industrial 的領域，像是機台的資訊擷取、參數的調整等等，我認為現在 LLM 的應用，多數場景是用得上的。

但回到 Robotic AI，這反而是另一個還在等「Robotic AI 的 ChatGPT moment」出現的時刻。也就是說，當機器人、外骨骼這些硬體與日俱進的同時，大家的想像是：未來會不會有一個好的控制機械的大腦？它可以看到外在環境、接收人類的指示，然後做出相對應的活動與行為。

機器人的大腦：VLM 與 VLA

所以會不會有機器人大腦版的 LLM 出現呢？介紹幾個例子跟大家分享。目前比較常見的是所謂的 VLM 跟 VLA，分別是 Vision-Language Model 和 Vision-Language-Action。一個是透過視覺知道你看到什麼、理解現在場域的狀況，接下來要做出哪些反應；反應的話，就需要一個模型去 trigger 身體的部件、肌腱怎麼動作，才能讓機器人有效活動，而且拿了東西不會掉，很精準地執行它該做的事。只是外在環境會一直變動，所以要邊看著環境、邊調整、邊執行。

在 VLM、VLA 的世界裡，特別是 VLA，還沒有一個像 Claude、ChatGPT 或 Gemini 這樣的模型出現，很多新創也就大量投入這件事情。

從代工到代訓：台灣的切入角度

回到模型本身，模型就是需要大量的資料訓練。所以台灣的切入角度，也許可以先從「資料提供商」的角色開始：可能是提供資料，或者有一種商業模式，類似從以前的代工延伸到「代訓」。也就是我們把手上有的資料，經過某種貼標、標註或處理，讓它更符合可以拿來訓練的樣子。

當然，台灣過去比較欠缺的，是對這種模型有理解的人才。我認為大公司可以透過投資等方式，接軌矽谷一流的模型訓練人才，特別是在 Robotics AI 領域，再搭配我們有的資料，也許是台灣往軟體 AI 前進的一種好方式。

從硬體接到軟體的人，來聊聊

以上是近期的一些小分享。Addin Ventures 這幾年看的案子，越來越多落在 Vertical AI 和 Physical AI 這兩條線上。如果你正在做 Physical AI 或 Robotic AI 相關的題目，或者想從硬體層面接進軟體，都歡迎跟我們聯絡。我相信我們有很多想法和資源可以互相交流，讓你的創意有機會更成功，打進國際市場。

有興趣的話，歡迎透過 contact@addin.vc 與我們聯繫。

本文同步刊於 ntu.vc。

From Vertical AI to Physical AI: Is Taiwan's Strongest Vertical in Hardware?

LLMs went from 70 points to 85

Over the past few years, large language models have played a major role in the online world. Since ChatGPT entered most people's awareness, around 2022, it has been the breakthrough that let ordinary users approach AI in a friendlier, more accessible way. The major model companies have each evolved since, whether Anthropic or Gemini. Anthropic recently announced an IPO and published its S-1; I may share some of that S-1 another time. 〔source to add: S-1 link〕

Broadly, the original idea behind large models was to collect every kind of public data, train on it, and keep improving. The result of those years of improvement is that both AI responses and reasoning models have reached a level most people find genuinely usable. The catch is that in a specific vertical, their performance doesn't always look that good to an expert in that field.

In plain terms, a general LLM went from "wow, it's already a 70" to 75, 80, even 85 points. For most people, getting an 85-point answer in a field you don't know is more than enough to get a first grasp of the subject.

The vertical AI moat: the engine is someone else's, the raw material is yours

Over these same years, most investors and founders have gradually come to see the importance of vertical AI. You can take your proprietary data from a unique domain, or that domain's know-how, and write it into your AI agent. The external large language model, open or closed, is just the engine that pushes you. As long as the raw material you feed in is uniquely yours, the engine only makes you go faster, and you won't be replaced by the big LLMs in the short term. That is a vertical moat.

Of course, every vertical can tell this same story about its moat. What other, extended moats look like is something to examine case by case.

What is Taiwan's strongest vertical for AI?

From Taiwan's point of view, I've wondered: with the advantages we already have, can we produce real results in software and plug into this paradigm shift driven by large language models?

I believe the gains from a paradigm shift like this are only enjoyed in big markets. That's why you've seen so many teams head to the U.S. in recent years, hoping for more traction, even local investors, and the chance to build a world-class software company.

Back to Taiwan: if we pick a vertical to apply AI to, what is our strongest card? To me, the clearest answer right now is hardware-related AI. I don't mean producing hardware, the compute, power efficiency, and cooling themes. I mean this: within the vertical of hardware and industrial applications, what software AI can we build?

Back to where LLMs started, then on to Physical AI

Think back to how LLMs took off online. It happened because massive amounts of data were collected and trained into strong models, which is where the 70-to-75-to-80-to-85 story came from. When ChatGPT arrived, the so-called ChatGPT moment, everyone was amazed at first, and then after a while thought, "it gets things wrong a lot." Today the error rate on most content has dropped sharply, hallucinations are far less obvious once you know how to control them with other techniques, and usefulness has jumped.

But seen from the angle of Physical AI or Robotic AI, this fight has only just begun.

If the LLM is the brain of software, imagine whether that brain could be installed in a robot and become the robot's brain, or the brain inside our industrial application scenarios.

Let's take it in stages. If you apply today's LLMs to industrial settings, pulling information off machines, adjusting parameters and so on, I think most of those scenarios already work.

But for Robotic AI, we're still waiting for Robotic AI's own ChatGPT moment. As robots, exoskeletons, and other hardware improve by the day, the question everyone imagines is: will there be a good brain for controlling machines? One that can see its surroundings, take instructions from humans, and act accordingly.

The robot's brain: VLM and VLA

So will an LLM for the robot's brain appear? A couple of examples. The most common today are VLM and VLA, Vision-Language Model and Vision-Language-Action. One uses vision to know what it's seeing and understand the state of the scene, then decides what to do next. Acting requires a model that triggers the body's parts and tendons to move, so the robot can act effectively, hold things without dropping them, and execute precisely. And since the environment keeps changing, it has to watch, adjust, and execute all at once.

In the world of VLM and VLA, especially VLA, no model like Claude, ChatGPT, or Gemini has emerged yet, which is why so many startups are pouring into it.

From contract manufacturing to contract training: Taiwan's way in

Back to the model itself: models need huge amounts of training data. So Taiwan's entry point may start from the role of data provider. That could mean supplying data, or a business model that extends from the old contract manufacturing into "contract training": taking the data we have, labeling, annotating, or processing it so it's fit for training.

What Taiwan has lacked is talent that truly understands these models. I think large companies can use investment and similar routes to connect with top-tier model-training talent in Silicon Valley, especially in robotics AI, and pair that with the data we already have. That may be a good path for Taiwan toward software AI.

If you're bridging hardware to software, let's talk

That's my recent reflection. More and more of what Addin Ventures looks at these days falls on two lines: Vertical AI and Physical AI. If you're working on Physical AI or Robotic AI, or trying to move from the hardware layer into software, please reach out. I believe we have plenty of ideas and resources to exchange to give your idea a better shot at success and at breaking into international markets. Reach us at contact@addin.vc.

Also published on ntu.vc.