Thoughts & Investments
Andrew Yang
*This is my personal space to think & write about what I'm investing in, what I'm seeing in technology, and other random thoughts.
Table of contents
  • About Me
About Me
My current obsession is building AI tools and investing in them. I've personally invested in Gamma, Blaze AI, Please (formerly MultiOn), Better Brain, etc. Though I also invest to learn about unique business models (Beyond Health) and trends (Shield AI). I'm also a tiny LP in a couple of incredible emerging funds such as Shumer Capital (by Hyperwrite founder Matt Shumer).
My day job is Senior Director at Darwin Ventures, where besides investing I build AI-powered workflows for the team and handle external communications.
Prior to Darwin I had my own consultancy, Presentality, where we helped startups in Taiwan raise over $100M in venture funding and worked with public companies on their communication strategy. Even prior to that I was a political staffer.

www.linkedin.com

My Investments

Gamma A new medium for presentations, powered by AI When I invested: Seed, pre-revenue Status: 100M users, $100M ARR, Series B led by a16z at $2.1 billion valuation Instant Presentations, Websites, and More with AI | Gamma Shield AI AI powering drone swarms and other key military applications When I invested: Late (secondary) Status: $5B valuation, winning mega contracts Shield AI Beyond Health Primary clinic roll-up and digitization play When I invested: Series A Status: Fast growing, and profitable. Revenue is higher than the valuation I invested at, that's always a good sign 🤯 Home One Phaseshift AI to discover new materials for industrial use When I invested: Seed Status: Secured two large industrial applications and proving out its technology and approach Home - Phaseshift Technologies | Alloy Design | Materials | Toronto Please When I invested this was MultiOn. Its pivoted from browser agents to life planning. When I invested: Series A Status: Pivoted and restarting Please Better Brain When I invested it was SQL agent for data teams. Now its workflow automation. When I invested: Seed Status: Pivoted and finding PMF BetterBrain Blaze AI Marketing autopilot - generating weeks worth of content across Instagram, Tiktok and other platforms When I invested: Series A Status: Growing fast Blaze | AI That Does Marketing For You Parker Data platform for US mobile home park investment When I invested: Seed Status: Announced $50M to invest into >1,000 mobile home units Invest in real estate's best kept secret

Why haven't we seen a gen AI calendar unicorn yet?
December 9, 2025
In the early days of Gen AI wave (circa late 2022-early 2023), there was a strong argument that we'd soon see an AI-native calendar unicorn.
After all, the calendar is central to our productive life: it's where we coordinate meetings, deadlines, and figure out where to fit in productive hours or even personal time. It tells us what to do and where to go next. It's therefore easy to imagine the calendar as a command center: managing our time and issuing task orders.
And there were strong contenders too: Motion, Amie, Howie, and there's also pre-AI darlings like Clockwise and Vimcal.
Calendly as a COVID-time darling was in danger.
But two years later, no AI calendar unicorn has emerged (yet). Instead, here's what's happened:
  • Motion pivoted to broader AI agents.
  • Amie became yet another AI notetaker.
  • Howie became "AI secretary" (though heavily focused on scheduling)
  • Calendly is alive and well
Why is this?
I find this question hugely interesting, so below are arguments and questions to help me think through it all.
Read More
Auto-complete: the next big thing in human-AI interface?
December 4, 2025
Andrej Karpathy said something in this interview that got me thinking.
[I'm paraphrasing]:
he doesn't like having to type so much English to LLMs to vibe code stuff.
He prefers "auto-complete": typing short phrases that indicate intention, LLM gets it, and completes the rest.
Btw, if you don't know who Andrej Karpathy is, he was the former director of AI at Tesla and co-founded OpenAI, and is one of the most influential minds in AI.
Oh and he also coined the term "vibe coding".
To go into more detail, he shared that a project he worked on "grew out" in sections. Essentially
  1. He defined the sections
  1. "Grew" each of them out
  1. Using a combination of hand-coding and LLM auto-complete
I think this has a lot of promise as a form of human-AI interface, probably because it fits my own belief that artificial intelligence is in the end all about human intelligence.
But let me enumerate the reasons:
You're still doing the thinking: AI isn't doing the thinking for you, it's just completing your thought. This way, you're still training your brain and being creative, and that's really important.
Let AI complete stuff you don't need to recall: The portion that AI is completing, often you don't need to remember anyway, e.g. in legal context, a standard clause that any lawyer would know and has written out hundreds of times, but it's a waste of time to even open another file and copy+paste.
You remain engaged: While you're doing all this, you're engaged. Because you write something, AI completes it, you review it, and onto the next bit - you're still locked in.
Btw, Gamma also has this feature, if you start a sentence and type three "+" in a row, AI will finish the sentence for you, so simple, so useful. This sentence was written that way too.
What if we added connectors?
Now, I have this theory slash dream of combining this with connectors. Connectors = all those integrations everyone is building, e.g. ChatGPT being able to read your Gmail and G Drive, etc.
Imagine that you're managing 50 customers, you can't possibly remember the founding year, CEO, or pipeline status of every single one, but you need to write about them in reports.
If your document is directly connected to the data sources, and has auto-complete?
  1. Type out the company name + what you want, like "Linear sales conversation summary"
  1. And AI auto-completes it for you
I think that'd be pretty awesome.
People intelligence and intelligent people
December 3, 2025
AI 浪潮下，另一個值得關注的領域是 “people intelligence"：
  • 首先，用戶群廣大到不行（Like, everyone who has LinkedIn?）
  • 而且，剛好 AI 非常擅長這種工作：能夠處理跟「人」還有「人際關係」相關的海量資訊
  • 還沒有明顯的「新創」贏家
最後那一點，你可能會想：「LinkedIn 肯定全吃啊，還有什麼好玩的？」
也許吧，但我認為不一定！
新創還是有很多可以做的，不過並不是開發出更強大的 AI（It's never about the AI actually...），而是找到新的使用情境/流程/介面。
比如 YC 出來的 Happenstance AI，在我看來就展現了這樣的能力：這兩天推出的 calendar research agent，就自動 research 你會議要見的人，然後把資訊直接放在你的 calendar event 裡面。
底下，串到你的 calendar 之後，就什麼都不必做，它就自動去工作了：
然後結果，就自動出現在你的 calendar 裡面
*我的直覺是「喔還不錯！聽起來很方便！」，是否真的好用而且會繼續用還未定，畢竟大家都有用過直覺很棒，但用起來的現實不佳的產品。
你可能會想：這哪有什麼創新？以前也有其他工具會這樣啊！
沒錯，以前有很多其他新創都有用這些「元素」（現在矽谷很流行說 “primitives"），但我個人感覺是 Happenstance 把這些元素重組的很棒，很直覺，讓我一看就懂，這個很重要！

Happenstance

Happenstance

Find people in your networks. Connect your Gmail, LinkedIn, or Twitter accounts to get started.

Just read this
December 1, 2025
I'm not gonna say anything else. Just read this.

First Round

How Superhuman Built an Engine to Find Product/Market Fit

Superhuman founder and CEO Rahul Vohra walks us through the framework his startup used to make product/market fit more actionable, detailing the survey and four-step process that were key to measuring and optimizing it.

Why building the AI version of legacy software isn't enough
November 26, 2025
As a VC you literally can't avoid running into AI startups everywhere, mostly building the "AI version" of some software. These can be either vertical software (law, shipping, finance, compliance, school accreditation, etc.) or horizontal (writing, presenting, researching, etc.) but the pitch is the same:
Take a vertical or horizontal
Point out the market size (wow 100 billion!!!)
Point out how ancient the software is (trash!!!)
So of course we should build an AI-powered version of it.
But IMO that's exactly the wrong way to go about it.
For one, every incumbent is doing precisely that: building an AI version of their existing software, and I hate to break it to you: they can do it just as well as you, AND they have more data, more distribution, and more engineers.
*Yes yes, every startup believes incumbents are just bolting on AI, while they are “true AI natives.” Frankly I don’t know what that means, and in too many cases we’ve seen, it’s just talking points without tangible differentiation to back it up.
Worse, in many cases there are competitors much scarier than the old guard - growth stage startups that have the weight and influence of incumbents, plus the agility and innovation of startups.
So the question is: if you’re a startup, what exactly is your edge?
And to frame the question for us investors, which “AI version of xyz software” should we think about backing?
Here’s my take:
AI is actually the ONE thing you shouldn’t focus on.
Google AI 產品大爆發，但為何只有 NotebookLM 是真創新？
November 24, 2025
Google NotebookLM 生成 slides & infographics 強大到可怕，但我反而思考一件事情：為何Google 做的很多產品中，（我認為）只有 NotebookLM 是真的 "category creator"？
Opal 明顯就是 Gumloop 跟 n8n 翻版、Pomelli 則是 Blaze AI 等行銷內容生成工具的翻版，有其他真正創新的東西嗎？（Mixboard looks interesting think I'll try it）
唯獨 NotebookLM 明顯開創了一個新的 category，跟之前的 notetaking tools 不太一樣，Why？關鍵是 Google Labs 的 Steven Johnson。他 2022 才年加入 Google，之前是個媒體人，都在研究 “where good ideas come from" 也甚至出了一本以這個為主題的書，還給了個很不錯的 TED talk。如果要總結他生涯研究的重點，就是「思考」及「靈感」的過程。
難怪他能夠創立 NotebookLM！
如果你整天想「打造出很棒的 AI 工具」，你應該做不出這種東西。但如果你整天都在鑽研思考的本質、人+人+機器一起 brainstorm 的過程等問題，才有辦法拆解這些基本的元素，並試圖重新想像，重組一個能協助人類思考，發想，並創作的工具。
想像未來，Gemini 有可能變成整個 Google 體系的核心：使用者可以跟 Gemini 對話，並直接從 chat 裡面啟動任何工作流程（生成一整份財務預測模型、更新一整份研究報告、撰寫 email 給1000個人）。
但另一個可能，是 NotebookLM 變成那個核心，因為 chat interface 用來整理你所有的素材跟筆記，說實在有點不大合適（我連昨天我跟 Gemini 對話的內容都找不到）。但 NotebookLM 介面就是用來組織你的思考的。就變成 NotebookLM 為中樞，可發動任何其他 Google 產品的工作流程，然後 Gemini 無所不在。
Running: Vibe-coding tools
December 11, 2025 (Updated)

Lovable

Using Lovable and its new "chat before you build" feature has a nice twist. It dynamically generates multiple choice questions for the user to sharpen ideas around what they'd like to build - instead of the old question lists that require typing long responses. I like it!

Running: No-Code Agent Builders
November 30, 2025 (Updated)
This is just me testing out various no-code AI agent builders. Latest thoughts:
AI will soon be able to wire everything for you: Barely a year ago, you had to configure everything yourself. Now AI can set up and tweak all the boxes and drop-down menus for you based on your natural language description of what you want to do.
Error rates in my experience are about 10%, which doesn't sound bad, but when errors happen, They're very hard to fix because you either have to prompt your way to a solution, which may take 30 min on by itself, or you configure the boxes yourself. And this means these tools do not go beyond the technically inclined population.
Running thoughts while using the tools

Google Opal

Conclusion: Same as any other visual programming builder, some neat design upgrades, but still impossible to fix when the agent makes mistakes 趁小孩睡覺玩 Google Opal，幾個很不錯的地方：😃 每一個步驟沒有一堆複雜的選單，下自然語言 prompt 就好。 右邊可以直接把你的 app 叫出來測試，協助 iteration 可是... 問題來了 ☠️，他們做這個的同時，Zapier 跟 Gumloop 早就都進化了（也回應每次 Google 一推出新產品大家就一窩蜂說新創都完蛋了）。 拿 Zapier 來說，今早嘗試用它來做 Gamma 簡報生成的自動流程，結果發現功能又躍進了： ⚡直接跟 Zapier Copilot 說你要什麼，它就把 zaps 的步驟先規劃出來，還幫你把每個步驟設定好，然在 chat 裡面給妳一個 “test" 按鈕，可以直接按下去來測試。 ⚡如果不 work，你直接跟他說看到什麼問題，它就跟你推論發生什麼問題，然後幫你 fix。不像 Google Opal 出了問題，你跟他說他還完全沒有反應。 *Google Brain 原團隊弄的 Lutra AI 一年前就已經弄出 self-healing 的 agent，自己會去看工作結果，判斷，並回去修復。一年前。 最後，seriously... 這種一堆框框線線的使用者介面，對大部分人來說都很困難，這麼簡單的 workflow 都已經眼花撩亂，但一堆 app 都還是只會拷貝這種介面，而不是從根本 (primitives) 來重新想像 building workflows 可以長什麼樣子。

Zapier Co-pilot

Conclusion: It can now set up almost everything for you, and because Zapier built an editing environment that predates gen AI, editing features are MUCH stronger than other no-code builder tools. Highly recommend 還不是 Zapier Agent 喔！那個更厲害 1.5 年前，用 Copilot 根本就是個廢物，只能大略把 zaps 的步驟叫出來，每個步驟還是要你自己設定那些選單項目（有時候幾十個！） 現在呢？不一樣了喔！你跟他說你要什麼，比如下面的，它就自己去設定咯 Whenever an email is received with the subject line containing keyword, please take the attached PDF file and create a Gamma presentation, and email it back to the sender. 如果出了錯誤，直接跟他說發生什麼事： Hey I realized that maybe Gamma API is not taking the PDF file as input, and instead seeing no data is just going with a dummy prompt, so perhaps we need a way to extract content from the file first? How do we do that? 然後它就直接幫你解決，然後在 chat 裡面給妳一個按鈕 “TEST"。

Airtable

Prompt: I want to build an app to help me track companies. Specifically, there are around 10 indicators I want to track and the input is: Company name, Website, Company registration or incorporation page By the way, all of these companies are in Taiwan. So all of them have company incorporation pages where you can look up their operating status. And so operating status is one. Another one is their registered capital amount and then another input is HR recruitment page profile where they'll list how many people work there, their main product technology or service and how many open positions they have to indicate their appetite or need for growth. It built the following in 2 minutes. Not bad. It has the right data structure and also an indicator timeline that seems to fit what I'm trying to track. However, I soon ran into an issue because I was hoping Airtable would help me pull data from the website that I gave it, but it does not yet have this capability. You need to plug in third-party browser data extraction tools in order to pull data into Airtable, and then you can use Airtable to essentially build dashboards.

變成 AI 的主管，會跟主管一樣煩躁
November 20, 2025
一直覺得那些在寫什麼 AI 代理人的，大部分都是騙人，但最近驚覺，如果你的工作需要處理大量資訊，搜尋，並把資料整理成如 Google Sheet 或 Airtable 等，那工作形態真的會完全不一樣。
以我們公司為例：創投不時就要篩上百家甚至千家新創公司的資料，現在的工作流程時常是這樣（而且是同時進行）：
⚡要篩選至少 1000 家日本新創，所以 Lutra AI 幫忙瀏覽每一家的網站並撰寫總結然後放到 database 裡面（重點是要找到很穩定，而且碰到任何公司的網站都會 work 的 prompt）。
⚡那1000家公司我想要一些額外的數字/資訊比如人才招募，但資訊散落所以 AI 無法很穩定的找到，而且一堆搞混的公司，所以請 Perplexity Labs + ChatGPT Agent 同時幫我用其中 30 家來測驗什麼 prompt 能夠取得品質比較好的數據，用 table 列出成果回報，然後給我一個 master prompt。
⚡同時美國同業推薦一家有趣公司，決定要約來 call 一下，直接請 Perplexity Comet 去徹底研究並撰寫成我們要的格式（這樣開會才有準備而不是笨笨的進去）。
⚡同時，因為要準備投資評估的簡報，在 Gamma 裡面製作，但很多細節要調整而且是在不同頁面調整（比如所有 $ 數字要 highlight，就請 Gamma Agent 去調，再回來看。
效率的確提升了，但新的挑戰是：一次要盯好幾個 AI 任務同時進行，常常心情變得很浮躁，尤其當它們同時卡住時，壓力反而更大。😡
而且，工作還會在「監督 AI 執行任務」和「密集吸收資訊」之間不斷切換。因為 AI 回傳的資料真的非常多，最後還是需要人一比一地看過，所以原本以為用 AI 會比較輕鬆，結果根本不是，反而變得超累。
I used to think that all the talk about “AI agents” was mostly hype. A marketing trick to make software sound more autonomous than it really is. But recently I’ve realized: if your work involves handling huge amounts of information—searching, sorting, structuring it into something like Google Sheets or Airtable—the way you work can change completely.
Take our team as an example. Venture capital often means scanning through hundreds, sometimes thousands, of startups. Here’s what a typical day looks like now (and yes, this often happens all at once):
⚡ We need to review at least 1,000 Japanese startups. Lutra AI scans each website, summarizes it, and dumps the output into our database. The trick isn’t just doing it—it’s finding a prompt that works consistently, no matter what kind of website it encounters.
⚡ For those same 1,000 companies, I want extra information—like hiring activity—but the data is scattered and messy. So I had Perplexity Labs and a ChatGPT Agent tag-team the task. I asked them to experiment on 30 companies, compare which prompts produced the cleanest data, and report back with a “master prompt” plus a results table.
⚡ Meanwhile, a U.S. colleague flagged an interesting startup we should call. Instead of diving in blind, I sent Perplexity Comet to research the company thoroughly and produce a report in our preferred format—so when the meeting comes, I actually know what I’m talking about.
⚡ At the same time, I needed to prepare an investment memo in Gamma. But formatting across multiple slides is tedious. For example: every dollar amount had to be highlighted. Rather than doing it myself, I let a Gamma Agent handle the cleanup, then reviewed its changes.
The efficiency gain is real. But here’s the catch: instead of doing the work myself, I’m now juggling multiple AI tasks simultaneously. And when they get stuck (which happens more often than I’d like), my stress level spikes. 😡
Even when they don’t get stuck, the workload shifts. I’m constantly switching between managing AI agents and absorbing the firehose of information they produce.
And that second part is brutal—because no matter how good the summaries are, you still need to go through the data one by one to really understand it.
I thought AI would make the job lighter. In reality, it just made the job… different. Sometimes, even heavier.
How I source AI-related investments
October 18, 2025
I joined Darwin Ventures in 2022.
I was a newbie and didn't have much deal activity going on, so I did lots of little things, e.g. writing the newsletters, updating the website, organizing LP conferences, etc.
And… since ChatGPT soon burst onto the scene, I became the one building AI-enabled, automated workflows for myself and the team to use.
  • Turning meeting notes into investment memo drafts
  • Turning verbal narrations into press releases
  • Auto-gathering portfolio news to post onto the website
  • And others
Building AI tools was really, really hard.
I don't have any technical background or coding ability. It was a slog. Hours banging my head against the wall using "no code" tools… Zapier, Make, Gumloop, etc.
But the result? I learned exactly what was possible (or impossible) for users like me
  • Where LLM models excelled (and struggled)
  • How to talk to users (our colleagues)
  • And best of all, I met lots of incredible founders.
Why? Because in order to build those tools, I needed lots of other tools. And to use those tools, I needed lots of help. Whenever I discovered a new, exciting tool, I'd schedule demo calls with their builders.
*Yes, lots of top AI startup founders have links you can use to book 30-min calls with them, right there on the websites.
I live in Taiwan. Most of these startups were in the US. So it was either post 11pm or 6am for me. But whatever, you do what you gotta do.
Later on, I realized that a lot of these founders had already raised tons of money, and didn't need VCs or even want to talk to VCs. If I'd emailed to say "hey can I see your pitch?", it'd have been crickets.
But they still talked to me, because I wasn't asking to invest. I wasn't asking them how much ARR they had, and which big shot VC had invested. I was just another person trying to build stuff and running into problems that we could solve together.
That's still what I do today.
  • Try lots of tools
  • Build useful stuff (well mostly not so useful).
  • Talk to their creators.
  • If I hit things off with one of them, and think they're simply amazing, I try to invest.