DearFlow (Techstars '24)

Ben and I are together in person again. Excited for what’s next!

Today marks 1 year since I first met my close friend, also my co-founder, Ben. I still remember vividly that it was an afternoon when I was reaching out to people on YC Cofounder Match, hoping to find a technical co-founder for my new product idea. I then came across a really impressive guy and decided to send him a message (though I thought he might never reply, but let's just try anyway!) To my surprise, he replied in just 3 minutes. We then had our first meeting 30 minutes after that and started working together the next day. That’s Ben, and that’s how it all started. Fast forward one year, and so much has been built, changed, and grown at DearFlow. One thing that still remains, though, is how much I enjoy working and growing together with Ben. He is smart, driven, empathetic, and most importantly, a super nice human being. I don’t know what is waiting ahead of us, but I’m so grateful to have Ben by my side. Thank you so much, Benjamin Drury. Here’s to many more years to come! 23:45 PM 30 June 2024 Ho Chi Minh City

What does it take to make GenAI useful for Businesses? When I first started building GenAI products for businesses more than a year ago, I was driven by how impressive the GPT-3 model was at the time. It was a very typical solution-first approach at best. However, as I delved deeper into the journey, I realized the most important thing about AI for business isn't how advanced it is, but how useful it is. And to be useful, AI needs to be three things: reliable, affordable, and easy to use. This seems obvious, but it's not how most people think about AI at the moment. Most people get excited about new AI capabilities. But capabilities are only half the equation. We learned this lesson while building DearFlow. We were inspired by multi-agent frameworks like Autogen and CrewAI. These are impressive technologies, but we realized they weren't ready for business use, at least in the near future. Businesses need reliability above all. If you're automating a crucial business process, it can't fail. Ever. A fun demo is one thing. A system that runs your business is another. The surprising thing we discovered is that to make AI reliable, you often have to use less of it. Our typical AI workflow on DearFlow is only about 10-20% AI, the rest is ordinary software infrastructure. The other crucial factor is cost. AI doesn't just need to be cheaper than humans. It needs to be much cheaper. If it's close, businesses will choose humans every time. Humans are more flexible, can do multiple jobs, and contribute to company culture. An AI system needs to be so cheap that it's worth the hassle of setting it up. Finally, there's ease of use. Some people may disagree with me on this, especially the very technical ones who know it all, but I think it's no coincidence that all of the most popular AI tools now are chat interfaces. Yes, they can be inconvenient, not scalable, and not customizable, but the key is how easy it is to use a chat box. And that is enough for millions of people to use it daily. But chat interfaces have their flaws, and I’ve started to see many more apps move away from them to build a dedicated UI for specific use cases. This may not be very obvious now, but I believe when all the foundation models get “saturated”, the main differentiation between AI apps will be how easy and delightful they are to use. This is a very interesting topic, I may write more on this later! So, what do you think? What makes AI useful in your opinion? Let’s discuss! Image: Pawel Czerwinski

Today, I found a fascinating way to think about “Tokens” in LLMs. It came from a conversation between Jensen Huang and Patrick Collison at a recent Stripe event. When asked about the future of compute capacity in five years, Jensen smartly dodged that question a bit, but then gave an excellent explanation and analogy about why tokens are the new forces that will power the next decades of humanity. He explained that we're now producing something unprecedented (and at scale): floating point numbers that possess “value”, which we now call tokens. These tokens are valuable because they encapsulate “intelligence”. People are now taking these tokens and transforming them into English, French, images, videos, chemicals, proteins, robotic movements, etc. And many are working hard to expand the range of concepts and ideas we can create with these tokens. He then goes on to made a compelling comparison between tokens and electricity. In the past industrial revolution, we successfully found a way to convert “atoms” into “electrons” (by boiling water to power electricity turbines). And now, we've discovered a way to convert "electrons" into "tokens" (by using energy to power data centers that train and run LLMs). When electricity was first introduced, few people understood its value. Today, paying for kilowatts is routine. The same will happen with tokens. Right now, only early adopters and builders are paying tokens. Soon, everyone will be paying for tokens on a daily basis to supercharge productivity and power new products and services. Many new industries will be born from and built on top of tokens. When I first heard about this way of thinking, I got goosebumps. Perhaps it’s because Jensen has a good storytelling skill that helps him sell what he builds, but I have to admit this way of thinking gave me a brief surge of pride. I'm proud of humanity's collective effort and creativity, which have taken us from living under rocks to building machines that can "think." That's nothing short but a miracle.

Lately, I've been thinking a lot about what work will look like in the near and distant future. As I explore the agentic workflow rabbit hole, I've noticed things moving at a surprisingly fast pace (which, when you think about it, isn't all that surprising ;) Fully autonomous agents that are reliable and consistent for complex tasks are still a ways off, but their arrival seems inevitable. This leads to an interesting question: as we automate more work, what will be left for us to do? What new roles will emerge? I don't have a definitive answer yet (hopefully, I will soon and can write a post about it), but I've realized there's an “interesting” problem we need to tackle first. As we delegate more work to AI, I suspect we'll experience an overwhelming period of constant change and iteration in our work approaches, both individually and within companies. It sounds counterintuitive, doesn’t it? AI is supposed to simplify things for us. Let me explain. At first, we might just use ChatGPT to draft a marketing post. Next, we might have an AI workflow that generates drafts, improves them, and posts them across all platforms simultaneously. Soon, we’ll have an "agent" to automate most of the process with human supervision. Eventually, we'll have a team of agents handling everything from start to finish. And there is the potential for many stages that we don’t know yet in between. Each stage requires a different approach to work, both technically and mentally, I believe, as we shift from human-human work to human-AI work and then to AI-AI work. The later stages always seem better productively, but early adoption is tempting (for people like me) or if your competitors choose to do so and you do not. So, the next few years will be a period of constant evolution in our work practices as this technology advances. Whether this is overwhelming or exciting is up to you. But one thing is certain: there will be a lot of internal meetings on how to use AI ;) What do you think? This challenge isn't just for AI users, by the way, it’s also for AI builders, especially those working at the application layer. But that’s a topic for another post. Quick update: DearFlow just closed its first enterprise deal recently. Benjamin Drury and I are thrilled to have made this progress. We look forward to building an exciting tool for the future of work with AI.