business – Jing Conan Wang

Beyond the Hype: Three Lessons from a Startup Rollercoaster

My first startup journey was a rollercoaster – a wild ride that began with a spark of an idea.

Drawing from my experience at Google Brain, I had a strong instinct that combining human feedback using reinforcement learning would significantly improve the experience of LLMs in dialogue. I started to explore this idea to build an assistant for knowledge workers in 2021. My gut told me there was a big opportunity in the space, but I wasn’t sure when the mass market would recognize it. Later, it turned out that this was one of the fundamental ideas behind ChatGPT.

However, everyone said the space was tiny back then, and potential investors and peers saw little value in my concept. As a first-time founder, the unanimous skepticism I made me doubt my vision. I ended up wasting a lot of time and getting distracted by other directions.

I eventually pushed forward, though not fast enough. We launched in September of 2022, right before ChatGPT was released. When we launched, we were amazed by the positive feedback and delight we got from customers. However, within months, ChatGPT’s release completely transformed the landscape. Suddenly, customers began to have much higher expectations and hesitated to sign contracts.

It was obvious that we had to build more to make ourselves stand out. However, the process felt like a constant chase of whimsical hope. We tried to tackle different parts of our customers’ workflow. Customers told us they liked our AI capability but missed features they longed for from their existing vendors. We tried to build those, but existing incumbents would quickly add AI capabilities similar to ours, making our new solutions seem redundant.

I ended up starting a new startup in a completely different space. But I learned three crucial lessons from the first startup journey.

You need to trust your gut.

I wasted a precious year that could have helped us build a more defensible moat and establish ourselves as the market leader, better preparing us for when ChatGPT’s storm hit. A few months of lead time is not enough—you often need more. When everyone says your idea is impractical, it is the best time to build your competitive advantage. Innovation rarely comes from following the crowd. It emerges from the courage to pursue ideas that seem impossible—until they aren’t.

Unique insight means nothing without a defensible moat.

Consider your competitive moat early on. While talking to customers helps identify valuable problems to solve, it doesn’t guarantee that your solution will be the only one customers choose. You need to figure out why customers would pick your solution over alternatives.

Technology is not a silver bullet; it is actually a very poor moat because ideas diffuse naturally.

At the beginning, the only competitive advantage of a startup is the time you get because of the ignorance of big players. But you need to turn it into an actual moat — reasons why customers should use you, not other players. For B2B companies, the reason is often data or customer relationships. For B2C, the reason is often branding or better user experiences.

Plus, you need to make sure you have the resources to build your moat, which requires strategic planning if you are already in fierce competition. (Unfortunately, it often creates conflicts with the customer-first culture in B2B).

Don’t be afraid to restart from zero.

If you’re facing strong headwinds and haven’t had time to build a moat, take a step back. Rethink what other valuable problems exist that you believe in but others haven’t yet recognized.

My story isn’t unique – it’s a microcosm of the startup ecosystem. Innovation isn’t about having the perfect idea from day one. It’s about resilience, adaptability, and the willingness to transform setbacks into insights.

Ultimately, I started a new venture in a different sector, carrying these lessons like a compass. Each “failure” was actually a sophisticated learning experience, and helped me transform into a true entrepreneur.

How 500 Lines of Code Challenged a $500M AI Giant, and What Moats GenAI Startups Should Have

Recently, there’s been an interesting development in the big model industry. Perplexity AI, a hot big model company in Silicon Valley, completed a financing round two months ago, valuing it at over $500 million. However, using Lepton AI’s middleware, Lepton’s co-founder Jiayang Qing managed to create an open-source version with just 500 lines of code over a weekend, sparking a heated discussion in the industry. The related demo on GitHub quickly garnered five thousand stars in just a few days. This incident reflects a broader trend, and I’ll analyze it based on my own year and a half of entrepreneurial experience.

Currently, big model companies can be categorized into three types:

Base model companies, which primarily provide big model capabilities. This area requires significant capital and resources are highly concentrated among leading companies and giants.
Middleware companies, which offer middleware between big models and applications. Jiayang Qing’s Lepton AI falls into this category.
Application-layer companies, which directly provide consumer-facing applications. These can further be divided into platform-type application companies, like Perplexity AI, and vertical application companies focusing on niche markets, such as Harvey AI, which recently completed financing for its legal application.

The incident with Perplexity AI and Lepton AI highlights a pain point for application-layer companies — high competitive pressure with insufficient moats. For instance, Perplexity, which aims to solve general information search problems, faces challenges from four fronts: pressure from giants like Google, competition from vertical knowledge applications like Harvey, market encroachment from other knowledge service companies, and disruptions from middleware companies like Lepton AI. Vertical application companies face slightly less pressure, but they still confront these four forces and have a smaller market size, resulting in fewer resources.

So, what can be done? Many entrepreneurs believe that accumulating proprietary data can create a sufficient moat. This is very sensible, but it presupposes a systematic methodology for acquiring proprietary data. Here, I propose a methodology: using contrarian insights to gain a time advantage, leveraging the founder’s personal strengths for breakthroughs, and focusing on data-first products or operational capabilities.

First, contrarian insights, or insights not commonly understood, are essential. Entrepreneurs must find less trodden paths to build their core competitive advantage with minimal resources. What was once a contrarian insight can become common knowledge, such as the combination of big models with chat interfaces, which was novel before ChatGPT but is now common.

Second, the founder’s personal advantage is crucial. While contrarian insights offer temporary protection, they quickly become common knowledge once proven useful. Here, a deep understanding of user pain points in vertical applications can be a personal advantage. For example, one of Harvey’s co-founders was a lawyer. Even if a team doesn’t have a co-founder from a specific industry, previous experiences that can be leveraged as industry advantages are valuable.

Finally, building a data-first product or operational capability is key. The founder’s personal advantage must be systematized to sustain. There are two strategies:

Product-driven: The founder uses their deep understanding of user needs to design a product that naturally accumulates high-quality data, enhancing the product experience and creating a flywheel effect.
Operation-driven: The founder uses their resources and experience to build an operational system that continually acquires proprietary data, making operations or sales more efficient and faster.

The former suits products focused on Product-led Growth (PLG), while the latter suits those driven by Sales-led Growth (SLG). Both must prioritize data. If product-driven, each feature should contribute to data accumulation. If operation-driven, operations should focus on data, not just revenue or other metrics.

Returning to Perplexity’s case, Jiayang Qing could replicate Perplexity’s main functions over a weekend but not its data accumulation. As a middleware company, Lepton likely doesn’t intend this as a core strategy. However, many new application startups may use this to challenge Perplexity further. Whether Perplexity can withstand this depends on its ability to build a moat with proprietary data.