Long Time No See

Andi, Founder • June 5, 2024

The last Chatend blog post was July 22, 2023. What happened since then?

To recap, let me recite the elevator pitch for Chatend 2023:

Chatend aims to provide a simple and powerful interface for building and managing chatbots without requiring significant technical expertise.

In particular, Chatend would allow users to define commands as simple TypeScript functions, and then integrate those tools into a chatbot that could call them automatically. Additionally, Chatend would provide a key-value data store for storing data that persists across multiple sessions. This would allow users to build very useful chatbots that integrated into their daily workflow, using custom code to connect to e.g. Google Drive or Salesforce and work with any platform they could dream of.

This was a lofty goal, and one that I'm still fond of. From May to August of 2023, I built the MVP.

Chatend TodoBot

Chatend was the largest codebase I ever built. It was started on February 2, 2023 - my 24th birthday, and the day my resignation from my first job after college went into effect - and launched 246 commits later on July 22, 2023 on board the Amtrak Zephyr from Chicago to San Franciso.

LanguageFilesBlank LinesComment LinesCode Lines

Source: AllDaniel/cloc (Count Lines Of Code)

June 2023

I went to San Francisco with my friend and Spellcraft cofounder Christian Lewis to seek venture funding (me for Chatend, and Lewis for Spellcraft). My hope with seed funding was to build out a team that could turn my MVP into a production ready platform while I focused on finding the first customers. I was privileged enough to leverage my Twitter connections into meetings with several major VCs who were probably confused as to why I ended up on their calendar. Unfortunately, nobody bit on Chatend. My path forward was muddy and unclear (although to be clear, it was obvious from the start that it'd be hard for me to find funding). I'd have to sell the chatbot platform myself.

Sidenote: I want to thank Lewis for being as supportive as he was with Chatend, and for helping get me off the ground and survive in San Francisco. This is my solo founder journey but I would have lost right after I started without Lewis having my back. I also want to thank Varadh Jain and his roommate Shahan for giving me a place to crash in the city while I was still figuring all of this out.

There was just one minor issue: It's hard to sell a chatbot platform.

Most individuals don't want to put in the effort to build agentic chatbots, and even if they could do it easily, they wouldn't know what to use it for. Small organizations may want a chatbot, but it's difficult to trust a solo founder without an Ivy League degree and no connection to FAANG or AI labs (apart from meeting OpenAI's CTO that one time). Enterprise definitely wants chatbots, but they're going to buy from venture funded C corps, not a random dude with a .ai domain.

I tried cold emailing a bunch of local businesses, talking to mid level managers, and even offering to build a chatbot end to end myself, but nothing stuck.

I had used my solo founder time to build an impressive technical platform, but I hadn't scoped out the market and spent too much time building instead of talking to potential users. When I was done building, I couldn't picture the type of person that would want to use Chatend. I hadn't built something people loved. One month into launch, I had less than 50 users. Womp womp. Game over.

OpenAI Dev Day • November 6, 2023

Then, something interesting happened. OpenAI announced Custom GPTs and the GPT Store. A much better execution of Chatend's core thesis, brought to market by a billion dollar team of expert software and AI engineers with direct access to the underlying foundation models. Anyone could easily build their own GPT-powered chatbot now.

Chatend was cooked.

But then something even more interesting happened: GPT Store flopped. It's not nearly as popular as ChatGPT. It doesn't have hundreds of thousands of developers serving millions of users. It's unclear if anyone is making money from it, and it's certainly not competing with the App Store or Play Store. People don't know what to use it for. It dawned on me that the most popular AI company in the world was relearning the same lesson I had learned on a shoestring budget sleeping on a friend's living room floor less than two months ago.

In a weirdly twisted way, I was vindicated. GPT Store showed me that even if I had worked 1000x harder and sunk in 10000000x more dollars, it wouldn't have changed the history of Chatend.

That was when I realized I needed to double down.

The Pivot • December 2023

Selling shovels in a shovel rush Source: Turk on X

Chatend the chatbot platform was obviously hosed. But in the process of developing the platform, I had ended up with all of this reusable infrastracture in the form of, well, a platform, that I could build into something else. I figured I had built a very good shovel. If it was really so good though, why not use it myself to find gold?

That's Chatend 2024. I'm pivoting to building products, not platforms. I'm using the shovel to deliver gold to you, the users. I probably won't find gold the first time, but with enough shots on target I can increase my surface area until I hit something.

Probably the most important reusable component I built is what I call PanelChat. Here's a screenshot from the MVP:

Chatend PanelChat

PanelChat provides really important functionality that I think makes it the best GPT wrapper ever. It supports streaming LLM responses from OpenAI's GPT-4, obviously. But more importantly it supports Panels which allows me to build user experiences that live inside of PanelChat. A panel is like a window in a desktop operating system, and PanelChat is the window manager that allows the user to switch between windows and handles opening and closing them. Through panels, I can build any arbitrary user interface, like the image gallery panel demoed above.

Lewis described the first iteration of Chatend as "an operating system for LLMs" which is a good way to think about it and probably the tagline I should have gone with instead of "a chatbot platform."

That's useful enough, but where it gets really exciting is when we integrate panels with GPT-4 function calling. And that's what PanelChat does. PanelChat gives GPT-4 the ability to open and close panels, edit their properties, and send messages to them. GPT-4 thus becomes a Copilot for whatever interfaces I can integrate into it. This is the Chatend Assistant I dreamed of in my very first blog post for Chatend.

Panels can also send messages back to the model. This opens up agentic capabilities, where a model can query a panel for something, the panel can respond, and the model can act on that response, just like a human does when they type a question into a search bar. It's still early days, but I think this is the future of AI interfaces. Cognition Labs's Devin is a great example of how this concept can be used in practice.

So now I have the underlying technology behind all of my great future ideas: Chatend PanelChat. It took quite a while to get PanelChat to a point where I could build on top of it, but I'm finally there, 720 commits later. That's why it's been so long since the last post! I haven't given up on Chatend, I was just grinding until I had something impressive to show for it.

Chatend is dead! Long live Chatend. It's time to start building end user products that deliver real value to users.

LanguageFilesBlank LinesComment LinesCode Lines

Source: AllDaniel/cloc (Count Lines Of Code)

The Next Step • July 5, 2024

I'm releasing my first product built on top of PanelChat tomorrow. It's called Chatend Codegen. It's built on top of Groq Cloud's LPU inference engine for super fast token streaming and focuses on TypeScript. I built it to be helpful for generating repetitive code like forms or API clients.

You can sign up now and start using Codegen today. I hope you like it! But this isn't all; I have a lot more products in the development pipeline.

See you soon.

Codegen Announcement ›