How a Funny Bot Became Our AI Workmate

Before Gepeto, a friend made a Discord bot called Marko. The name came from Markov chains, which is what he actually used to make the bot use our past messages to send new ones with the words we used the most.

The messages were basically random, and most of the time it said nonsense, but sometimes it would reply with something incredibly funny.

This was before LLMs and ChatGPT were publicly released. I loved that bot. We had so much fun interacting with a chatbot, taking screenshots of the funny moments, and laughing at the weird things it would say.

Based on that idea, I created another version using LLMs, which we would call Gepeto in reference to GPT, the technology underneath this new bot. I wanted to make a robot that could actually interact with us, remember things about us, remember more of our conversations, and send more meaningful messages.

It was a complete success with my friends because we tweaked it so it could actually talk like us and be funny.

That was the first version of Gepeto, and it lived in a Discord channel.

At the time, Gepeto was literally an experiment. We mostly made it say stuff and be funny.

But very quickly, people started wanting more things from it.

At first, Gepeto couldn't even read images, so we added that. Then we added the ability to react to messages, edit its own messages, and search the internet.

We didn't really have some big product roadmap. People just wanted those features, and once we added them, they started finding more interesting ways to use Gepeto.

Some friends would share images and ask it for photography tips. Other people would just send memes to see how Gepeto reacted. Sometimes people would be in the middle of an argument or discussion, mention Gepeto, and ask it to look something up or explain a concept to someone—like what Grok is now, just long before it!

From this experiment, we realized Gepeto was actually really useful to have around.

It was basically ChatGPT, but with friends, which was very novel at the time.

Intelligence where you already work

From those first experiments, I realized the importance of having an agent in the places where you already work and interact with other people.

Of course, I have used AI a lot to code by myself, study by myself, research by myself, or test ideas and hypotheses alone.

But if you add that intelligence to a group chat, you can interact with other humans to ask bigger questions, have more insights, and do more ambitious work.

The intelligence is there with you to actually get things done, help you out, or move the project forward.

There was eventually an idea of turning Gepeto into a more work-focused project, but initially, because Gepeto was so fun to interact with, I simply wanted it to be part of my family and friends' group chats.

The use case of simply having it there, being able to make it remember things, say funny things, or browse the internet was useful enough that I thought: we should have this everywhere.

I started adding it to my family group, and later to my university class group because I wanted it to remember our schedule and remind us when we had important assignments to deliver.

And then we started doing something much more ambitious with it.

Gepeto at Nightingale

I made a separate version of Gepeto specifically for Nightingale.

Instead of the version running with my friends on Discord, this version had a lot more tools and much more access to the systems we were using.

It could operate Bash tools and update an SQLite database. It could browse the internet and interact with a Baileys integration so it could read information from the Amber Alerts WhatsApp channel which we were analyzing for our research. It could also use Apify to access Instagram comments and posts.

This meant Gepeto could actually do some of the manual work we previously had to do ourselves at Nightingale.

We could give it information, screenshots, posts, or places where information was being published, and use Gepeto to help us gather, structure, clean, and maintain the data we needed.

That was when Gepeto started becoming much more than a chatbot.

Because at Nightingale we have been doing more and more ambitious things, it has become really important for us to have the right systems and reminders so we can actually do our homework, study for classes, and at the same time build these huge projects at Nightingale.

We use Gepeto for both.

We use it to organize ourselves, manage reminders, and keep track of our academic life.

And we use it to keep track of our professional work at Nightingale.

Gepeto can do much more now

The models themselves have progressed so much in just a few months.

At the same time, we have done a lot of work to make Gepeto cheaper and able to run for much longer at a lower cost.

We have been able to make it build things for us, manage more documents and much more context, and remember information for much longer.

We now give Gepeto a computer so it can actually be doing things in the background.

Gepeto can do many of the things other agents like Codex, Claude Code, and similar tools can do because it has access to many of the same kinds of tools.

That means it can be so much more efficient than those first versions.

It's no longer something you only interact with that happens to have some integrations with your calendar and works with your friends and family.

Making Gepeto useful to talk to

We have also worked a lot on something that sounds simple but has turned out to be extremely important, which is making sure Gepeto is someone you want to keep around.

Because the first versions of Gepeto live inside WhatsApp, you don't want it sending you a huge message with tons of context tangled together with complicated formatting. We want Gepeto to understand what information is actually important for you at that moment and send it in a very summarized, scannable way.

This is one of the best things we have done. You notice it especially when you ask a question that requires Gepeto to go and research a lot of documents.

It can go through all of that information and then explain it very simply to you while also giving you the sources where it actually got that information.

Cheap, capable, and proactive

There are already other agents and solutions that are extremely useful.

What we're focusing on with Gepeto is making it really cheap and really capable, so you can actually afford to have Gepeto working across your team and use it to do much more work.

We're also focusing heavily on making Gepeto proactive.

We want it to understand you, know what is important to you, and be something that is very easy and engaging to interact with.

We don't want Gepeto to feel like another AI that talks like an AI.

We want it to behave more like a really good workmate and have a personality that fits and understands your culture.

That idea actually goes all the way back to those first versions of Gepeto with my friends.

Being capable matters, but being something that people actually enjoy having around matters too.

From an internal tool to a Nightingale product

We built Gepeto because, as a small team, it was important for us to find a way to leverage technology to give us a huge boost in productivity and capabilities, so we could actually achieve all of the ambitious ideas we have.

For a long time, Gepeto was something we were building for ourselves.

Now we're turning it into a Nightingale product.

Later this October, we're going to start opening a limited preview of Gepeto so other people can actually use it.

During the preview, you'll be able to try Gepeto for free for a limited period of time.

You'll be able to use Gepeto directly from WhatsApp and Slack, where it can work with the tools and information that are already part of your work and life.

This is still something we're actively building and improving, which is exactly why we want people to start using it and giving us feedback.

You can learn more and sign up at heygepeto.com.

We built Gepeto because we needed technology that could help a small team like ours achieve much more ambitious things.

Now we want to see what other people can do with it, while we keep improving this technology infrastructure that can help all of us achieve even more.