Making a Man in the Mirror

1. Digital Engagement

I was out grabbing a bite with a couple of mates of mine a couple of weeks ago when one of them asked me a pretty solid question — “Soban, is it possible to cook up an AI that sounds exactly like you?”

Of course, the answer is no — a computer can’t match my charm no matter how much data it’s fed. Can it get close, though? Absolutely. The results speak for themselves.

Here’s a quick video summary of what we cooked and how we did it!

If you’d like to see more content of this sort, subscribe!

After that fateful conversation, I had a couple of the boys at Antematter cook up a digital me. They were given free reign on how to do it, all I cared about was getting something in the ballpark of my behavior.

How’d they do it? Eliza.

2. I Remember You

No, I don’t mean the Eliza from the 60s that Weizenbaum cooked up. I mean ElizaOS, the ‘operating system for AI agents’.

If you pop open the documentation for it, you might see this diagram.

Let’s break it down in layman’s.

Eliza sets up an agent — a software entity powered by AI with the ability to interact with the digital environment around it. It achieves this through four components:

• Plugins let you hook your agent to AI-driven services you’re most fond of.

• Adapters let you store desired data your agent should draw from.

• Characters let you give your agent a unique persona it should maintain in its interactions.

• Clients let your agent interact through popular applications.

Now the great thing about Eliza is that setting it up is pretty easy — takes no more than fifteen minutes to get it up and running on Twitter / X!

But that’s just the framework — there’s still the question of how to get an agent to act like me in the first place. The answer will shock you!

3. 4D Chess

Turns out the folks in-charge of making a digital me wanted to test out a crackpot theory they were cooking — “what if the key to good prompting is gaslighting?”

They had seen some of their buddies use Eliza before — it’d work, but the text would be sniffed as AI-generated from a mile away.

To prevent this from happening, these guys decided to go the extra mile and gaslight an agent into digital engagement by first convincing it that it’s a pickup artist.

I know, I couldn’t believe it either.

In the dead of night, these guys grabbed a bunch of books on pickup artistry and seduction and then used it to outline the character file of my Eliza twin.

In particular, they settled on defining my twin first as a ‘coquette-ish male’ and then building a conversationalist off of that.

Bonus. Prompt Optimization w/ Ant-AI

In retrospect, I realize that these guys were tackling prompt optimization in an obscure way. Interestingly enough, we’ve already taken a crack at prompt optimization and even made our tool public!

Introducing Ant-AI, our custom prompt optimizer!

Tired of having to write and rewrite prompts, almost as if you’re at war with a psychotic infant?

Do you drown yourself in tears when your model breaks character midway?

No worries — we have the tool for you! Pop in your API key and jot down what kind of persona your model should embody along with the constraints. Ant-AI will go from there and cook up a set of system prompts that drown your sorrows away!

4. I Have No Mouth

Of course, there’s no point in having a character if it can’t even say anything. This is where you start chugging in API keys.

We experimented with a dime-a-dozen models — GPT, Llama, Qwen, you name it.

Unfortunately, the guys underestimated the effects of turning a pickup artist into a tech enthusiast — some of the models were more prone to flirt with random Twitter handles (No, I will NOT call it X!) whilst others were too hostile.

We settled on good ol’ GPT for the job — it had the right balance of poetic language (after all, who doesn’t like to sound impressive in a conversational setting) and puckish charm.

Took us a while to get there though, since each model reacts to system prompts very differently.

5. Touch — Suddenly Alive

I’m a simple guy — I wanted to automate my Twitter interactions. There’s a ton of things I find interesting, but being a founder means I have little-to-no time to sit down and write meaningful posts.

So, in this attention-starved era, I thought it’d be pretty funny to replace myself with an agent on Twitter. That’s essentially what the boys were meant to achieve with Eliza.

Of course, Twitter doesn’t make it easy — couple of the test accounts ended up getting the ban hammer the moment they stepped foot.

But a couple of swanky workarounds later, we got me on Twitter.

6. Money for Nothing

Of course, the astute among you might be asking — why on earth didn’t I just make a vanilla chatbot.

There’s a reason for that, dearies! If you’ve dealt with chatbots before, you might’ve noticed that the text it produces is…less than satisfactory, in polite terms.

Why’s that the case? Well,

• The text is a sloppy mess — it’d take too much refinement to get it up to snuff.

• The text is inauthentic — it lacks the contextual knowledge needed to maintain a facade.

• The text is unreliable — it’s much harder to control the hallucinations on the thing when there’s no systemic oversight.

So our reasoning for picking Eliza ought to make sense at this rate.

7. Knowledge is Power

To sweeten the deal, Eliza lets you inject your agent with a ton of knowledge. You can do this however you want, e.g. using a RAG pipeline.

What’s the point? Your agent gets to use it as the paint in its brushstrokes.

I popped in a small fraction of the reports I’ve read or cooked up into the Eliza agent, and it already sounds like me in the investor meetings.

Could you picture the power on this baby if it had much more to work with?

8. Me, Me, and Me

Of course, that’s not the end of the tale — rest assured I’ve got some of the engineers at Antematter locked up in my basement making my twin more and more like me.

Heck, agentic stuff impresses me to the point that I start to empathize with Pygmalion.

But enough of that, we have ourselves a newsletter to run.

If you were amused what we had on offer today, consider subbing!