Day 1 experience with the AI Twin as a closed beta user

Training an AI twin actually takes patience. I found that out while testing Tsuin, which builds an AI twin that learns how you think. I’m one of the closed beta users, so part of my job is QA and giving the team feedback on the experience, the good parts and the frustrating parts.

What I expected

At first I thought it was a regular chatbot. I expected it to ask me questions about a topic, discuss it with me, and learn as I corrected its understanding.

I asked for the rules first

Before I start anything, I like to know the dos and don’ts and the limitations. That way I get the big picture before working on the details, and I avoid training the wrong thing and spending time fixing it later.

So my first message asked the twin for the best way to record and learn from me. It replied with a long list of general tips, like being specific and giving context.

Testing whether it makes things up

Next I tested whether it would make up facts about me. One of the biggest issues with AI is hallucination. For example, when I asked Gemini what James’s birthday is, it told me about LeBron James.

I asked the twin the same question, and it said I don’t have enough from James on this yet. Honestly, that was better than I expected.

Then I gave it a fake birthday on purpose, 9 September 1999. I wanted to see if it would remember the fact, and if it would update it when I later told it something that conflicts.

Some of my answers didn’t save

When I told the twin my birthday, it replied I don’t have enough from James on this yet again, which was confusing.

Later we found another issue. If my answer had a ? anywhere in it, the twin read it as a new question. It gave the same reply, and my answer wasn’t saved for training.

Now I take out every ? before I send an answer, so even ?! becomes !.

It waits to learn from me

By then I started to question whether the twin knows anything at all. I even asked it what 1 plus 1 is, and it said it didn’t have enough context.

That’s when I realised the twin waits to learn from me first. A regular chatbot would just give me an answer.

There’s no user manual yet

I had no guide on how to use the twin. I didn’t know when it trains.

When I asked the team for a list of commands, they said it isn’t ready yet, since they’re still working on the user experience.

Training doesn’t happen in real time

Then the admin told me training doesn’t happen in real time. Nothing I had asked or answered so far had been trained yet. The twin only learns when a training run starts, and you can’t use it while it trains.

The admin also said to stop asking it to tell you back what it learned.

The admin’s advice was to focus on teaching the twin about you, and to think of it as cloning yourself.

Answer why questions

Once I knew that, I figured the best way to train my reasoning is to answer why questions. For example:

Writing the why questions myself got troublesome. So I asked Claude to generate the questions, and I typed my own reasoning for each one.

One answer won’t stick

After my first training, which took around 10 minutes, I tested the twin on what I had just taught it. It didn’t remember the facts I had shared, and that confused me at first.

That’s when I learned it takes a lot of examples. Most of the time, the twin trains on your reasoning about certain things, so a fact you share once won’t just stick. According to the admin, it takes around a thousand examples from you before the twin performs like you.

What I’d tell myself on day one

The twin starts in its pre-school phase. Give it enough examples of your own understanding and reasoning.

• Focus on building more examples, instead of training it as a fact retainer.
• After you train enough examples, around 1,000, you can expect it to start becoming useful.
• Keep ? out of your answers, or they won’t be saved.
• Ask the team when you’re stuck, because the commands aren’t written down anywhere yet.