Amarbot no longer has a WhatsApp number. This number now belongs to Sentinel, the custodian of Sanctum.

This number was originally wired up directly to Sanctum functions, as well as Amarbot's brain; a fine-tuned GPT-J model trained on my chat history. Since this wiring was through Matrix it became cumbersome to have to use multiple Matrix bridges for various WhatsApp instances. I eventually decided use that model on my actual personal number instead, which left Amarbot's WhatsApp number free.

Whenever Amarbot responds on my behalf, there's a small disclaimer. This is to make it obvious to other people whether it's actually me responding or not, but also so when I retrain, I can filter out artificial messages from the training data.

Ever since I discovered the Discord server for AI music generation, I knew I needed to train a model to make my voice a great singer. It took some figuring out, but now I'm having a lot of fun making myself sing every kind of song. I've tried dozens now, but here are some ones that are particularly notable or fun (I find it funniest when things glitch out especially around the high notes):

Amarbot was using GPT-J (fine-tuned on my chat history) in order to talk like me. It's not easy to do this if you follow the instructions in the main repo, plus you need a beefy GPU. I managed to do my training in the cloud for quite cheap using Forefront. I had a few issues (some billing-related, some privacy-related) but it seems to be a small startup, and the founder himself helped me resolve these issues on Discord. As far as I could see, this was the cheapest and easiest way out there to train GPT-J models.

Unfortunately, they're shutting down.

As of today, their APIs are still running, but the founder says they're winding down as soon as they send all customers their requested checkpoints (still waiting for mine). This means Amarbot might not have AI responses for a while soon, until I find a different way to run the model.

As for fine-tuning, there no longer seems to be an easy way to do this (unless Forefront open sources their code, which they might, but even then someone has to host it). maybe#6742 on Discord has made a colab notebook that fine-tunes GPT-J in 8-bit and kindly sent it to me.

I've always thought that serverless GPUs would be the holy grail of the whole microservices paradigm, and it might be close, but hopefully that would make fine-tuning easy and accessible again.

It's official — Amarbot has his own number. I did this because I was using him to send some monitoring messages to WhatsApp group chats, but since it was through my personal account, it would mark everything before those messages as read, even though I hadn't actually read them.

My phone allows me to have several separate instances of WhatsApp out of the box, so all I needed was another number. I went for Fanytel to get a virtual number and set up a second WhatsApp bridge for Matrix. Then I also put my profile picture through Stable Diffusion a few times to make him his own profile picture, and presto: Amarbot now has his own number!

In case the profile picture is not clear enough, the status message also says that he's not real. I have notifications turned off for this number, so if you interact with him, don't expect a human to ever reply!

As of today, if you react to a message you send me on WhatsApp with a robot emoji (🤖), Amarbot will respond instead of me. As people in the past have complained about not knowing when I'm me and when I'm a bot, I added a very clear disclaimer to the bottom of all bot messages. This is also so I can filter them out later if/when I want to retrain the model (similar to how DALL-E 2 has the little rainbow watermark).

The reason I was able to get this to work quite easily is thanks to my existing Node-RED setup. I'll talk more about this in the future, but essentially I have my WhatsApp connected to Matrix, and Node-RED also connected to Matrix. I watch for message reactions but because those events don't tell you what the actual text of the message is that was reacted to was, only the ID, I store a small window of past messages to check against. Then I query the Amarbot worker with the body of that message and format and respond with the reply.

This integrates quite seamlessly with other existing logic I had, like what happens if you ask me to tell you a joke!

Amarbot has been trained on the entirety of my WhatsApp chat logs since the beginning of 2016, which I think is when I first installed it. There are a handful of days of logs missing here and there as I've had mishaps with backing up and moving to new phones. It was challenging to extract my chat logs from my phone, so I wrote an article about this.