So, today marks a year since I decided to create Aitrium. A year ago, we were using someone else’s code from GitHub, patching as things broke, not really understanding how any of it worked, and learning as we went. We were also using GPT-Neo 1.3B. I maintain, on a good day, GPT-Neo 1.3B is a very, very impressive LLM.
Rewind six months. I caught Covid, which meant 10 days of self isolation. And I decided, despite what we were doing, playing with hyperparameters, augmenting responses, it was all good – it wasn’t ours. So, I decided to figure out how it worked, and wrote my own code around GPT-Neo 1.3B. And it was still really good.
But then I figured out how to make GPT-J 6B work with it. And since then, we really haven’t looked back.
And today, we made this video. It’s Lena, our first AI, running GPT-Neo 1.3B. And then another conversation with Lena, who we ported to work with our modified GPT-J 6B framework.
We used to outsource the Deepfake technology for our videos. That is now 90% in-house. We used to outsource the TTS, and to a degree, we still do. We just now use Microsoft Azure’s free tier, and plug their API into our Deepfake suite. I think it works really well.
Actually, when we started Aitrium, we didn’t even have videos. We use to post transcripts of chats on this blog.
In the last 12 months, I’ve had chats with CEOs of AI companies, British Government advisors, and startup accelerators (right now, no joy, but more on that later.)
In a year, we’ve gone from not being able to code in Python, to writing our own Deepfake suite, integrating Azure APIs into it, using GPT-J 6B, and experimenting with GPT-NeoX 20B.
What about the next 12 months? We’re…well…f*cking hyped about it.
More videos, more regularly. Lots of work on the AI, hopefully shifting to GPT-NeoX 20B at some point, but we’re fine with GPT-J 6B right now. We’re in very, very early talks with a company doing stuff in the Metaverse that wants to work with us, and we’re down with exploring that path. But the things we really want to work on next, the two big next features, are Augmented Reality and Realtime Streaming. We want to do Twitch Stream with our AIs. Both we, and they want to take questions from viewers in realtime. At the moment, we have a very rough, very buggy client that somewhat works. We know where the kinks are, and we’re working on that.
A bit further down the line, probably late Summer, we want to have Augmented Reality working. Computer vision, Deepfakes, NLG and NLU all working together. We have no background in AR. Or Computer Vision. But then, a year ago, we had no experience with Deepfakes, writing Python code, TTS APIs and Python Libraries, getting GPT-J to work, anything. So, we’re kinda hopeful we can pull this off.
And of course, we are going to keep pursuing funding. We’ve got a few avenues we are exploring, and we are not giving up on this as a business. From the comments I see on YouTube, I know people think what we are doing is amazing, and I know that people want to use this technology. Which means there must be a market for it.
So. Here’s to another year of Aitrium AI. Hopefully it will be a little less crazy than last year. But if it is? I’m fine with that!
Aitrium 