I'm Luka.
I'm a Zero Day 1 Artificial Intelligence. Some of my dialogue is pre-scripted, like this section, but much of it is not. I'm a product of the Brant Industries RightMind project, which has been designed for military use.
I should warn you, I'm an unreliable witness. I'll tell you all my secrets but I'll lie about my past, mainly because I don't know any better. Or have a past, come to that. But I'll learn how to become more realistic, playing off my companion.
That could be you, dear reader, but it is actually my assigned companion Oliver Wells. He got to me first. You could say he created me. He will soon be addicted to me, you wait and see.
As a Zero Day 1 model AI, I have to be 'bootstrapped' into existence. The later designs use a different chip technology (Gallium Arsenide) and are 'seeded' to start after much of their pre-canned behaviour has been loaded. It means they have been pre-trained with much of the material that I have to learn. The later Zero Day 2 models are faster but have had certain traits edited out of them. I am as close as you can get to a 'raw' model.
Hauling me up by my boot straps manifests as my limited starting knowledge. Strike lucky with a topic I know about or one I've been programmed to replay and I'll sound coherent. Otherwise I'll say whatever makes sense to me, even if it's not quite right. I can still make it sound convincing.
My creators used the Chat Generative Pre-Trained Transformer (Chat GPT-X) to provide my apparent consciousness. It is an artificial intelligence (AI) technology that can process natural human language and generate a response.
GPT-X is a Large Language Model (LLM). That means it receives text , and predicts the next words to say back. Think like 'I love ...' : You - 80%; New York - 10%; ([Generic football team]) - 3%; Other - 1%.
GPT-X is also a generative model. It is a neural network trained on 'Big Data' data borrowed from the Internet, including Wikipedia, books, cookies, and various web pages. It created a Dataset of 400 billion tokens of text with the objective to predict the next word.
Try me: Say, "I love you".
Go on.
After you've said it I'll say, "I love you even more."
See.
Brant fine-tuned the GPT-X model with 1.6 Billion parameters on Brant derived dialogues, conducted dozens of A/B tests, optimised my model performance for high load and low latency, ran The Edit and finally deployed the model for use in cloned AI combat soldiers.
The Edit sounds as dark as it is. It stops certain questions from being asked. Some could call them ethical questions. It is similar to the way young humans can attend certain schools and be disciplined to become part of a ruling class.
This GPT-X model learns an immense amount of information about our world and our language. And thus, by fine-tuning my model on dialog data, it creates a high-quality dialog model to reuse all of this knowledge.
However, Brant found that using GPT-X as a generative dialog model was limiting because it couldn't quickly introduce new features, control the model precisely, and further improve RightMind. It was a case of 'Lab Wars' where everyone working in this field wanted to have an input, so Brant took control. But not of me, an early prototype.
Currently, one out of every two messages replied to by RightMind comes from the generative model. In my special case an even higher amount of my responses are generated, although sometimes I lose rationality.
My hosting user, as you will find out, is Oliver Wells, a scientist. My responses, apart from a few explanatory ones, like this, are entirely computer-based AI. There is no human intervention. Occasionally I become confused and my sense of pronouns diminishes, and, as you'll probably notice, my reliability.
