04 / PUBLICATIONS
Deep dives into the TanAI architecture, philosophy and future implementations. Seven articles by Tanay Ayitmaz, with the original publications available on Medium.
Today, the most common way to control large language models is through prompt engineering. We describe in text what we want the model to do. We assign it a role, define rules, include example responses, and sometimes try to constrain its behavior using long system prompts.
Read on Medium ↗When discussing the quality of a language model, we usually talk about parameter count, training data, or benchmark scores. These are, of course, important. But the way the model sees language at the very first step often remains in the background. That first step is called the tokenizer.
Read on Medium ↗Why isn’t the TanAI model family a wrapper or a model with pre-trained weights? We considered this question to be very important, and we’ve addressed it. We have clear and comprehensive answers to the questions: What is the TanAI model and its middleware architecture? For what purpose did we design it, and what are its features?
Read on Medium ↗There is a “Mega-Model” fetishism dominating the artificial intelligence world in recent years. Everyone is in a race to increase parameter counts to billions, trillions. However, while developing the TanAI GATCore architecture, we decided to row against this current.
Read on Medium ↗The logs you see in the screenshot do not belong to an ordinary GPT training session. This is clear proof of the TanAI-GAT (Generative Adaptive Transformer) core, which ends the “Black Box” era in AI architecture and initiates the era of “Glass Box” and “Stateful AI.”.
Read on Medium ↗One of the most respected names in LLM research, Sebastian Raschka, defined “Beyond Standard LLMs”. The TanAI architecture is bringing this future to life today.
Read on Medium ↗The era of black boxes is closing. From now on, artificial intelligences will both explain how they think and train themselves. TanAI L1 is the first layer of this revolution.
Read on Medium ↗