The AI Agent Architecture: Going Beyond Transformers
GPA-1 (Generative Pre-trained Agent) is a neuro-symbolic foundation agent model, surpassing GPT-4 by 100x in quality, capabilities and safety of AI Agents.
Join the wait-list
About GPA-1
GPA-1 (Generative Pre-trained Agent) is a foundation agent model, vastly outperforming GPT-4 in both safety and overall performance. GPA-1 Agents offer unprecedented control over their behavior, precisely aligned through human feedback in key interactions. They are capable of working on behalf of any organization - from retail and airline sales to insurance, healthcare, and government services - and serve their customers and users responsibly and safely. With complete control over agent’s actions, behaviors, and specific do’s and don’ts, GPA-1 effectively addresses the safety concerns in commercial AI deployment.
GPA-1 does not rely on a transformer for its reasoning. Its novel architecture, replacing traditional transformers with a neuro-symbolic approach, harnesses the strengths of LLMs while resolving control, reliability, and safety challenges with a system of symbolic parameters and conditional rules. GPA-1 possesses advanced situational awareness, relying on a network of sensors to identify the state of each user-agent interaction. This enables precise alignment of agents' actions and behaviors with desired outcomes through human feedback, facilitating continuous, real-time enhancements.
Developed proprietarily by AUI™ from 2017 to 2023, in collaboration with 60,000 human agents, GPA-1 marks a new phase in the evolution of controllable, safe and practical AI agents.