Intelligence only emerges from the interaction of specialized units — the brain contains neurons, organizations contain members, and LLMs contain transformers and MLPs.
· Our bet is that the next type of intelligence will emerge from self-coordinating, super-fast, and small AI agents.
Small agents are not limited to single LLM models — sometimes they are composites of LLM models or multi-stage layering techniques.
· At ProxAI Labs, our first research area is an automated and semi-automated agent training platform. We are developing distillation models optimized for Speed (under 200ms), Size (~350M parameters), and Robustness (>0.95 AUC).
· We are using these models to create composite structures to achieve extreme performance according to speed, scale, and robustness.
Agent self-improvement extends beyond reinforcement learning and architectural gradient updates. It ranges from finding the correct inputs from other agents to optimizing the kernels.
· We focus deeply on the economics of this performance improvement — our pipeline accounts for every layer of optimization, balancing trade-offs to maximize cost-efficiency.
Fact-1: Neural network architectures and optimizations have been well-studied over the last few decades.
Fact-2: Orchestration tools for microservices have matured after extensive battle-testing in the real world.
Hypothesis: The orchestration of multiple small agents is a very new area and will be the next big problem for the upcoming decades. The architectures, hierarchies, and objective functions governing agent swarms are still in their infancy.
· At ProxAI Labs, we are building our entire platform with this research and optimization flexibility from day one.
Our foundation is built by applied researchers from Google Brain and DeepMind, bringing over seven years of dedicated experience in small distillation models and high-speed AIs that moderate other AIs.
From securing medals at the International Olympiad in Informatics (IOI) to competing in the World Universities Debating Championship (WUDC), we recruit elite, competitive talent.
The world's hardest problems are solved by smart, intensely passionate teams moving at extreme velocity.