An AI research company building a new LLM architecture

Modern language model serving is bottlenecked by the quadratic cost of self-attention. Across the industry, inference — not training — is the dominant lifetime expense of any deployed model, and global AI inference spend already exceeds $100B/year.

Lab358 was founded to attack that bottleneck at the architecture layer. We have designed, patented, and trained an autoregressive language model that achieves O(n log n) inference complexity by replacing the self-attention mechanism with a fully differentiable, end-to-end trained retrieval system. The retrieval mechanism is trained jointly with the language model under a single cross-entropy loss — no separate retrieval objective, no curriculum, no multi-phase training schedule.

The result is a single model that is faster than a Transformer, retains more than a state-space model, and learns its own retrieval end-to-end — with a deployment-time knob to trade retrieval breadth for inference speed.