The company
Built on research that pre-dates the LLM era
Inception was founded in 2023 to bring score-based generative modeling from academic theory to production serving infrastructure.
Why we started
The autoregressive architecture that powers today's language models was designed for modeling probability distributions, not for serving latency. Its left-to-right dependency chain is a mathematical constraint imposed by the training objective, not a serving requirement.
Score-based generative modeling offers a different formulation: model the joint distribution over all output positions simultaneously, refine iteratively, converge to a coherent sequence. The theoretical foundation existed before LLMs dominated the field. The gap was the engineering work to make it production-ready.
Inception is that engineering work. We are building the infrastructure to run diffusion-style decoding in production, starting with the API layer teams use to integrate generation into their products.
- Founded
- 2023
- Stage
- Angel-backed
- Focus
- AI inference acceleration through diffusion-style parallel decoding
- Team
- Stefano Ermon, Yuna Park, Marcus Dubois
- Headquarters
- Palo Alto, California
The people
Researchers and engineers who built the foundation
Stefano Ermon
CEO & Co-Founder
Stefano co-founded Inception in 2023 with a focus on translating generative modeling research into production inference systems. He drives product direction and works directly with early-access design partners.
Yuna Park
Head of Research, Co-Founder
Yuna co-founded Inception with a background in discrete diffusion models and parallel decoding algorithms. She leads research into refinement schedules, convergence criteria, and the token-posterior estimation at the core of the approach.
Marcus Dubois
VP of Engineering
Marcus spent several years building GPU-cluster serving infrastructure for large language models before joining Inception. He owns the serving kernel, the batching system, and hardware utilization across the inference stack.