Inception Labs is a US startup, founded in 2024, that builds large language models using a diffusion framework rather than the autoregressive transformer architecture standard across frontier AI. The company emerged from stealth in early 2025 (Source: TechCrunch / Inception emerges from stealth with a new type of AI model).
Overview
Inception positions its work around diffusion language models, applying the diffusion framework used in image and video generation to text generation. This contrasts with the autoregressive transformer architecture used by models including GPT, Claude, Gemini, Llama, Qwen, and DeepSeek (Source: TechCrunch / Inception emerges from stealth with a new type of AI model).
Architecture
The diffusion approach is an alternative to the autoregressive transformer that has dominated frontier AI since the introduction of the Transformer in 2017. According to Inception, diffusion-based generation offers potentially faster parallel decoding, different compute tradeoffs, and possibly different quality and failure-mode characteristics relative to autoregressive models (Source: TechCrunch / Inception emerges from stealth with a new type of AI model).
Relationships
- related: Attention Is All You Need, The Bitter Lesson, Scaling Laws, A.I. Brainiacs (architectural-alternatives discourse).
Sources
- TechCrunch / Inception emerges from stealth with a new type of AI model