A London-based startup has compressed the collective knowledge of human cooking into just 2 megabytes using an AI model, according to Decrypt. The team trained their system on 4.1 million recipes spanning seven languages, achieving a feat of data compression that rivals the file size of a typical MP3 song.
The accomplishment highlights advances in machine learning efficiency and the growing capability of compressed neural networks. Instead of storing massive datasets directly, modern AI systems can learn patterns and distill that knowledge into remarkably small models. This approach reduces computational overhead and makes deployment practical on resource-constrained devices.
The startup's recipe dataset encompasses global cuisines and cooking methods across multiple languages, suggesting the model captures cross-cultural culinary patterns. By training on such breadth, the AI likely learned ingredient relationships, flavor combinations, cooking techniques, and recipe structures in a generalized way. The compressed output can presumably generate or reference recipes without maintaining the original 4.1 million recipe database.
This development sits at the intersection of AI optimization and practical application. Smaller models consume less energy during inference, run faster on mobile devices, and cost less to deploy at scale. For food tech applications, a 2MB model could power recipe recommendation engines, dietary apps, or cooking assistants on smartphones without requiring constant server connectivity.
The multi-language training also demonstrates how compressed models can handle diverse inputs efficiently. Rather than storing separate models for each language, the single 2MB system learned to represent culinary knowledge language-agnostically.
While the startup hasn't disclosed specifics about accuracy or limitations, the project reflects a broader industry shift toward efficient AI. Companies increasingly compete on model compression rather than raw parameter count, particularly as edge computing and on-device AI become competitive advantages. The culinary application may seem niche, but the underlying compression techniques apply across healthcare, finance, and other domains where deploying powerful models on limited infrastructure matters.