All of human cooking compressed into 2 megabytes
Researchers have released a dataset and model that claim to compress global cooking knowledge into about 2 MB by mapping millions of recipes into roughly 1,800 standardized ingredients and their flavor relationships. Commenters find the flavor-network idea and its demos useful for things like ingredient substitution, pairing suggestions, and structured recipe representations, but argue that the title is clickbait and the coverage is incomplete, with major gaps in African, Middle Eastern, and some European cuisines. The thread also branches into broader questions about how far cooking can be formalized, the limits of AI-generated recipes, and the cultural and sensory richness that may be lost when food is reduced to data.
Overall Reception of the Paper & Title
- Many find the technical idea interesting but strongly criticize the title “All of human cooking compressed into 2 MB” as misleading and clickbait.
- Commenters note the paper focuses on ingredients and their relationships, not full cooking techniques, procedures, or proportions.
- Several people say the inflated title reduces trust in the work, despite the underlying dataset being “cool” and potentially useful.
Scope, Coverage, and Biases in the Dataset
- The corpus is from 11 sources and a limited set of languages; commenters argue this cannot represent “all” human cooking.
- Missing or underrepresented areas mentioned: African cuisines, Arab/Middle Eastern, Indian/South Asian, some Southeast Asian, and non‑translated French/Italian sources.
- Others point out that English-language recipes for these cuisines do exist, but likely aren’t fully representative or authoritative.
- There is concern that non‑English ingredients were machine‑translated, introducing ambiguity and error.
Ingredients vs Techniques & Compression Claims
- Multiple comments stress that the model captures ingredient co‑occurrence and flavor compatibility more than real “cooking,” which depends heavily on technique and ratios.
- Some argue the space of ingredients and techniques is actually small enough to compress; others demand empirical proof via taste tests and stress that “crib notes” aren’t true mastery.
- People highlight that subtle technique (e.g., fried chicken variations, stew timing, use of acid) is crucial and often missing from algorithmic approaches.
Applications: Flavor Pairing, Substitution, and Tools
- The dataset is seen as promising for:
- Exploring flavor pairings and ingredient embeddings.
- Suggesting substitutions or next-best ingredients.
- Interactive tools such as flavor maps and recipe generators (several linked demos and side projects).
- Some see this as groundwork for specialized cooking models; others think generic LLMs are already “overpowered” for cooking if prompted well.
Cultural & Human Concerns
- Several comments object to omitting major culinary traditions while claiming universality.
- There’s ambivalence about automated or robot cooking: some are excited, others see cooking as core to human culture and creativity and feel automation “robs” something human.
Related Visual & Structural Representations
- A tangential but lively subthread praises schematic/graph-based recipe representations and flowchart-style cookbooks as clearer than traditional prose recipes.