NYT Cooking's new AI assistant will reliably get dinner on the table, but don't ask it to bake a cake. That one insight should guide how you think about integrating LLMs into creative or precision-driven workflows.
I've been testing the assistant integrated into the NYT Cooking app and website. It's a chatbot that can answer questions about recipes, suggest substitutions, and help you plan a meal based on what's in your fridge. The core model appears to be a fine-tuned version of GPT-4, trained specifically on the NYT Cooking recipe database and user questions.
Where it works: the weeknight scramble
The assistant shines when you're staring at a chicken breast and a half-used jar of salsa. I asked it "What can I make with chicken, salsa, and cheddar?" It immediately pulled up a recipe for chicken tinga tacos and offered to walk me through the steps. It correctly identified the salsa as a marinade base and suggested adding cumin and lime from the pantry. The response was concise, actionable, and saved me from scrolling through thirty recipes. For pragmatic, substitution-based cooking, it's genuinely useful.
Where it fails: precision baking
Then I tried to bake a cake. I asked about swapping buttermilk for regular milk in a classic yellow cake. The assistant confidently said to use a 1:1 ratio with a tablespoon of lemon juice. That's correct for a quick substitute, but it didn't ask what kind of cake I was making. For a dense chocolate cake, that works. For a delicate sponge cake, the extra acid can throw off the crumb structure. The assistant gave a generic answer when the recipe context mattered. The lesson: LLMs are terrible at tacit knowledge. A baker knows that a chiffon cake's structure depends on exact pH and protein content. An LLM trained on general cooking knowledge will give you a safe, but often suboptimal, answer.
The takeaway for developers
If you're building a cooking assistant, treat it like a sous-chef, not a pastry chef. Use it for meal planning, ingredient substitution, and step-by-step guidance. Don't rely on it for precision baking, fermentation, or any recipe where small variations in technique or chemistry matter. The NYT Cooking assistant is a great example of a focused, domain-specific LLM application. It succeeds because it stays in its lane.
Source: The Verge - NYT Cooking’s AI assistant saved my weeknight chicken, but almost ruined my cake
Source: https://www.theverge.com/tech/1029489/nyt-cooking-ai-assistant-review
Top comments (0)