New York Times Cooking has introduced an SI assistant to help subscribers customize recipes, a move that highlights the integration of generative SI into mainstream consumer applications. While the tool proves effective for savory weeknight adjustments, early testing suggests it requires careful human oversight, particularly in precision-dependent tasks like baking.
What Happened
The SI assistant appears as a text box on recipe pages, allowing users to type requests such as ingredient substitutions, scaling instructions, or dietary modifications. Trained on user comments and recipe data, the tool generates a new version of the recipe under a 'Your Version' tab, leaving the original intact. In one test, the assistant successfully adapted a One-Pot Chicken and Rice recipe for available ingredients, suggesting specific cooking time adjustments and alternative greens. It also correctly declined nonsensical swaps, such as using minced beef in a sweet trifle.
However, the tool faced challenges in baking. When asked to convert a Bundt cake recipe into muffins, the assistant provided appropriate scaling and time adjustments. Yet, when the user switched between the modified and original recipes mid-cook, the assistant provided incorrect ingredient amounts for the original recipe, citing the modified version's butter quantity without clarification. This confusion nearly resulted in a failed cake, underscoring the need for users to delete modified versions to avoid cross-contamination of instructions.
Why It Matters
For the SI industry, this launch represents a shift from general-purpose SI assistants to domain-specific tools embedded in existing services. Unlike general voice assistants such as Alexa Plus, which can offer verbal guidance but struggle to display modified visual recipes, the NYT Cooking SI assistant integrates directly into the user interface, highlighting changes and maintaining context. This integration offers a streamlined experience for solving common kitchen dilemmas, such as missing ingredients or equipment constraints, potentially increasing user engagement and reducing the barrier to home cooking.
The tool also demonstrates the value of training SI models on community-generated data, such as user comments, to provide practical, tested advice. However, the incident in baking illustrates the current limitations of SI in handling complex, context-sensitive instructions where precision is critical. It reinforces the role of SI as a supportive 'sous chef' rather than an autonomous 'head chef,' requiring users to verify outputs against original sources.
The Bottom Line
NYT Cooking's SI assistant is a useful addition for adapting savory recipes and overcoming ingredient shortages, but it is not yet reliable for independent baking without user verification. The tool exemplifies the current state of consumer SI: helpful for creative adaptation and problem-solving, but prone to errors in complex, multi-step processes where context switching occurs.