Louis’s experiment centred on a fundamental question: Can AI accurately judge subjective culinary tasks, such as determining the perfect doneness of a steak? Preferences for steak doneness vary widely, from those who prefer a barely-cooked, pink-centred steak to those who insist on a well-charred, fully-cooked version. Louis aimed to train GPT-4V to navigate these subjective waters and provide accurate assessments.
The core of Louis’s experiment involved creating a ‘meat model’ trained on an extensive array of images and descriptions of steaks cooked to various levels. These images, ranging from rare to well-done, were stored on a Streamlit web server. Users could upload photos of their cooked steaks to this server, and GPT-4V would analyse the images, comparing them to its trained database to deliver a verdict on the steak’s level of doneness.
To make this tool more accessible, Louis set up the entire system on a Raspberry Pi 4. With this setup, anyone could snap a photo of their steak using a smartphone and receive a quick assessment from GPT-4V – like having a mini version of Gordon Ramsay in your pocket.
Understanding the risks associated with undercooked meat, Louis incorporated an additional safety measure. He employed an RP2040-powered sensor, specifically the W6100-EVB-Pico, to measure the steak’s temperature. This temperature data, transmitted to the Raspberry Pi 4, combined with GPT-4V’s visual analysis, provides a more reliable indication of whether a steak is adequately cooked and safe to eat.