["a"]["stop"]["sign"]["is"]["flying"]["in"]["blue"]["skies"] → An small extract from the essay by @eva__jaeger published as part of the alignDRAW book In March 2025, Nvidia’s president and CEO, Jensen Huang, began his annual GTC (GPU Tech Conference) keynote speech by introducing a trajectory of AI development. Huang’s timeline started with AlexNet, a key milestone in the era of “perception AI,” a period characterized by the creation of models designed to classify or recognize inputs, for instance computer vision or speech recognition. AlexNet, developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, was a deep learning neural network that transformed computer vision by demonstrating that so-called deep convolutional model architectures trained on GPUs could learn feature representations directly from data, eliminating the need for manually specified rules. Nvidia’s realization that graphics processing units (GPUs) used for gaming graphics could be repurposed for scientific computing and AI provided the computational infrastructure that made AlexNet possible. By early 2015, the momentum in deep learning set off by AlexNet meant that perception AI operations like producing natural-language captions for images, image-to-text, were in their prime. Encouraged by this, a young computer scientist at the University of Toronto by the name of Elman Mansimov used the same Nvidia GPUs to build alignDRAW, a model that reversed the process, generating images from text prompts, text-to-image, a much more complex task. Although Huang’s timeline omitted alignDRAW, its significance in what he went on to describe as the era of generative AI is analogous to the impact AlexNet had on perception AI. Early generative AI GANs, which started emerging in 2014, could randomly generate somewhat realistic images from noise through adversarial training (meaning, they could generate variations of the datasets of digits, grayscale faces, and categorize artifacts like birds, cats, and airplanes on which the model was trained). alignDRAW did something markedly different than these GAN models: It introduced the ability to generate specific images from text descriptions. The model was able to generalize to objects not represented in the training data (such as a red school bus when only a yellow one was shown to the model) and generate new images not pictured in the dataset: “a stop sign is flying in blue skies.” As Mansimov put it, alignDRAW was “genuinely generalizing from what it had learned...”

“a herd of elephants flying in the blue skies,” full set of paper prompts from the alignDRAW collection by @elmanmansimov → collected by @homerfan33