Alec Radford

Overview

Alec Radford is a machine learning researcher who worked at OpenAI from 2016 to 2024 and is best known for the GPT line of language models. He was also first author of papers on image and speech models, including CLIP and Whisper.^[1][5][6]

Education and early career

Radford attended Olin College of Engineering and left in 2014 without graduating. While there, he co-founded the startup Indico with fellow students.^[1]

Key contributions

Radford was the first author of the DCGAN paper, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,” published in November 2015 with Luke Metz and Soumith Chintala.^[3] A generative adversarial network pits two neural networks against each other: one generates images, while the other tries to tell real images from generated ones. The title describes the goal of the work, which was to learn useful image features without human-labelled examples, using convolutional networks, the standard architecture for image tasks.

Radford joined OpenAI in 2016, a year after it was founded, and worked on natural-language processing.^[1] His 2018 paper, “Improving Language Understanding by Generative Pre-Training,” on which he was lead author, introduced the GPT approach. In that approach, a model is first trained on large amounts of text to predict the next word, and then adapted to specific tasks.^[1]

In 2019 he was first author of “Language Models are Unsupervised Multitask Learners,” the GPT-2 paper.^[4] The 2020 GPT-3 paper credits him with originally demonstrating that few-shot learning occurs in language models, and with collecting and filtering the training data. Few-shot learning means a model can perform a new task from a handful of examples placed in its input, without retraining.^[7] He is also a co-author of “Scaling Laws for Neural Language Models” (2020).^[1]

In 2021 he was first author of the CLIP paper, which learned images and text together, and a co-author of the DALL-E paper.^[5] CLIP learns to match pictures with written descriptions, which lets it classify an image by comparing it with text labels.^[5]

In December 2022 he was first author of the Whisper paper, “Robust Speech Recognition via Large-Scale Weak Supervision,” which scaled training to 680,000 hours of multilingual and multitask audio.^[6]

Positions

Radford was at OpenAI from 2016 until December 2024, when he left to pursue independent research. In 2025 he became an adviser to Mira Murati’s Thinking Machines Lab.^[2]

Impact

The GPT approach he helped introduce underlies the chat-style AI systems that many people use today. CLIP and Whisper became common reference points for later work on linking images with language and on speech recognition.^[5][6] Read together, his papers show a recurring pattern: train a model on very large amounts of data with a simple objective, then reuse the result for many tasks. That pattern links work on text, images and speech that might otherwise look unrelated.