Alec Radford is a machine learning researcher at OpenAI widely regarded as one of the most technically influential yet publicly low-profile figures in modern generative AI. He is the first author of the papers that introduced OpenAI's original generative pretraining approach and its 2019 successor, and he later led or co-led two of the lab's most widely reused open models.
GPT and GPT-2
Radford was first author of "Improving Language Understanding by Generative Pre-Training" (2018), the paper that introduced what became known as GPT, establishing the recipe of unsupervised Pretraining on a large text corpus followed by task-specific fine-tuning. He was also first author of the 2019 follow-up paper behind GPT-2, which demonstrated that a sufficiently large Autoregressive model trained only to predict the next token could perform many language tasks with no task-specific training data at all, a result central to the "unsupervised multitask learner" framing that shaped later work on large language models such as GPT-3.
CLIP
In 2021, Radford was a lead author of CLIP (Contrastive Language-Image Pre-training), a model trained to associate images and text descriptions using a contrastive objective over large web-scraped datasets. CLIP became a foundational component of the text-to-image boom that followed, since its image-text embeddings were used to guide generation in systems including early versions of DALL-E and were adopted widely across the Text-to-image generation research community.
Whisper
Radford was also first author of Whisper (2022), an open automatic Speech recognition model trained on a large amount of weakly supervised multilingual audio. Whisper was released with its weights public and became one of the most widely deployed open transcription systems, used in products far outside OpenAI itself.
Profile and influence
Unlike some OpenAI colleagues such as Ilya Sutskever or Sam Altman, Radford has kept a comparatively low public and media profile despite being first author on several of the field's most cited papers. Colleagues and outside researchers have repeatedly pointed to his GPT-2 and CLIP work as pivotal in showing that scale and self-supervised objectives, rather than increasingly specialized architectures, were the more productive direction for the field, a thesis later formalized in work on Scaling laws by colleagues including Jared Kaplan.