Alec Radford

Alec Radford is an AI researcher whose collaborative work includes DCGAN, GPT, GPT-2 and Whisper. In 2026, he coauthored research on shaping language-model capabilities through training-data filtering and helped develop talkie, a language model trained on historical text to study generalization beyond its intended knowledge cutoff.[5][6][8][11][12][14]

With Neil Rathi, Radford authored Shaping capabilities with token-level data filtering, first submitted on January 29, 2026. They studied reducing undesired capabilities during pretraining rather than intervening after a model has learned them. On a proxy task of removing medical capabilities, they reported that filtering individual tokens caused less damage to benign capabilities than filtering whole documents for the same reduction in the targeted capabilities. Their method included labeling tokens with sparse autoencoders and distilling classifiers to perform the filtering.[12]

Radford developed talkie with Nick Levine and David Duvenaud. Their April 2026 report introduced a 13-billion-parameter base model trained on a reported 260 billion tokens of historical English text, with an intended cutoff at the end of 1930. The team chose that cutoff because of US public-domain availability. They proposed using vintage models to investigate forecasting, possible rediscovery of later inventions, and learning beyond the material available during pretraining.[13][14]

A concrete experiment asked vintage models to write Python programs after seeing in-context examples. The team reported a rotation-cipher decoding function produced from an encoding example by changing addition to subtraction—a single-character edit. They also stressed the narrowness of the result: all correct vintage-model solutions were simple one-line programs or small modifications of examples. This was an early generalization experiment, not a demonstration of broad programming competence.[13]

The team supplied a FineWeb-trained “modern twin” with the same architecture and training FLOPs for comparison, while cautioning that the corpora differed in subject matter as well as temporal coverage. Talkie’s historical boundary was also imperfect: a document-level n-gram anachronism classifier did not prevent the 13B model from knowing some World War II and postwar facts. For instruction tuning, the team built examples from historical reference works, then used modern Claude models during post-training; they acknowledged that AI feedback introduced anachronistic behavioral influence.[13][14]

These 2026 projects followed Radford’s earlier collaborative research on learned image representations, language models and speech recognition, including work published with an OpenAI affiliation.[5][6][8][11][12][13]

In November 2015, Radford, Luke Metz and Soumith Chintala submitted Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. They introduced DCGANs, convolutional generative adversarial networks with architectural constraints, and reported hierarchical representations ranging from object parts to scenes in both the generator and discriminator. They also reused learned features for other tasks, testing their usefulness as general image representations.[5]

Radford’s GPT collaborations concerned language understanding through generative pretraining and language models as multitask learners. He coauthored Improving Language Understanding by Generative Pre-Training with Karthik Narasimhan, Tim Salimans and Ilya Sutskever, and wrote its June 11, 2018 announcement. For GPT-2, Language Models are Unsupervised Multitask Learners, he and Jeffrey Wu were credited as equal contributors.[6][7][8]ⁿ¹

In a March 2024 Latent Space interview, David Luan recalled sprinting to help Radford finish the GPT-2 paper and writing its unusually short modeling section together. Luan remembered their concern that established researchers would say, “They're going to say no novelty.” He interpreted the episode as prioritizing solving AI problems over architectural novelty—a retrospective account of their collaboration.[3]

Radford’s later collaborations extended to connecting text and images and to speech recognition. He coauthored the January 5, 2021 CLIP announcement, Connecting text and images. In Whisper, Robust Speech Recognition via Large-Scale Weak Supervision, Radford and Jong Wook Kim were credited as equal contributors, placing him among the researchers responsible for the speech-recognition work.[10][11]ⁿ²