According to The Information, Alec Radford, a researcher at OpenAI and the father of GPT, announced in his circle of colleagues yesterday that he will leave his job to conduct research independently. In the future, he plans to work in cooperation with OpenAI and other artificial intelligence developers.

At the recently concluded NeurIPS conference, OpenAI co-founder Ilya Sutskever reviewed the important historical moments he called the "pre-training era" and also mentioned: the birth of landmark models such as GPT-2 and GPT-3. , thanks in large part to the pioneering work of Alec Radford and Anthropic founder Dario Amodei.

I wonder if IIya already knew about Radford’s departure at that time. . . .

OpenAI director Mark Chen confirmed the news in a statement:

“We have deep respect and appreciation for Alec and his contributions and look forward to continuing to work with him as he explores independent research.”

Is OpenAI really okay without him...

This person's resignation is really serious for OpenAI! Check out these sensational papers at that time:

GPT-1:

GPT-2：

CLIP：

Google Scholar personal page:

Sam Altman's evaluation of Alec Radford:

The guy who created GPT-1 was somewhat independent in solving the problem, which was somewhat impressive, but (at the time) there wasn't a deep understanding of how it worked and why it could work

This is truly the father of GPT!

Alec Radford graduated from Franklin W. Olin College of Engineering in 2016 and joined OpenAI after graduation to begin his seven-year research journey.

During his 7 years at OpenAI, Radford has always focused on the research of large language models. It can be said that his work not only laid the development direction of OpenAI, but also led new changes in large language models and even the entire AI industry.

In June 2018, he led the publication of an epoch-making paper, proposing for the first time a language model generation pre-training method based on Transformer. This preprint paper, published on OpenAI's official website, can be considered a milestone achievement in the field of AI, laying the theoretical foundation for the explosive development of AI in the future.

Radford named their model Generatively Pretrained Transformer, or GPT for short.

This model with 117 million parameters integrates massive data such as books, Q&A, and middle and high school exam essays, demonstrating unprecedented abilities in language comprehension and generation.

Radford is not only the creator of GPT-1, but also led the development of GPT-2 as the first author. In addition, he has also made outstanding contributions in projects such as the CLIP model and the speech recognition system Whisper in the field of computer vision.

Pioneers of CloseAI are really gone

OpenAI keeps losing its earliest and most groundbreaking contributors

Some of OpenAI’s veterans: CTO Mira Murati, research director Bob McGrew, co-founder and researcher John Schulman, genius Ilya Sutskever, and security chief Lilian Weng have also left in recent months for various reasons.

And now, Radford.

This trend appears to be no accident.

The departure of these pioneers of the "pre-training era" is not only a matter of personal choice, but also represents the new trend of this era, suggesting that the industry may be undergoing a paradigm shift.

Previously, Dario Amodei, who co-created the era of pre-trained models with Radford, has established a separate company, founded Anthropic and made remarkable achievements in the field of AI security.

An anonymous Wall Street analyst’s evaluation of the departure of “tech soul figures” like Radford is very far-sighted:

This may indicate a major shift in the paradigm of AI research. We may be witnessing the beginning of a new era in which independent researchers will play an increasingly important role in AI technology innovation.