OpenAI's senior executive resigns to start a new business.

In the early hours of today, William Fedus, the head of the post-training team at OpenAI, announced his resignation on the X platform and plans to start a new startup focused on using AI to discover new materials.

According to the media outlet The Information, Fedus' former employer OpenAI is expected to invest in and collaborate with his new company.

Fedus joined OpenAI from Google in 2022. As the head of post-training, he was primarily responsible for overseeing the development of models for ChatGPT, OpenAI's developer tools, and other products.

During this period, he led the post-training research and development of several important models at OpenAI, including but not limited to 4o, o1-mini, and o1-preview.

After the departure of OpenAI's previous head of post-training, Barret Zoph, Fedus was promoted to the position of head of the post-training team. The paper authorship shows that Fedus was one of the main responsible persons for the o1 and o3 models.

In terms of educational background, William Fedus obtained a bachelor's degree in physics from the Massachusetts Institute of Technology from 2010 to 2013.

From 2013 to 2017, he earned a master's degree in physics from the University of California, San Diego. Subsequently, he pursued a doctoral degree in computer science at the University of Montreal, with Yoshua Bengio and Hugo Larochelle as his advisors.

It is worth mentioning that Yoshua Bengio is one of the pioneers in the field of deep learning and, together with Geoffrey Hinton and Yann LeCun, was awarded the Turing Prize in 2018.

William Fedus has had multiple work experiences, such as interning at Google Brain and researching generative adversarial networks (GANs).

According to public information, William Fedus' research interests include supervised learning, unsupervised learning, and reinforcement learning, with a focus on building robust and reliable algorithms.

In 2018, he co-proposed MaskGAN, a generative adversarial network for text generation, which generates higher-quality text by filling in blanks.

In 2019, he participated in proposing an unsupervised graph neural network method that learns node embeddings by maximizing the mutual information between node representations and graph summaries.

In 2022, he, along with Barret Zoph and OpenAI researcher Noam Shazeer, proposed Switch Transformers, an efficient architecture for building trillion-parameter models through sparse activation, addressing issues such as model complexity, communication cost, and training stability.

After the resignation announcement, friends such as Logan Kilpatrick, the product manager of Google AI Studio, and Noam Brown, an OpenAI researcher, left comments in the comment section to offer their blessings.

GPT-5 is on the way

While there is a "seismic" change in OpenAI's management, the company's technical research and development is still ongoing.

Recently, OpenAI's Chief Product Officer, Kevin Weil, gave an interview and discussed various hot topics, including GPT-5. He stated that OpenAI is integrating the GPT and o-series models to launch a more powerful GPT-5, but the specific release timeline is not yet clear.

GPT-5 will simplify the model series and unify the functionality, encompassing all the tools and capabilities of the previous model series, such as processing file uploads, web searching, and even using Python programming during the reasoning process.

Regarding Deep Research, Kevin stated that it is the best product they have launched since ChatGPT, capable of completing tasks that users cannot do on their own. He pointed out that this success is the result of close collaboration between the research, product, engineering, and design teams.

Regarding the future of programming, Kevin believes that at the current rate of development, AI will surpass humans in the field of programming earlier than Anthropic's prediction of 2027, possibly as early as 2026.

He explained that when GPT-4 was released, it was already an excellent programming model, widely used in tools like GitHub Copilot. However, compared to o1 preview, o1 performs better in programming because reasoning ability is crucial when writing code.

Reasoning ability, to a certain extent, is necessary for solving problems, completing crossword puzzles, solving Sudoku, and writing code, as it allows breaking down problems into smaller parts, forming hypotheses, and verifying or refuting them.

In his view, AI will surpass humans in competitive programming standards within this year, just as computers surpassed humans in multiplication calculations 70 years ago, which will democratize software development and allow more people to create software.

Additionally, Kevin pointed out that while content generation is becoming easier, people will still value works that have been created with a great deal of time and effort. AI will change the nature of work, but it will not replace human work; instead, it will complement and enhance human capabilities.

Regarding future new job categories, Kevin believes that almost all jobs will be supplemented by AI. He also proposed a specific startup idea - providing personalized AI tutors for every child to improve the quality of education and living standards.

When asked about the possibility of developing physical products, Kevin stated that OpenAI has started to focus on robotics technology. Although the specific form is not yet determined, he believes that robotics technology is an important direction for the future. After achieving artificial general intelligence in the digital world, the next step is to extend the technology to the physical world and impact the physical environment.

The original interview video can be found at: https://www.youtube.com/watch?v=SnSoMh9m5hc&t=88s&ab_channel=Overpowered

This article is from the WeChat public account "APPSO", written by APPSO, and published on 36Kr with authorization.