San Francisco, California, United States
Combining scaling and long-term AI alignment methods to build steerable, interpretable, and robust AI systems.
* I led the efforts to build GPT-2 and GPT-3 (https://arxiv.org/abs/2005.14165). * I am one of two people who sets overall research direction at OpenAI and writes its annual research roadmap. * In addition I have built and lead several teams focused on long-term safety research, including how to make AI systems more interpretable and how to embed human preferences and values in future powerful AI systems.
I worked as a deep learning researcher on the Google Brain team, working to extend the capabilities of neural networks. I have also worked on the safety and reliability of AI systems, and recently published a paper (with Chris Olah and several other researchers) laying out some key problems for preventing accidents in AI systems. Our work (https://arxiv.org/abs/1606.06565) is described here (https://www.technologyreview.com/s/601750/google-gets-practical-about-the-dangers-of-ai/).
I work with Andrew Ng and a small team of AI scientists and systems engineers to solve hard problems in deep learning and AI, including speech recognition and natural language processing. * Performed the majority of machine learning research for Deep Speech 2 (http://arxiv.org/abs/1512.02595), listed by MIT Technology Review as one of the top 10 technological breakthroughs of 2016 (https://www.technologyreview.com/s/600766/10-breakthrough-technologies-2016-conversational-interfaces/) * Conceived, prototyped and implemented a neural network architecture that achieved a 35% relative reduction in the word error rate (WER) of the lab's English speech system. * Conceived, prototyped and implemented a neural network architecture that achieved a 15% relative reduction in the WER of the lab's Chinese speech system. * Increased speed of decoder by 10x in Chinese and English, removing a key obstacle to deployment of the lab's speech system to Baidu users. * Increased speed of training for Chinese system by 27% and reduced memory usage by 50%, resulting in a substantial increase in lab-wide productivity and computational efficiency. * Co-developer (with Jim Fan and Jesse Engel) of Baidu's internal neural network library * Awarded "Star Employee of the Quarter" for Q2 2015.
Skyline is a comprehensive software suite essential to the research of a large fraction of the proteomics community: 30,000 unique installs, 7000 instances/week, 300K lines of C# code, 10 professional programmers • Personally contributed 14K lines of code (net, human-generated) + 50 pages tutorials and docs. • Invented and implemented peak picking, binary classification, Storey-Tipshirani, and signal processing algorithms for mass spectrometry data, now widely employed by Skyline’s user base. • Instructor at weeklong course (Zurich, February 2014) attended by 40 on how to use Skyline, invited to teach second course in Barcelona (October 2014). • Featured speaker at Skyline user group meeting (June 2014, attendees: 250), and IMSC Work- shop (August 2014, attendees: 100+) • Regularly handle support requests from Skyline users (3-4 requests/week)