On the 15th, the news of a researcher joining OpenAI attracted significant attention.
The researcher named James Campbell recently obtained a PhD in Computer Science from CMU in 2024. Now, he suddenly announced that he would abandon his doctoral studies and join OpenAI.
On social media X, he stated that his research focus at OpenAI is "AGI and ChatGPT's memory + personality", and that memory will fundamentally change the relationship between humans and machine intelligence. He will work hard to ensure its correct implementation.
His joining was even welcomed by OpenAI co-founder and president Greg Brockman.
Some also asked, "Why did it take you so long to decide to join OpenAI?"
So, who is this guy, and why did his joining attract so much attention? Let's look at his background.
He graduated from Cornell University with a major in Mathematics and Computer Science. During his undergraduate studies, he was dedicated to research on LLM interpretability and authenticity, and was the primary author of two papers: "Representation Engineering" and "Localizing Lying in Llama".
The former paper studied representation engineering: a top-down approach to AI transparency, while the latter researched locating lies in Llama: understanding dishonest instructions on true-false questions through prompting, probing, and patching.
- Paper 1: Representation Engineering: A Top-Down Approach to AI Transparency, https://arxiv.org/abs/2310.01405
- Paper 2: Localizing Lying in Llama: Understanding Instructed Dishonesty on True-False Questions Through Prompting, Probing, and Patching, https://arxiv.org/abs/2311.15131
He also worked on adversarial robustness and assessment at Gray Swan AI (an AI safety company), and has previously worked in computational neuroscience, physics, and deep learning theory. He is the founder of ProctorAI (a multimodal AI system monitoring user work status), AI-Timeline.org, and a co-author of AidanBench (AI model benchmarking work).
ProctorAI once sparked widespread discussion. The system monitors the user's screen and alerts you if it detects procrastination. Its working principle is to capture computer screen screenshots every few seconds and input them into multimodal large models such as Claude 3.5 Sonnet, GPT-4o, and LLaVA 1.5. If it finds the user is not focused, it controls the screen and reminds with a personalized message.
Project address: https://github.com/jam3scampbell/ProctorAI
In 2024, he entered CMU to pursue a PhD, focusing on understanding the components of general intelligence and ensuring their safety.
Now, he has chosen to join OpenAI and begin his exploration in the industry.
In fact, James Campbell's joining was not without precedent. A few days ago, they began interacting under OpenAI's post about forming a model behavior research team.
Earlier, James Campbell had published some views on ChatGPT's memory.
He agreed that research should be conducted on ChatGPT's memory and the issue of stronger hallucinations it causes, and believed that context pollution and model collapse are real but currently under-researched. If thousands of tokens of memory are stuffed into each prompt, they will become more active.
It seems that James Campbell's joining is a mutual pursuit.
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: Du Wei, published by 36kr with authorization.
