Just now, Harvard Chinese researcher Jeffrey Wang officially announced to join OpenAI

Will work as a foundation team researcher, responsible for model pre-training and inference.

After the announcement of the news, Greg Brockman, president of OpenAI Joint Ventures, also led the team to welcome the first time.

Before that, Jeffrey Wang worked on machine learning while teaching at Harvard.

Joining OpenAI from Harvard, High school finalist for US "Junior Nobel Prize"

After graduating from a high school in California, Jeffrey Wang enrolled at Harvard in 2021 as an undergraduate majoring in math and computer science, with a minor in English.

He has been working for OpenAI since September, and it may be the experience of these few months that has finally made him decide to join OpenAI now.

In fact, while still a student at Harvard, Jeffrey Wang began teaching and researching machine learning.

Since January last year, he has lectured to 61 students in the Computer science and statistics department.

At the same time, he also worked as a researcher in the School of Engineering and Applied Sciences, focusing on the robustness and privacy of machine learning models.

Two co-authored papers were also presented at the NeurIPS SoLar (2023) and ICML (2024) conferences.

MoPe: Model Perturbation-based Privacy Attacks on Language Models
MoPe: Language model privacy attack based on model perturbation

This paper proposes a new detection tool called MoPe (Model Perturbations), which can help people determine whether a particular text is being used to train a large language model.

This tool works by adding some disturbance (noise) to the model parameters and then observing the effect of this disturbance on the model predictions. The study found that MoPe is more efficient than previous methods (in the 70M to 12B parameter range) and can more accurately simulate the impact of changes in model parameters on predictions.

In addition, the study found that just looking at the loss value of one training point is not enough, because some points with average loss values can be successfully recovered by MoPe, challenging previous studies that rely on the loss value to determine whether the information is remembered by the model.

Part Two: Bias Begets Bias: the Impact of Biased Embeddings on Diffusion Models
Bias breeds bias: The impact of bias embedding on diffusion models

This paper examines how biases in embedded Spaces affect the fairness of diffusion models.

The authors propose that unbiased text embeddings are necessary to generate a balanced distribution of images, and find that biased multimodal embeddings (such as CLIP) lead to lower scores when assessing the alignment of images with text prompts, inadvertently rewarding unfair behavior.

Of course, the paper also presents a theoretical framework to study bias in assessment and suggests ways to mitigate bias.

In addition, Jeffrey Wang won a national award in high school for "developing statistical methods to detect structural changes in 3D genomes."

In 2021, he was a finalist for the (top 40) Regeneron Science Talent Search (STS), the oldest and most prestigious high school research competition in the United States, known as the "Junior Nobel Prize" in the United States.

The top 40 finishers will each receive a minimum prize of $25,000.

The research is the result of a two-year collaboration with Dr Abhijit Chakraborty, who spends his weekends and breaks visiting each other's laboratories.

They studied the structure of DNA in chromosomes. Because many developmental and disease processes, including cancer, are caused by changes in the three-dimensional arrangement of DNA, Jeffrey created a computer application that identifies and arranges significant differences in this arrangement across the genomes of different cell lines.

The program makes it easy to locate important genes and pathways based solely on the three-dimensional shape of DNA.

He has since made the model public (it is now used by more than 50 laboratories) so that more people can use it to discover new pathways, regulatory mechanisms, and cancer targets.

And a writer.

BTW, in addition to his obvious love of development, Jeffrey Wang is an accomplished writer.

Back in 2019, he signed on to The Adroit Journal, a New York-based publication founded in 2010 that showcases poetry, prose, and art from emerging writers around the world, with an inclusion rate of less than 1%.

After receiving a literary and artistic education, he was signed by PBH Network long-form history writer, and his multiple syndicated works have been viewed more than 250,000 times.

To this day, he shares stories about history, science and statistics on Quora, a popular online question-and-answer platform in the United States, and his work has been viewed more than 6 million times.

Well, it's doge again.

However, some netizens took the opportunity to ridicule and "quit school" to join the OpenAI group +1.

You know similar news in OpenAI is not uncommon, a while ago also recruited a high school dropout from Sweden, joined the Sora team to specialize in AGI.

Even OpenAI CEO Altman dropped out of his Stanford computer science major at age 19 to start his first company.

……

It's almost a corporate culture. bushi

Paper 1:https://aclanthology.org/2023.emnlp-main.842.pdf
Paper 2:https://arxiv.org/pdf/2409.09569