After securing a Nobel laureate, Anthropic poached the head of UC Berkeley's CS department, recruiting four top talents in just two weeks

On the afternoon of July 1st, a tweet caused a stir in the Silicon Valley academic circle.

Jelani Nelson, the head of the Department of Computer Science in the Electrical Engineering and Computer Sciences (EECS) at UC Berkeley and a professor of theoretical computer science, temporarily put down his office keys and went to Anthropic.

He posted on X:

I have joined Anthropic and taken a leave from the university. I'm excited to work with many talented and mission - driven people to research the defining technology of our era.

These two short sentences are packed with information: he has started work at the new place while retaining his faculty position through a leave of absence. As for his position, team, and research direction, nothing was mentioned.

Nelson's profile on X has been updated accordingly: Member of Technical Staff at Anthropic, becoming a colleague of Karpathy, who joined in May.

 Jelani Nelson, the Chair of the Department of Computer Science in the Electrical Engineering and Computer Sciences (EECS) at the University of California, Berkeley

The person in charge of one of the top computer science departments in the United States just left without hesitation.

AI companies have been poaching talents for three years, from engineering to product, from alignment to multimodality.

This time, they have reached into the pinnacle of theoretical computer science.

The Man Who Rose from MIT to Berkeley and Achieved the World's Best in "Counting"

Nelson's resume is almost a standard full - set for theoretical computer science.

He self - taught HTML to build websites in junior high school, learned programming in high school, and proved himself capable of writing bug - free code the fastest through competitions during college.

He completed his undergraduate, master's, and doctoral degrees at MIT, obtaining his Ph.D. in computer science in 2011, with a focus on efficient algorithms for massive data.

He described the appeal of this discipline to him as "almost religious": it is both the most fundamental core problem of human thinking and closely related to the real world.

After graduating with his doctorate, he did post - doctoral research at Berkeley, Princeton University, and the Institute for Advanced Study (IAS) in Princeton successively, and joined Harvard as a faculty member in 2013.

In 2019, Nelson left Harvard and moved west to UC Berkeley.

The Harvard University Gazette directly expressed its regret in the title: His departure left a big hole in the computer science department.

At Berkeley, he thrived and immersed himself in the theoretical circle centered around the Simons Institute for the Theory of Computing.

In the fall of 2024, Nelson took over as the Chair of the Department of Computer Science in EECS, leading one of the world's top CS departments.

His main research areas are streaming algorithms, dimensionality reduction, and randomized algorithms.

In plain language, Nelson is studying the same kind of thing: how to perform calculations when the data is too large to fit.

A few years ago, he set his sights on a problem that seems like a primary - school math problem: teaching computers to count.

This may seem simple, but when the numbers are so large that mobile phones and servers can't remember "where they left off", the costs of storage and speed will get out of control.

His team came up with a mathematical formula to prove the minimum amount of memory required for any algorithm to solve this problem.

Nelson's team's paper, which proves the memory lower bound for the approximate counting problem. https://arxiv.org/pdf/2010.02116

Engineers make programs run faster, while Nelson proves how fast a program can run at most. This is the job of a theoretical computer scientist: setting the physical lower limit for computation.

Nelson's contributions to the academic circle are far more than just "counting".

First, he and Kasper Green Larsen proved the optimality of the Johnson - Lindenstrauss lemma. This is a cornerstone in the field of dimensionality reduction, and he nailed down the theoretical lower bound. Previously, he and Daniel Kane proposed the sparse JL transform.

Second, he, along with Kane and David Woodruff, provided an asymptotically optimal algorithm for the count - distinct problem (how many different elements there are in a data stream).

In his view, even something as simple as "counting" that everyone can do has a theoretical optimal solution behind it.

These achievements have brought him a long list of honors: the Sloan Research Fellowship, the Presidential Early Career Award for Scientists and Engineers (PECASE), and more.

Beyond academia, Nelson has another side.

In 2011, while still a Ph.D. student at MIT, he went to Ethiopia and founded the free programming summer camp AddisCoder.

Over the past fourteen years, nearly 700 students have graduated from here, and some have gone on to pursue Ph.D.s at Harvard, MIT, and Stanford.

Later, Chronixx, a famous Jamaican reggae star, voluntarily donated money, which led to the sister project JamCoders.

The free programming summer camp AddisCoder founded by Nelson in 2011 has trained nearly 700 students. (Source: AddisCoder official website)

Nelson is also one of the most vocal opponents of the math curriculum reform in California. The reason is simple: his grandfather came from a poor family and became a doctor through high - quality public education, changing the trajectory of the entire family.

Therefore, in his view, removing rigorous math courses from public schools is like taking away the ladder for the next generation to climb up.

This "off - academic" cause later earned him the ACM Lawler Award for Humanitarian Contributions.

What Does Anthropic Need a Theoretical Scholar For?

What's the connection between a professor researching streaming algorithms and a large - model company?

Nelson's research areas (streaming algorithms, dimensionality reduction, randomized algorithms) are actually about the same thing: how to process the largest amount of data with the least amount of memory and computation.

Corresponding to large models, these are exactly the most costly aspects: training efficiency, data compression, and computational complexity.

Take the JL lemma, whose final piece was completed by him, as an example. It answers a question so simple that it's almost common sense: how small can high - dimensional data be compressed without losing fidelity.

The underlying intuition of vector retrieval and embedding compression used all over the world today is based on this lemma.

Training a cutting - edge model essentially involves compressing and screening astronomical data streams. On the inference side, video memory, cache, and context windows are all about struggling with memory and complexity.

And this is exactly the problem domain that Nelson has been delving into for twenty years.

When the model scale hits the ceiling of computing power and data, the value of "saving" begins to exceed that of "piling up". The focus of AI competition is shifting from "whose model is stronger" to "whose underlying algorithms are more cost - effective".

The toolbox of streaming and randomized algorithms naturally fits the problem of "approximating the optimal solution with limited resources", which exactly hits the common anxiety of all cutting - edge laboratories today.

From this perspective, Anthropic signing a theoretical computer scientist is more like making up for a deficiency: deepening the theoretical foundation beyond models, engineering, and alignment.

It's Now Popular for Top Professors to Join AI Companies Without Resigning

Regarding his joining Anthropic, Nelson's exact words were "taken leave from the university", which means a leave of absence from the university.

A leave of absence is different from resignation: the faculty position is retained, and he can return at any time.

This is a well - established system in the US academic circle. Professors can take paid or unpaid leave to go to the industry, start a business, or do anything else.

This path has been proven by others.

In 2017, Fei - Fei Li took an academic leave to serve as the vice - president of Google and the chief scientist of Google Cloud AI, and returned to Stanford two years later.

Nowadays, the revolving door between academia and industry is spinning faster and faster, and "joining a company on leave" is becoming a mainstream model.

For scholars, this is a guaranteed ticket, and besides, the industry has computing power, data, and real - world problems that academia doesn't have.

For AI companies, this is a low - friction talent recruitment channel. More importantly, signing a scholar means not only getting an individual but also his students, peers, and the entire academic network behind him.

The traditional one - way path of "getting tenure and working until retirement" is being replaced by the leave - taking model of "dipping a toe into the industry".

For universities, once this door is opened, it's hard to close.

After Poaching from Peers, AI Giants Start Poaching from Universities

How crazy was the AI talent market in June just passed?

On June 18th, Noam Shazeer, the author of the Transformer paper and the co - leader of Gemini, announced that he was leaving Google to join OpenAI.

It should be noted that Google bought him back from Character.AI in 2024 with a $2.7 - billion deal, but less than two years later, he left again.

On June 19th, John Jumper, who won the 2024 Nobel Prize in Chemistry for AlphaFold, officially announced that he was leaving DeepMind, where he had worked for nearly nine years, to join Anthropic.

Constrained by the non - compete clause of DeepMind's senior management, he may not be able to start work until next year.

On June 24th, Bloomberg reported that Jonas Adler and Alexander Pritzel, core researchers of Gemini, would also join Anthropic. Both of them are collaborators of Jumper on protein structure research.

Alphabet's stock price dropped immediately, and investors began to publicly question whether Google could retain its talents.

So far, the battle was still between AI companies. Soon, the battle spread to universities.

On June 25th, Dawn Song, an AI safety scholar who had taught at Berkeley for 19 years, announced that she was joining Meta's Super Intelligence Laboratory as the vice - president of AI research.

On July 1st, it was Nelson's turn.

In just two weeks: a Nobel laureate, two core Gemini researchers, a senior professor, and a current department chair.

Among them, Jumper, Adler, Pritzel, and Nelson all moved to Anthropic.

The background of this crazy talent flow is not hard to guess.

OpenAI has secretly submitted its IPO documents, and Anthropic is also pointed to by multiple sources as being close to going public. For top researchers, joining now means getting equity before the company goes public, which is a price that big companies can't offer.

Berkeley's role in this migration is particularly eye - catching.

The Simons Institute for the Theory of Computing is located here, as well as one of the top EECS departments in the United States. The three lines of theory, machine learning systems, and AI safety are continuously supplying talents to Anthropic, OpenAI, and DeepMind.

In the previous round, AI giants poached people who could train models. In this round, they are poaching people who know the limits of models.

As top scholars flood in, AI companies are actually growing into a "second research institution system".

If the best theorists are all on "leave" at companies, what will be left of universities? No one knows.

The only thing certain is that the focus of the AI competition has shifted from model capabilities to the theoretical foundation of algorithms.

References:

https://x.com/minilek/status/2072322757908664728?s=20 

https://www2.eecs.berkeley.edu/Faculty/Homepages/minilek.html 

https://vcresearch.berkeley.edu/news/jelani-nelson-considers-human-thought-computer-science-tools 

https://arxiv.org/pdf/2010.02116 

This article is from the WeChat official account "New Intelligence Yuan", author: ASI Revelation. Republished by 36Kr with permission.