Yang Jing Mingmin was sent from The Temple of Oufei
Qubits | Official account QbitAI
After questioning the results of his paper published in Nature, the Google researcher was fired.
Last June, Google published an article in Nature: A graph placement methodology for fast chip design, of which Jeff Dean was one of the authors.
The article said that AI can generate chip design drawings in 6 hours, and it is better than humans.
But the researcher argues that some of the article's assertions are untenable, and that the experiments have not been adequately tested.
However, he only expressed his ideas internally, and the paper that verified his views was directly intercepted by Google and failed to publish it.
(Netizens can't see the preprint)
Google's latest response confirmed that the researcher had been fired "for some reason" in March.
It is worth mentioning that this is not the first time that Google employees have been "convicted of words".
In just a year and a half, Google has reported that three technicians have been fired.
AI chip design capabilities beyond humans?
The paper involved was accepted by Nature on April 13 last year and published on June 9.
It mainly discusses a way to use deep reinforcement learning to quickly design chips.
According to the paper, ai can design a chip in less than 6 hours, while manual work often takes weeks or months to do so, the paper says.
Specifically, this is a chip layout method with generalization capabilities.
By learning 100,000 chip layouts, ai can design new solutions on their own, and all key indicators (including power consumption, performance, and chip area) are comparable to manual design.
In order to improve the learning efficiency of AI, the researchers also designed a reward mechanism that calculates based on the approximate cost function of line length and wiring congestion.
Specifically, macros and standard units need to be mapped onto a flat canvas to form a "chip netlist" with millions to billions of nodes.
The AI model then optimizes the power, performance, and area (PPA) and outputs a probability distribution.
The following illustrations are the effects of zero sample generation and fine-tuning based on a pre-training strategy, where each small rectangle represents a macroblock. In the pre-training strategy, there is room in the middle for standard units.
Anna D Goldie, co-corresponding author of the paper, said,
This approach works for any type of chip design and is already being used to produce the next generation of Google TPUs.
The skepticism began in 2020
But behind the "major breakthrough", the internal doubts about this technology within Google actually began in 2020.
It was Dr. Chatterjee, a Google Brain employee who was fired in March, who raised the issue.
He graduated from the Department of Computer Science at UC Berkeley and worked at Intel, where he focused on high-level modeling and verification of communication protocols.
In 2020, Google proposed a way to design chips using machine learning, which can be seen as a precursor to the Achievements in Nature.
At the time, Google asked Dr. Chatterjee if the method could be sold or licensed to chip design companies.
In response to the email, Dr. Chatterjee said he had reservations about some of the claims in the paper and questioned whether the technology had been rigorously tested.
However, these doubts do not seem to have affected Google's footsteps.
A year later, they submitted the questionable work to Nature and published it.
Unlike the original study, the paper on Nature made some tweaks to the previous approach, while also removing the names of the two authors.
Because they had worked closely with Dr. Chatterjee and were equally skeptical about the results.
And Google's riot operation is not over, they also personally demonstrated, what is called "rules are dead people are alive."
On this side, papers submitted to Nature were questioned and did not strictly follow the publication approval process.
In response, Anna D Goldie, Google and co-corresponding author of the paper, said that because the paper did not make much changes to the previous work, it did not need to go through a full approval process.
But on the other hand, Dr. Chatterjee's paper questioning this achievement ultimately failed to pass the review.
They submitted the paper refuting this approach to a resolution committee for approval for publication.
As a result, a few months later, the paper was rejected.
The reason is: not up to standard.
According to The New York Times, Dr. Chatterjee's team was told by Google that they would not publish a paper questioning Nature's work.
And a written report proves that Dr. Chatterjee has been fired by Google.
Google Vice President Zoubin Ghahramani responded to the matter by saying:
We thoroughly investigated the manuscript of the Nature paper and insisted on the results of peer review.
At the same time, we also rigorously investigated a subsequent submission that did not meet our publishing standards.
At the same time, people familiar with the matter revealed that one of the authors of the results said that "the dismissed people harassed her and questioned her work."
The lawyer for the "dismissed man" responded: He was defending the integrity of science.
Dismissal if you disagree?
As mentioned earlier, this is not the first time that Google management has clashed with researchers.
The most typical one before was the dismissal of Timnit Gebru, co-head of Google's AI ethics team.
At that time, this incident directly triggered 1400 Google employees and 1900 AI academic circles to condemn Google's behavior and make Jeff Dean the target of public criticism.
According to my tweet and the response of my brother-in-law, the core contradiction is the difference between the two sides in the internal paper review.
According to the contents of the public email, she had always wanted to publish a paper on "the big language model is biased", but she has been opposed by her superiors.
In addition, she revealed in the email that Google treats black people as a vulnerable group unfairly, and does not attach importance to AI ethics.
It was because of this email that it was revealed that it "did not meet the expectations of Google's managers" and was fired.
Two months later, Margaret Mitchell, another head of the AI ethics team, was fired for using scripts to search the company's intranet for evidence in support of Gebru.
In April last year, Samy Bengio, one of the founding members of Google Brain, also left to join Apple, and many people speculated that the reason for the departure was related to this matter.
Now, another Google Brain member has been fired, which is considered the latest example of internal turmoil among Google researchers.
Previously, DeepMind, another major research team of Google, also broke rumors from time to time about making independence and breaking up with the parent company.
On the other hand, as the netizen said, the entire technology industry is dealing with the problem: "adaptation" between researchers and companies.
Even a big manufacturer like Google can't completely solve this problem.
From the perspective of enterprises, of course, it is the rapid commercialization of scientific research results to achieve revenue.
But from the perspective of a technician, any scientific research is a long-term investment, how can it be allowed to be degraded for short-term benefits.
Once they encounter a mismatch, there are only two paths left for them:
One is to remain in the industry, change companies or start your own business;
For example, Baidu chief scientist Wu Enda, who is responsible for the leadership of Baidu Research Institute, especially the Baidu Brain program, left after three years and had his own entrepreneurial project.
The other is a return to academia.
Li Feifei, a professor at Stanford University, joined Google in 17 years and became the head of Google Cloud AI, leaving his job a year later to return to Stanford.
For enterprises, there is no optimal solution to this problem.
Reference Links:
[1]https://www.nature.com/articles/s41586-021-03544-w
[2]https://www.cnbeta.com/articles/tech/1264813.htm
[3]https://www.engadget.com/google-fires-ai-researcher-over-paper-challenge-132640478.html
[4]https://www.nytimes.com/2022/05/02/technology/google-fires-ai-researchers.html
[5]https://news.ycombinator.com/item?id=31235152
[6]https://static.poder360.com.br/2020/12/About-Googles-approach-to-research-publication.pdf