When asked about Chatterjee, Google spokesperson Jason Freidenfelds issued a company statement confirming it was “terminated for a reason.” Freidenfelds also issued a statement from Zoubin Ghahramani, vice president at Google Research, saying that “we firmly maintain our standard of respectful discourse among our researchers.” Chatterjee was not mentioned by name in Gharamani’s statement.
The episode adds to a series of recent internal conflicts at Google that suggest that the freewheeling, engineer-centric culture it fostered as a startup has left the company unprepared for a number of challenges as a multinational corporation with more than 100,000 employees.
Google hired Satrajit Chatterjee in 2018 as a senior researcher in machine learning. Before that, he was senior vice president at hedge fund Two Sigma and also worked at Intel. When Chatterjee joined Chatterjee, Mirhoseini and Goldie were already working in the company’s premier machine learning lab, Google Brain. Chatterjee joined a separate, smaller research group within Google’s research division.
The two women did not work directly with Chatterjee, but in 2019, Goldie’s internal document claims, he asked to manage the Morpheus project. After being politely rejected, employees say, Chatterjee began expressing doubts about the couple’s work with senior researchers they needed to collaborate with or get support from, suggesting their work was wrong or even fabricated.
As a senior employee, Chatterjee’s questions could be influential. As a result, employees say, other senior employees sometimes became skeptical of Goldie and Miroseini’s work and doubted their results.
The effect was that Miroseini and Goldie’s work at Google turned into a stressful, split reality, insiders claim. While running a successful project with support from Google’s chip designers, they said the pair had to do extra work to respond to allegations that their results were wrong or even untrue.
Chip design teams at Google and elsewhere are generally cautious by nature, as nanoscale fabrication is expensive and any flaws in a chip cannot be rectified once it has been cut into silicon. Google has said TPUs have enabled breakthroughs in its AI research and services, renting the chips through its cloud unit. But Chatterjee’s criticism of Morpheus continued even after Google’s hardware leaders decided they were confident enough to let it help design the company’s next-generation TPUs.
In May 2021, a Google employee posted to an internal email list asking if anyone had applied machine learning to designing circuit boards. Mirhoseini replied that Morpheus could help. But Chatterjee stepped in, claiming that older techniques outperformed machine learning tools and that commercially available chip design tools produced the best results.
Jeff Dean, Google’s head of AI, joined the discussion to say that Morpheus was already being used to design the next generation of TPU chips. The technology had won in extensive tests against human chip experts and commercial chip design tools, Dean said, also including a slideshow of the results.
Dean also linked to the recently published peer-reviewed Nature study. It reported that the Morpheus team’s code formats blocks of TPU circuits better than Google engineers using commercial chip design tools. The authors did not disclose the details of those chip segments, saying they were confidential to Google, but also included results for an open source processor design that was freely available to everyone. The results of the paper were later replicated by another research team at Google, and the code for the experiments was open source.
