Two additional researchers from Gemini are departing Google to join Anthropic.

      According to a Bloomberg report from June 24, two prominent AI researchers from Google are set to transition to Anthropic, marking the latest exits from the team responsible for Gemini and the second pair to join the Claude maker within a week. This time, the individuals are Jonas Adler and Alexander Pritzel, both seen internally as vital contributors to Google’s flagship model. Adler focused on the company’s AI coding initiatives, while Pritzel specialized in pretraining—the initial phase where a model learns from vast data sets.

      At the time of reporting, neither departure had been officially confirmed, and the sources cited by Bloomberg requested anonymity. Google and Anthropic did not provide comments. However, it is evident that the trend has been consistently moving in this direction for some time.

      In recent weeks, Google would prefer to overlook the events that transpired. Last week, Noam Shazeer, a vice-president of engineering and co-author of the influential 2017 paper “Attention Is All You Need,” which laid the groundwork for the Transformer architecture pivotal to the field, departed for OpenAI. Shortly after, John Jumper, who won the 2024 Nobel Prize in Chemistry for AlphaFold, declared he was leaving Google DeepMind after nearly nine years, also headed to Anthropic.

      Jumper’s exit alone resulted in a loss of around 6% in Alphabet’s share price, equating to over $245 billion in market value, within a single trading session. The four high-profile departures within just a few days suggest a trend that could lead to broader implications, with Anthropic emerging as a clear beneficiary.

      Anthropic has been aggressively hiring from Google DeepMind as it expands beyond general-purpose models into areas like coding, healthcare, and scientific applications, where talents like Jumper and Pritzel are highly sought after. This exodus contributes to a growing perception that Google’s dominance in sectors it once expected to fully control is diminishing.

      The ongoing movement of talent is driven not only by financial incentives, although that is a significant factor. Both Anthropic and OpenAI are nearing potential public offerings, offering a unique opportunity that may not be as accessible at a company already valued at two trillion dollars: the chance to acquire equity before going public and share in the potential upside.

      For researchers considering job offers, an equity package before an IPO at a rapidly growing private lab can surpass a comfortable salary at a larger corporation whose stock is already valued for success.

      Access to computing resources also plays a crucial role in this equation. Within Google, securing access to the company's proprietary chips has become a challenge for researchers, as external demand for its tensor processing units (TPUs) consumes much of the available capacity that could have been allocated to internal projects. A researcher unable to obtain the necessary hardware for experiments has a compelling reason to consider opportunities from competitors, with Anthropic being a notable external customer of that TPU capacity.

      For Google, the immediate implications are largely reputational, rather than operational. Gemini remains a competitive model, and the loss of two individuals does not negate its standing. However, a leading lab relies on a limited number of skilled individuals capable of conducting the toughest tasks, and losing four such talents in a week to its top rivals is a situation that may catch the attention of other researchers contemplating their future.

      The upcoming months will reveal how Google addresses its recruitment challenges rather than product-related issues. The company has the means to fill the vacancies and a historical track record of attracting talent, and it has previously managed workforce defections. The pressing question is whether it can mitigate a departure rate that increasingly resembles a trend rather than a series of isolated choices.