[Photo: Google DeepMind]
Google DeepMind has formed a dedicated team to improve its AI coding models, The Information reported on April 20, citing three sources familiar with internal matters.
Anthropic is said to have had a big influence on Google’s moves. DeepMind researchers have assessed Anthropic’s coding tools as ahead of Google’s Gemini model in code-writing ability, The Information said.
The team is led by DeepMind research engineer Sebastian Borgeaud. He previously led pre-training of AI models at DeepMind. The team is focusing on boosting model performance in complex coding tasks that take a long time, such as writing new software.
Google co-founder Sergey Brin and DeepMind Chief Technology Officer Koray Kavukcuoglu (코레이 카부코글루) are also said to be directly involved with the team.
In a recent memo to DeepMind employees, Brin stressed that “the final battleground in the AI competition is agents.” He said it is necessary to quickly narrow the capability gap in agents and make models the主体 of development. He meant models should be able to write code directly, rather than serving as simple support tools.
Google is also increasing its focus on developing coding models for internal use. It appears to be prioritising internal models trained on Google’s own code over coding models for external customers.
It is also urging employees to expand use of an A tool. Google is operating an internal leaderboard that tracks usage of its internal coding tool, Jetski. In the memo, Brin stressed that all Gemini engineers must use internal agents for complex multi-step tasks.
Anthropic is already handling most internal engineering work with coding tools. Boris Cherny, head of Claude Code, said in January that “Anthropic’s code is effectively 100 percent written by AI.” By contrast, Google Chief Financial Officer Anat Ashkenazi said at an earnings release in February that coding agents write about 50 percent of the company’s total code.