Google and Anthropic: The AI Battle for Programming Dominance

Google Deepmind Strengthens Its Programming Capabilities

Google Deepmind has recently taken a strategic initiative by forming a specialized team to enhance the programming skills of its Gemini models. This decision aims to tackle the challenge of creating new software from scratch, a complex task that requires a deep understanding of user needs and the ability to read and interpret files autonomously.

The origin of this approach lies in an internal assessment that revealed the programming tools developed by Anthropic currently outperform those of Google. To bridge this gap, Google has intensified the training of its AI models on its own internal code. Additionally, the company is monitoring the use of its programming tools by employees and has made AI training mandatory in certain teams.

An Elite Team for an Ambitious Challenge

According to reports from The Information, Google Deepmind has assembled an elite team of researchers and engineers to refine the programming skills of its Gemini models. This team is led by Sebastian Borgeaud, a Deepmind engineer who previously oversaw the pre-training of the company's models.

The primary goal of this team is to tackle complex, long-term programming tasks, such as writing new software from scratch. This approach is driven by the belief that Anthropic's programming tools are currently more effective, prompting Google to ramp up its efforts to catch up.

Competition Intensifies in the Programming Field

Programming has become a major competitive arena for large AI labs, with players like OpenAI and Google striving to catch up with Anthropic. Recently, OpenAI halted its Sora video generator project to free up computing resources, allowing it to focus on training and executing other AI models.

Sergey Brin's Direct Involvement

Google co-founder Sergey Brin, along with Deepmind's CTO Koray Kavukcuoglu, is actively involved in this effort. In an internal memo, Brin emphasized the urgency of closing the gap in agentic execution and transforming models into primary code developers. He also insisted that every engineer working on Gemini utilize internal agents to accomplish complex, multi-step tasks.

Brin informed employees that enhancing programming skills is a crucial step toward creating AI capable of self-improvement. An advanced programming agent, combined with an AI capable of solving mathematical problems and conducting experiments, could potentially automate much of the work currently performed by AI researchers and engineers.

Internal Monitoring and Training

Google has implemented a monitoring system for the use of its internal programming tool, "Jetski," and ranks teams based on their usage. This system is similar to that of Meta, which uses tokens as a metric. Some teams, outside of Deepmind, also require engineers to participate in AI training sessions.

According to sources from The Information, Google is increasingly relying on models trained on its own internal code. This internal codebase is very different from the public code typically used to train general-purpose programming agents, meaning that these internally trained models cannot be released. However, they could help Google develop better models that will eventually be made available to users while accelerating internal development.