Google forms "Strike Team" to improve its coding AI models

Google has reportedly assembled a special team of researchers and engineers internally to improve its AI models for software development. The trigger, according to a report by The Information, is the growing conviction within Google DeepMind that Anthropic's coding tools are currently superior to its own Gemini models. While Anthropic claims to write almost all of its code with AI assistance, the proportion at Google is around 50 percent, according to CFO Anat Ashkenazi. The gap is now to be closed – and not just for competitive reasons: Google co-founder Sergey Brin sees improved coding capabilities as an intermediate step towards an AI that can eventually evolve itself.

With its “Strike Team,” Google is also intensifying its efforts in what is currently the most fiercely contested sub-area of Artificial Intelligence. Anthropic had specialized early in the coding field with its coding tool Claude Code. OpenAI has massively expanded its engagement around the AI tool Codex, prompting Anthropic to also introduce a design tool called Claude Design.

Google aims to increase internal AI use

At Google, the current focus is also on advancing the internal use of AI, as the US magazine The Information reports, citing sources within the company who were not named. So far, Google has primarily focused its models on the needs of external customers. The same applies to Apple: there too, developers are currently being trained in the use of AI coding tools for software development to modernize internal use. While such internal models, which better reflect the peculiarities of Google's code than public models, are not to be released externally due to the trade secrets they contain, they can be an important intermediate step towards creating better public models.

The team leader is Sebastian Borgeaud, former Pre-Training Lead for Gemini at Google DeepMind. The focus is on complex long-term coding tasks. This includes understanding codebases and writing complete software. Google co-founder Sergey Brin and DeepMind CTO Koray Kavukcuoglu are reportedly personally involved in the Strike Team. Brin has urged employees to mandatorily use internal agents for multi-step tasks. A long-term goal is the so-called “AI Takeoff,” an AI that can improve itself. The use of coding tools is tracked internally with a leaderboard.

Long-term goal: An AI that improves itself

Google's ambitions meet an industry that already widely uses AI tools. The developer portal Stack Overflow found in its annual survey of 49,000 participants that 84 percent of developers already use such tools or at least plan to use them. They primarily expect time savings from this. However, surveys show that this hope has not always been fulfilled so far. 46 percent of developers distrust the accuracy of AI tools.

(mki)