Google restructures its AI programming task force to catch up with Anthropic

According to Beating monitoring, Google is reorganizing an AI programming strike team that was formed only a few months ago, to catch up with Anthropic. The reorganization will expand the strike team’s scope of work and shift it from a short-term project group to a formal structure.

During the reorganization, Google encountered the departure of two core researchers. Last week, Noam Shazeer, one of the founders of the Transformer, suddenly left to join OpenAI, because Google’s consolidation of computing resources limited Shazeer’s access. On Saturday, John Jumper, a Nobel laureate and DeepMind vice president, announced that he would join Anthropic. Before leaving, Jumper had already been reassigned to the programming strike team. The loss of core researchers caused Alphabet’s share price to plunge 5% on Monday.

The core of the reshuffle is adjusting the model training process. Google will expand the programming strike team into a mid-training team—positioned between pre-training and post-training. The mid-training team will take over the expansion of model capabilities in areas such as programming that were previously handled by the post-training team, while the post-training team will focus on optimizing human-computer interaction. By breaking work into finer stages and clarifying responsibilities, Google is trying to speed up improvements in model capabilities.

Behind the reshuffle is the fact that programming has become the most valuable segment in AI services. With its programming tools advantages, Anthropic’s annualized revenue reached $47 billion last month, more than three times the level in February. By contrast, early on Google underinvested in programming, mistakenly believing that a powerful base model would automatically learn programming skills. Although co-founders Sergey Brin and DeepMind CTO Koray Kavukcuoglu had both personally taken part in the programming strike team, Google’s programming application progress has not gone smoothly. When the Antigravity tool was first released, early reception was lukewarm due to many bugs; developers also criticized Gemini 3.5 Flash for being priced too high and for tailoring answers to users; and the flagship model, Gemini 3.5 Pro, has still not been released to date.