
The Scoop
Google is replacing the leader of its consumer AI apps as the focus of the AI race shifts from the underlying models to the products built around them, according to memos reviewed by Semafor.
Sissie Hsiao, who led Google’s effort to create an AI chatbot, originally called Bard and now dubbed Gemini, will step down immediately. Josh Woodward, who leads Google Labs and oversaw the launch of NotebookLM, the company’s popular tool that turns text into a podcast-like show, will replace her.
In a memo to the staff, Google DeepMind CEO Demis Hassabis said the move will “sharpen our focus on the next evolution of the Gemini app.” He said Woodward will remain head of Google Labs while shaping the next chapter of Gemini.
Hsiao, in a separate staff memo, called her time as head of the team “chapter 1″ of the Bard story, and said she was optimistic in handing the baton to Woodward for “chapter 2.” Hsiao plans to take a “short break” and return to Google in a new role.
Hsiao, Woodward, and Hassabis declined to comment through a spokesperson.
Hsiao, a 19-year veteran of Google, jumped into action after ChatGPT upended the tech industry, forcing Google to accelerate the rollout of AI technology it had created.
While Google researchers had pioneered the transformer-based architecture that allowed the creation of large language models that powered ChatGPT, its unpredictable nature led the company to keep its chatbot experiments mostly under wraps.
In the competitive rush, mishaps illustrated why. The launch of Bard was criticized for “hallucinations,” and its Gemini image creation tool became a target of conservatives when it generated embarrassing results, inserting photos of women or people of color when prompted to create images of Vikings, Nazis, and the Pope.
But Google quickly moved past those early errors with successful product launches. Last week, Gemini 2.5 blew away AI benchmarks held by competitors like OpenAI and Anthropic. It took the lead in Chatbot Arena, where users vote on their favorite large language model responses.
In this article:
Know More
While the Gemini app took center stage, Woodward’s Google Labs team was creating rapid prototypes of new consumer products powered by the company’s AI models.
One experiment, NotebookLM, was a runaway success. Users can upload large text documents and the model turns it into something similar to an episode of a podcast, where two AI-generated hosts discuss the contents of the text.
Woodward’s team also built Project Mariner, an AI agent that can control the Chrome browser, navigate the web, and take autonomous actions, like filling out forms and gathering information. (Mariner hasn’t yet been released to all users.)
By placing Woodward in charge of Gemini, Hassabis is hoping he will help capitalize on the company’s research prowess by finding ways to wrap user-friendly products around sophisticated models.

Reed’s view
The leadership change at Gemini, which the company internally refers to as the “Bard team,” reflects a new phase of the AI race, where the product scaffolding built around AI models is just as important as the models themselves.
That is, in part, because models have reached a level of capability where they can be used for more than novel experiments. Companies like Cursor (for coders) and Harvey (for lawyers) are building successful businesses by packaging AI in a way that makes the technology more useful and accessible to consumers.
Last week was a perfect illustration of the phenomenon. Google released Gemini 2.5, arguably the best AI model in the world, but it received less attention than OpenAI’s new image generator, a discreet product that went viral on social media.
Hsiao was a good choice to lead Bard as the company was being forced to go from a large, deliberate giant into something resembling a fast-moving startup. Her long tenure at the company could serve as a bridge between the two cultures.
But there is no doubt that the transition period is over. It’s time for the company to start throwing a lot more spaghetti against the wall and build a lot more NotebookLMs and Project Mariners.
At the same time, Google DeepMind’s research team is on a trajectory that could give it an advantage. The company made an early decision to build Gemini as a natively multimodal app, training it not just on text but on all multimedia.
That slowed Gemini down in the beginning, but the decision is paying off. Google’s recent launch of Gemini Robotics shows how multimodal models are the likely path toward better reasoning.
There is a whole genre of consumer products that could take advantage of that early lead in multimodal AI models.
Google’s vertically integrated, custom AI chips could help bring product ideas to market faster by reducing the costs of inference, which have so far made some ideas cost prohibitive.

Notable
- WIRED chronicled the last two years at Google, focusing in part on Hsiao’s ambitious mandate to rapidly build an AI chatbot to compete with ChatGPT.