Good news: Google's new AIs won't reach the public anytime soon

In the technology sector, few habits are as common as vaporware – present products long before they are ready for launch. In some cases, this happens because companies simply underestimate the time required to complete development.

In others, it is a deliberate strategy to create anticipation around something new, diverting focus from competing offerings. Either way, the result is always the same: products that take time to reach – or never reach – the market.

In the current race for leadership in artificial intelligence, this has been frequent. It is no surprise, therefore, that Google's new projects – Astra and Mariner – are not yet available to the public.

For now, Google DeepMind, the company's AI research arm, is only making these tools available to a select group of “trusted testers”. However, classifying them as another example of “vaporware” would be unfair.

Google has been transparent. Your goal is to understand how people would actually use these new technologies before offering them widely. Considering some of the problems the company has had with past releases, this strategy seems pretty sensible.

Astra is a vision of the future of AI assistants. It is not limited software like Google Assistant or a text-based chatbot like Gemini, but rather an assistant that listens, speaks and “sees” the world around it.

It's comparable to ChatGPT's Advanced Voice Mode, released by OpenAI in May — although its camera capabilities aren't yet off the ground.

Mariner is a Chrome extension that automates web browsing, typing and clicking to perform tasks that you would do manually.

Considering some of the problems the company has had with past releases, this strategy seems pretty sensible.

This idea is reminiscent of Anthropic's “Computer Use” feature, introduced in October as part of the Claude model, which allows the chatbot to control applications. Both tools are steps towards the future of autonomous AI, capable of acting independently on the user's behalf.

Google's learnings from Astra and Mariner could be as relevant to the quality of future experiments as the technical capabilities of the Gemini model – yet another sign that artificial intelligence is finally starting to deliver practical results.

“You benchmarks academics remain important, but today, when we say something is 'best in class', we are talking about the user experience. What matters is: do they think it's the best?” explains Koray Kavukcuoglu, CTO of Google DeepMind. “The capabilities of the model need to be integrated with the functionality and utility of the application – this is a big change for researchers.”

“When we say something is 'best in class,' we're talking about the user experience. What matters is: do they think it's the best?”

Of course, enthusiasts might find it fun to use AI for things like shopping, even if it doesn't save time. However, initial testers are generally people who are more familiar with technology than the general public, which limits what the company can learn from them.

“We are currently working with testers who already have experience with AI, as we are still in the early stages of the project,” comments Helen King, senior director of accountability at Google DeepMind.

“But as we move forward, it will be essential to include a greater diversity of voices: experts, academics and, above all, the general public. We want these tools to be accessible to everyone, not just those who have already mastered AI,” adds King.

“We are currently working with testers with AI experience because we are still in the early stages of the project.”

Everyone involved in the Astra and Mariner projects emphasizes that the tools will continue to evolve as Google learns from testing.

“We assembled the team to enable rapid exploration. We are focused on this”, reinforces Kavukcuoglu. The true proof of the value of these technologies will be in the final products that Google will release. But first impressions indicate that these tools seem promising and well-founded.

---

ABOUT THE AUTHOR

Harry McCracken is Fast Company's technology editor based in San Francisco. In past lives, he was editor of Time, founder and editor… find out more

---