Microsoft’s Copilot AI Gets a Voice, Vision, and a ‘Hype Man’ Persona

Microsoft’s Copilot Ai Gets A Voice, Vision, And A ‘hype Man’ Persona

Suleyman adds that Copilot might eventually critique a webpage, making a qualitative judgment based on a user’s interests and preferences. “It could actually just read the entire page in an instant and then talk to you about the page,” he says. “Like, you can ask ‘do you think this is an article I would enjoy?’ That’s kind of a different experience.”

Text interactions with Copilot are stored for 18 months, Microsoft says, although users can delete conversations. Copilot Vision will not keep a record of what users ask, Microsoft says, deleting the data at the end of a session. The feature will be limited to certain websites and will also be blocked from accessing copyrighted or NSFW content. It will be rolled out to Copilot Pro users in the US at an undisclosed date, Microsoft says. The company says no data is shared with OpenAI.

Another experimental feature called Think Deeper lets Copilot try to solve more complex problems through a process that mimics step-by-step reasoning. The technology is partly based on a new AI model called OpenAI o1 that was announced earlier this month by OpenAI. Think Deeper will be available to some Copilot Pro users in the US today.

The changes to Copilot are a sign of Microsoft’s desire to experiment with its AI tools and make them more compelling. They also reflect the rapid pace at which AI is developing, with most cutting-edge large language models—as the algorithms that power chatbots are called—now able to handle audio and imagery as well as text. OpenAI, Google, and others have in recent months all given their models the ability to converse naturally with different human voices.

Besides plenty of competition, Microsoft also faces some behind-the-scenes uncertainty.

The company has invested a reported $13 billion in OpenAI, and has a license that grants it access to its AI models. But although OpenAI is still widely considered a leader in AI, it has been beset by turbulence, most recently the departures last week of CTO Mira Murati and two senior engineers leading research efforts. Suleyman declined to comment on the situation at OpenAI.

Suleyman joined Microsoft in March, after the software giant signed a $650 million licensing deal with his startup, Inflection AI. He previously cofounded the British AI company DeepMind, which was acquired by Google in 2014. Last year, DeepMind was merged with Google’s AI effort to form Google DeepMind, and is now led by another DeepMind cofounder, Demis Hassabis.

Microsoft developed Copilot after seeing success with a tool for coders released in 2021, called GitHub Copilot. It autocompletes blocks of code and can answer programming questions.

Shane Greenstein, a professor at Harvard Business School who has studied Microsoft’s AI strategy, says it will be more challenging to design a useful general purpose helper; he adds that the company’s experimental new features still need to prove their value to users.

“It took five to 10 years of messing around with web interfaces to figure out how to get more than technically savvy people to buy something online,” says Greenstein of Harvard Business School. “I expect that type of time scale for the iteration here, too.”