DeepMind founder Demis Hassabis said he expects Google will be able to “unpause” the portrait feature of Gemini, the company's multimodal generative AI tool. The ability to respond to human image prompts should be back online in the “coming weeks,” he said today.
Google announced last week that the Gemini feature was being rolled out after users said the tool was producing historically inconsistent images, such as depicting the Founding Fathers as a diverse group of people rather than just white men. was temporarily suspended.
Hassabis was responding to questions about product snafus during an on-stage interview at Mobile World Congress in Barcelona today.
Asked by host Wired reporter Stephen Levy to explain what went wrong with the image generation function, Hassabis declined to provide a detailed technical explanation. Instead, he suggested that the problem was caused by Google's failure to identify instances where users were essentially looking for what he described as a “universal depiction.” The example also shows “the nuances that come with advanced AI,” he said.
“This is an area that we're all working on. For example, if you type in a prompt that asks for a photo of a person walking a dog or a nurse at a hospital, then obviously there's some type of ' We need a universal depiction. Especially considering that Google serves over 200 countries, i.e. countries all over the world, it is important to understand where the user comes from, what his background is, and what his situation is. I don't know if there are any. I want to show the possibility of a kind of universal scope. ”
Hassabis said the problem ultimately lies in the Gemini people's “well-intentioned function” of promoting diversity in image output being applied “too blatantly across the board.”
He added that prompts asking for content about historical figures “naturally” should have “a much narrower distribution of returns,” hinting at how Gemini might approach prompting people in the future. .
“Of course, we value historical accuracy, so we've kept that feature offline while we fix it, but next time we hope to bring it back online very soon. In the coming weeks. , another few weeks.”
When asked a follow-up question about how to prevent the misuse of generative AI tools by bad actors, such as authoritarian regimes seeking to spread propaganda, Hassabis did not have an easy answer. He suggested the issue was “very complex” and would likely require the mobilization and response of society as a whole to determine and enforce restrictions.
“There needs to be some really important research and discussion that involves not only technology companies but also civil society and governments,” he said. “This is a social and technical problem that affects everyone, and it needs to be discussed by everyone. What value do we want these systems to have? What does that represent? You're talking about someone with malicious intent accessing the same technology, and using them for harmful purposes not intended by the creators of those systems. How do we prevent it from being reused?”
Touching on the challenges of open source general-purpose AI models, which Google also provides, he added: “Customers want to use open source systems that they have complete control over…but the question is how do you ensure that what people are using downstream is not?” Isn't it harmful because it's getting more and more powerful?
“Today, I don't think that's a problem because the system is still in a relatively early stage. But if you look out three, four, five years and have the ability to plan and be active in the world and solve problems, I think society has to think seriously about these issues once we start talking about next-generation systems that can solve our goals.If this becomes widespread, bad actors from individuals to rogue states could take advantage of them as well. there is.”
During the interview, Hassabis was also asked about his thoughts on AI devices and where the mobile market is headed as generative AI continues to drive new developments. He predicts a wave of “next-generation smart assistants” that will be useful in people's daily lives, rather than “gimmicky” ones like previous generations of AI assistants. Packing mobile suggested that it might even change the shape of his hardware. about their personalities.
“I think it also raises the question of what type of device is appropriate,” he suggested. “But in five years or more, will mobile phones really be the perfect form factor? For AI systems to actually be a little more aware of the context you're in, so they can be even more useful in your daily life, whether it's glasses or something else. Things might be needed. So I think all sorts of amazing things will be invented.”
