Technology

➚ MOVE FAST: Bouncing back. OpenAI is relaunching its robotics unit after the team wound down in 2020. The capabilities of multimodal large language models have dramatically improved, and it’s looking to hire employees to develop new systems that can run on machines built by other companies.

➘ BREAK THINGS: Dying down. GM autonomous unit Cruise is driving slowly and more carefully; its cars won’t be back on the roads anytime soon. The company was the first to roll out an autonomous taxi fleet in the US, but is now falling behind its competitors following an incident where one of its cars dragged a woman underneath its wheels.

Media Circus of the Century, a Buzzfeed Comeback and Sleeping with Your Phone

The debut episode of Mixed Signals from Semafor Media, presented by Think With Google, is ready for your ears. Ben Smith and Nayeema Raza catch up with New York Times reporter Maggie Haberman and former presidential candidate Vivek Ramaswamy to talk about the media circus of the century, the future of BuzzFeed, and the case for sleeping with your phone, according to Editor Max Tani.

Listen wherever you get your podcasts

Hallucinations were not what drove Google’s AI Overviews feature to generate ridiculous answers to users’ queries. The Gemini-based large language model did not make up text, and, instead worked exactly as it should have even though it made mistakes, the company said.

It’s a reminder that people can’t always trust everything they read on the internet. But AI Overviews is still useful because it can generate clear answers so people don’t have to click on websites and sort through information manually .

It summarizes text from top webpages on Google Search, and can produce inaccurate information if there aren’t many details online about a particular topic or if what’s out there is already wrong.

Posts where AI Overviews told people it’s okay to smoke cigarettes when pregnant or to jump off the Golden Gate Bridge if depressed were faked, according to Google. Even though they were untrue, however, they still went viral online anyway. Pushing AI onto consumer products as widely used as Google Search is risky and the technology is never perfect even though it works exactly as it should be.

— Katyanna Quach

Inside DeepMind’s effort to understand its own creations

THE SCOOP

With missteps at industry leader OpenAI possibly providing an opening for rivals touting safety advances, Google DeepMind unveiled fresh details of how it’s building systems to catch potentially dangerous leaps in artificial intelligence capabilities.

OpenAI has tried to reassure the public, announcing a new safety committee earlier this week, after a top safety researcher joined rival firm Anthropic. That move came before actress Scarlett Johansson accused Sam Altman’s firm of using her voice without her permission for ChatGPT.

With AI guardrails becoming a possible competitive advantage, Google DeepMind executives told Semafor that the methods for predicting and identifying threats will likely involve a combination of humans and what the company calls “auto evaluations,” in which AI models analyze other models or even themselves.

The effort, though, has become particularly challenging, now that the most advanced AI models have made the jump to “multimodality,” meaning they were trained not only on text, but video and audio as well, they said.

“We have some of the best people in the world working on this, but I think everybody recognizes the field of science and evaluations is still very much an area where we need additional investment research, collaboration and also best practices,” said Tom Lue, general counsel and head of governance at Google DeepMind.

Google, which released a comprehensive new framework earlier this month to assess the dangers of AI models, has been working on the problem for years. But the efforts have ramped up now that foundation models like GPT and DeepMind’s Gemini have ignited a global, multibillion dollar race to increase the capabilities of AI models.

The challenge, though, is that the massive foundation models that power these popular products are still in their infancy. They are not yet powerful enough to pose any imminent threat, so researchers are trying to design a way to analyze a technology that has not yet been created.

When it comes to new multimodal models, automated evaluation is still in the distant horizon, said Helen King, Google DeepMind’s senior director of responsibility. “We haven’t matured the evaluation approach yet and actually trying to automate that is almost premature,” she said.

REED’S VIEW

It’s in every AI company’s interest to push hard on the “safety” front. The same effort to make AI safe will also help make AI models more reliable. And right now, reliability is the key factor that is holding the technology back.

As impressive as these new models are, they could, at any moment, embarrass a company or make an error. The fact they must be babysat makes them unusable for any task that is important.

Google and other makers of foundation models know this and they’re working overtime to address it.

The other interesting factor is that the models are somewhat commoditized. A small handful of companies are racing to become the best one, but from the end user perspective, they are almost interchangeable.

Most businesses don’t even want to use the most capable models because they are slower and costlier. As AI companies look for ways to differentiate themselves. AI “safety” is one of the biggest ways to do this.

Thanks to our friends at Interconnects, it’s possible to understand how the latest AI tools actually work. Their weekly newsletter covers the cutting edge models, what they can do, what they can’t, and what the researchers building them think about the future. Subscribe for free.

The number of short films made with OpenAI’s video tool Sora that will be shown at the Tribeca Festival in two weeks. The filmmakers, who were given early access to the technology, will abide by the AI terms of the contract reached last year between Hollywood actors and major movie studios.

State actors from China, Iran, Israel, Russia and other countries have used generative AI tools from OpenAI and Meta to launch disinformation campaigns online. Concerns that fake news and propaganda from foreign adversaries have risen ahead of the US presidential election.

But despite improved capabilities to create realistic-looking content across different types of media at scale, these campaigns haven’t really taken off. Meta said it shut down hundreds of Facebook accounts, while OpenAI has removed five users that have posted fake political comments, articles, or images across various platforms, including social media and chat apps.

It’s difficult to coordinate effective disinformation campaigns, considering there isn’t much to do other than spam the internet with AI-generated content. These fake posts probably get lost amongst all the other trash being uploaded to social media.

— Katyanna Quach

• Inside the NATO debate over protecting Ukraine’s skies.
• The reason why Donald Trump’s conviction might matter to voters.
• A taxonomy of Republican rage against the Trump verdict.