The week in AI: is China's Ernie 4.0 as good or better than GPT-4?

Presented by

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools or services, and exciting projects in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

Our takes on Ernie 4.0 and Baidu’s claim:

• Even before US export controls on advanced AI chips to China, Baidu had mass produced its advanced Kunlun II AI chip with the help of Samsung. Baidu is one of the most well-resourced tech companies on the planet, with a deep bench of AI talent. It shouldn’t be too surprising that they have an outstanding AI language model - especially considering how far along the open source/global community is in this space.

• This isn’t to say the US sanctions haven’t been effective: it’s telling that Ernie only went public a month and a half ago - nearly a year after ChatGPT. This is particularly surprising because back in 2019-21, Baidu appeared to be ahead of or on par with anyone else, including OpenAI, in LLM development. China’s generative AI regulation only went into law two weeks before Ernie went public, so it wasn’t the CCP preventing Ernie from going public before then - more likely it was a) Baidu’s fear that the chatbot wasn’t up to snuff and b) a lack of resources to train and deploy at scale - a multimodal chatbot like Ernie uses a massive amount of compute, something the sanctions have likely impacted.

• The ‘commercial’ seems disingenuous. At the beginning of the linked video with the commercial, you can see an Ernie-generated image of the same A07 EV. It’s not quite photo-realistic - it has that hyper-stylized characteristic of advanced AI image generators. AI-generated videos have been advancing steadily for months; but this looks much more like pieces of stock footage being attached to each other, or a one-off PR showcase.

• Ernie is still not available for use anywhere outside China. ChatGPT is available everywhere on the planet except China, North Korea, Iran, Syria, and Cuba.

It’s hard to say anything conclusively. Without widespread access to models like Ernie and better benchmarks for AI systems, we’re left to much speculation about China’s AI capabilities. It does appear that they are ‘mostly’ keeping pace, but that export controls have had an effect.

In related news this week: the Biden administration has further tightened restrictions on exporting advanced artificial intelligence chips and manufacturing equipment to China. Commerce Secretary Gina Raimondo continues to reiterate that the goal is limiting China's military access to cutting-edge semiconductor technology, not harming their economy. The vast majority of semiconductors will remain unrestricted.

Firefox is integrating a new AI feature into its browser to help online shoppers detect fake reviews. This new feature (powered by Fakespot (a company Mozilla acquired back in May) will analyze product reviews on sites and assign a letter grade indicating the reliability of the reviews from A to F. Google’s algorithm will push well-reviewed items, but can’t itself differentiate between real reviews and a fake ones; hopefully this feature will help more users navigate murky online review waters.

Fakespot will be an auto-update as part of Firefox 120 for Android and desktop; you can download the Fakespot browser extension for Chrome if you aren’t a Firefox user. It currently works on Amazon, Walmart, eBay, Yelp, and TripAdvisor.

The author wants her daughter to grow up in a world where people control technology - not vice versa. She concludes that aligning AI with human values, while challenging, is possible if we resolve disagreements about the best path forward. Ultimately, the mother hopes for an abundant future where AI assists rather than replaces, and where freedom and opportunity remain in human hands.

A new study published in Mayo Clinic Proceedings: Digital Health found that using AI to analyze just 10 seconds of a person's voice, along with basic health data, can predict if they have type 2 diabetes with up to 89% accuracy. The non-invasive screening method analyzing vocal features could allow much easier access to diabetes testing and detection, as current diagnostic methods like blood tests can require significant time, travel and costs.

Diabetes diagnosis is associated with an increased risk of mortality from: cancer, renal disease, infections, liver disease, nervous system disorders, and chronic obstructive pulmonary disease. With nearly half of the over 400 million diabetic adults worldwide undiagnosed, the voice analysis technology could revolutionize screening and early identification of type 2 diabetes.

New York Mayor Eric Adams has been using AI to making robocalls to residents in languages he doesn't speak, leaving some constituents confused and wary of the ethics behind the technology. City Hall is using voice cloning artificial intelligence to create audio public service announcements that sound just like Adams. The mayor says people frequently assume he speaks those languages based on the convincing AI clones of his voice. Adams argues the technology allows him to communicate with more New Yorkers, brushing off concerns that failing to disclose the AI origins could be misleading or problematic.

More in AI this week:

• How ChatGPT is transforming the postdoc experience

• Millions of workers are training AI models for pennies

• Foxconn and Nvidia are building ‘AI factories’ to accelerate self-driving cars

• Rapper Pras Michel of Fugees seeks new trial, contends former attorney used AI for closing argument

• Atlanta’s first AI supermarket opens

• YouTube gets new AI-powered ads that let brands target special cultural/holiday moments

• Margaret Atwood reviews a “Margaret Atwood” story by AI

• (Google Cloud) Protecting customers with generative AI intellectual property indemnification

• After ChatGPT disruption, Stack Overflow lays off 28 percent of staff

Read The Information first…or them later

The Information’s journalists get the news that other newsrooms envy. That’s why their reporting on tech, media and finance has been followed by the New York Times, Bloomberg, the Wall Street Journal and dozens of other outlets more than 400 times in the last year.

Get the news straight from the source.

That’s it for AI this week. We’ll see you next Thursday!