The Prompt: Google’s AI Suggested Adding Glue To Pizza

The Prompt is a weekly rundown of AI’s buzziest startups, biggest breakthroughs, and business deals. To get it in your inbox, subscribe here.

Welcome back to The Prompt.

Today, OpenAI announced that it recently began training a new AI model that would succeed GPT-4 (the model powering ChatGPT) and produce what it calls the “next level of capabilities” (though it didn’t say what those capabilities are.) The AI company also said it is creating a new Safety and Security Committee to assess the risks posed by the technology and make an initial set of recommendations to mitigate them within the next 90 days. The committee will be led by the company’s board members, including Bret Taylor, Adam D’Angelo, Nicole Seligman, and CEO Sam Altman. OpenAI said the committee will consult external experts like Rob Joyce, former cybersecurity director at the National Security Agency and John Carlin, a former Department of Justice official.

Now, let’s get into the headlines.

ETHICS + LAW

A fashion designer’s body was deepfaked using AI tools in order to sell rip-offs of her own products on platforms like Amazon and Walmart. Forbes spoke to Cassey Ho, CEO and founder of fashion brand Popflex and the creator of a viral clothing item, the “pirouette skort,” which she said is being counterfeited. Ho said she found the deepfake “incredibly violating.” Photos of her brand’s models and customers were also stolen and posted on Amazon listings to sell the counterfeits, she said.

MARKET MOVES

Chipmaker Nvidia had a terrific week on Wall Street. In an earnings call on Wednesday, the AI giant reported that its profits grew 630% and revenue grew about 270% from the same time last year, making it Nvidia’s most profitable and highest sales quarter ever. The company’s stock hit a record high on Tuesday and peaked above $1,000 per share for the first time. (The company’s stock closed at $1,064 on Friday.) The stock’s rally added over $10 billion to CEO Jensen Huang’s net worth, putting him in the top 20 in Forbes’ rankings of the richest people in the world.

HUMANS OF AI

Generative AI is increasingly finding its way into the world of real estate, whether it’s to design fantasy homes, write descriptions of listings or scan and process documents for home loan approvals. Now a new AI assistant called Sidekick aims to let thousands of real estate agents conduct research on properties, compare their prices and whip up descriptions of listings from uploaded photos. AI makes a powerful tool for real estate agents because they’re “salespeople, they’re not experts at data analysis which is something they have to do,” said Sidekick cofounder and CEO Michael Martin. Currently focused on the Miami and San Francisco areas, Sidekick is built on GPT-4 and trained on a massive property database called MLS.

AI DEALS OF THE WEEK

Defense AI company Helsing is in talks to raise $400 million at a $4 billion valuation, sources told Forbes. Led by General Catalyst, the round would bring the company’s total investment to $750 million. The Munich-based company has scored contracts to provide AI software to assist German fighter jet pilots and develop AI capabilities for drones in Ukraine.

Germany-based AI translation startup DeepL has raised $300 million at a $2 billion valuation to focus on enterprise growth. The company’s proprietary translation model is used by over 100,000 businesses to translate documents into more than 30 languages, and it recently launched an AI business writing tool.

French AI startup H announced that it raised $220 million in seed funding to build AI agents for business applications. Founded by former Stanford and Google DeepMind researchers in 2023, the company is backed by Eric Schmidt, Accel and Amazon, among others.

DEEP DIVE

Days after Google plugged AI overviews into its flagship search engine, the platform used by billions across the world produced an array of misleading answers. Google told people they could use glue to ensure that cheese sticks to pizza, that they could stare at the sun for up to 30 minutes and that geologists suggest eating one rock per day. It also produced other inaccuracies like stating that former President Barack Obama is a Muslim and failed to find existing information, such as the names of countries in Africa that start with K. After stories about these mistakes went viral on social media, Google removed some of these answers. Part of the issue? Google’s AI search overviews appeared to be pulling from unvetted sources like Reddit posts and treated satirical sites such as The Onion as legitimate news sources.

Representatives at alternative search platforms like DuckDuck Go and You.com said that even though Google’s AI is spewing out inaccurate information, the tech giant maintains a strong monopoly in search, in large part because it is the default search engine on popular devices and browsers alike. “Even if the results are worse, or people don’t like the product, can they switch?” DuckDuckGo spokesperson Kamyl Bazbaz said in an email that highlighted the U.S. Department of Justice’s antitrust lawsuit against Google, which alleges the search giant pays Apple over $20 billion to be the default search engine for Safari.

You.com CEO Richard Socher also expressed frustration with Google’s seeming stranglehold on search, stating that although he’s seen a five to 10 percent jump in usage of his AI-powered conversational search engine after Google rolled out its AI-generated answers in search, it’s not enough to expand much market share. “I do think there’s an underlying current of people just being fed up with Google,” Socher told Forbes. “But it feels impossible. The monopoly is very entrenched.”

AI INDEX

1 out of 6

Responses generated by AI models for legal queries are incorrect, a new Stanford study shows.

75%

Responses to questions about a courts’ rulings created by general purpose large language models like OpenAI’s GPT 3.5, Google’s PaLM 2 and Meta’s Llama 2 are also incorrect, according to a previous study.