Weekly AI Recap: Google retools AI Overviews, OpenAI announces safety committee

The feature, which was released to all US users earlier this month, leverages generative AI to scrape the web for relevant information and provide summarized responses to queries at the top of the search results page. Some publishers have protested, claiming that the new tool will divert valuable web traffic.

Last week, a flurry of screenshots started showing up across social media showing AI Overviews apparently providing inaccurate, bizarre and in some cases potentially dangerous information. In one instance, the system responded to a search about how to prevent cheese from slipping off of a pizza by recommending that Elmer’s glue be mixed in with the sauce. The apparent source? A jokey comment on a Reddit thread from more than ten years ago.

Google has tried to provide an explanation, along with a plan for fixing the problem. “We hold ourselves to a high standard, as do our users, so we expect and appreciate the feedback, and take it seriously,” Liz Reid, the company’s head of search, wrote in a blog post published Thursday.

Reid pointed out, for example, that the AI model began acting unexpectedly when it started receiving absurd queries – like, “How many rocks should I eat?” – which people wouldn’t normally search for (and which were obviously intended to generate an embarrassing response). In that particular case, the system recommended “at least one small rock per day.”

In response, Reid says that the company has “built better detection mechanisms for nonsensical queries that shouldn’t show an AI Overview, and limited the inclusion of satire and humor content.”

This isn’t the first time that Google has rolled out an AI-powered feature, only to have to go back to the drawing board after inciting a public uproar. In February, it paused Gemini’s ability to generate images of people after users reported that the model created historically inaccurate images and, as some have suggested, was prone to avoiding representations of light-skinned people.

OpenAI training a new model & debuting ‘safety and security committee’

OpenAI has “recently begun training” the AI model that will eventually succeed GPT-4, predicting that “the resulting systems [will] bring us to the next level of capabilities on our path to AGI,” the company announced in a blog post on Tuesday.

The news follows the first-ever public demo of GPT-4o – a model capable of responding to text, audio and visual cues – earlier this month. GPT-4o comes with a variety of human-like voices, one of which, according to the actor Scarlett Johanssen, sounds “eerily similar” to her own. Johanssen wrote in a statement that OpenAI CEO Sam Altman had twice asked her to provide the voice for GPT-4o, and that she had declined both requests. OpenAI has denied that the voice has been modeled on Johanssen, but has paused the use of the model’s voice “out of respect for Ms. Johanssen.”

The Tuesday blog post also announced the formation of a “safety and security committee” within OpenAI. The committee, led by Altman and his fellow company boardmembers Bret Taylor, Adam D’Angelo and Nicole Seligman, “will be responsible for making recommendations to the full board on critical safety and security decisions for OpenAI projects and operations,” according to the company.

Ilya Sutskever and Jan Leike, who formerly led OpenAI’s superalignment team – tasked with ensuring that AI models do not behave contrary to human interests – recently resigned from the company. In a post on X, Leike accused his former employer of putting safety on the backburner while it ran headlong towards the development of “shiny products.”

Leike announced in another X post on Tuesday that he had joined the superalignment team at Anthropic.

I'm excited to join @AnthropicAI to continue the superalignment mission!

My new team will work on scalable oversight, weak-to-strong generalization, and automated alignment research.

If you're interested in joining, my dms are open.

— Jan Leike (@janleike) May 28, 2024

The Atlantic and Vox Media sign deals with OpenAI

OpenAI’s roster of publishing partners continues to lengthen.

Earlier this week, both The Atlantic and Vox Media announced that they had signed separate, multiyear content licensing agreements with OpenAI. Both deals give the AI company access to the publishers’ content in order to train large language models, like the one that’s currently powering ChatGPT. The publishers, meanwhile, now have permission to incorporate OpenAI’s technology into new products.

Founded in 1857, The Atlantic is one of the longest-running publications in the United States. Vox Media owns several major online news outlets, including Vox, The Verge and Vulture.

Financial terms for the deals haven’t been disclosed, but as a report from Axios pointed out, the fact that other publishing companies have been compensated in their own deals with OpenAI makes it likely that a similar arrangement is at play here.

News Corp, owner of The Wall Street Journal and some other major publications, announced last week that it had signed its own multiyear content licensing deal with OpenAI.

Musk’s AI startup raises $6bn

xAI, the AI company founded by billionaire Tesla CEO Elon Musk last summer to compete with OpenAI, announced on Sunday that it had raised $6bn in its series B funding round, bringing its total valuation up to around $24bn.

The company revolves largely around Grok, a chatbot that’s been positioned as a less ’woke’ alternative to competitors like ChatGPT. Since Grok’s debut, xAI has been pushing to quickly expand the chatbot’s capabilities.

“xAI will continue on this steep trajectory of progress over the coming months, with multiple exciting technology updates and products soon to be announced,” the company wrote in a blog post on Sunday. The funds from the [series B] round will be used to take xAI’s first products to market, build advanced infrastructure, and accelerate the research and development of future technologies.”

The company added that it’s currently hiring for a number of roles, seeking professionals who are looking “to join a small team focused on making a meaningful impact on the future of humanity.”

Nvidia’s ascent reaches a new peak

American chipmaker Nvidia, whose GPUs have become an immensely valuable building block in the race among big tech companies to build generative AI-powered products, is inching closer to overtaking Apple as the second most valuable company in the world.

Nvidia’s shares hit a record high on Tuesday, leaving the company’s total valuation about $100bn shy of Apple’s, Reuters reported. At the time of this writing, Nvidia’s valuation stands a little below $2.7tn, while Apple’s is about $2.9tn, according to stock market tracker Companiesmarketcap.com.

Microsoft overtook Apple as the most valuable company in the world in January, an achievement that was driven in large measure by the company’s quick efforts to invest in OpenAI and commercialize new AI products.

For more on the latest happenings in AI, web3 and other cutting-edge technologies, sign up for The Emerging Tech Briefing newsletter.