Technology

➚ MOVE FAST: 3-D. Leading AI researcher Fei-Fei Li has raised $230 million for a new venture, World Labs, which launches today. Its aim is to use images and other data to build software that gets how the three-dimensional real world works. Investors include actor Ashton Kutcher, former Google CEO Eric Schmidt, and Salesforce CEO Marc Benioff.

➘ BREAK THINGS: Second round. TikTok will make its case next week against a US law forcing parent company ByteDance to divest of the American version of the app or face a ban. A federal appeals court will hear oral arguments Monday. Just like when it faced similar restrictions during the Trump administration, TikTok will again assert that the rule is unconstitutional.

You’ve probably seen the AI demos from companies like OpenAI and Google where a chatbot can analyze a phone’s video in real time. You can try something kind of like that with a new app called Lloyd from a company called EndlessAI.

But Endless co-founder Thomas Pompidou tells me the app is really a proof of concept for a technology that makes it more efficient for AI models to analyze video. Running huge amounts of video through powerful AI models could get prohibitively expensive. Endless’ technique involves running the video through its own proprietary algorithm first to reduce the data being fed into the LLM. They hope to sell the tech to AI companies, enabling widespread use of video.

It’s the latest example of how companies are optimizing practically every aspect of big foundation models to make them more practical to employ.

How the UAE got the US to bless its AI ambitions

THE SCOOP

ABU DHABI — The US has approved the sale of cutting-edge Nvidia chips to G42, the company at the center of the United Arab Emirates’ efforts to become a global leader in artificial intelligence, according to a person with knowledge of the matter.

The approval, which came earlier this year but has not been reported, is a major milestone for the ambitions of the UAE and the broader Gulf region. The area is becoming an increasingly important global technology hub — but Biden administration concerns about their ties to China have been obstacles to gaining access to key assets, like cutting-edge chips, which face US export curbs to the region.

G42 is only just beginning to deploy the chips, including a sizable order of Nvidia H100 models, but the effort has been years in the making. Long before ChatGPT captivated the world and global demand for Nvidia’s most advanced AI chips exploded, the UAE was laying the groundwork to be a serious competitor in the global AI race.

It has made significant, long-term investments in innovations that protect sensitive data, making it exceedingly difficult for anyone but the owner to access. Semafor has learned new details of those efforts, which helped lead US officials to approve the sale of the AI chips.

New data centers in the country are physically locked down and have been built from the ground up with hardware made exclusively by Western nations to avoid the possibility of a Chinese backdoor, people familiar with the matter said.

Older data centers have been stripped of any parts with ties to China, even if those parts are unplugged and not being used.

The compute portion of its data centers, which use military-level FIPS 140-2 encryption, is physically separated from all other systems, from security cameras to cooling. That separation reduces the likelihood that attackers could use vulnerabilities in ancillary hardware to worm their way into the servers.

Customers, workers, and visitors are all screened extensively, and Chinese nationals are not allowed to work at the data centers.

The facilities employ wide-ranging telemetry of everything that happens inside, giving G42 the ability to monitor anomalies in real time, including fluctuations in compute levels that might suggest a customer is building AI models large enough to create national security concerns.

It has also hired US Department of Defense contractors to “red team” its data center defenses to find any holes that might allow intrusions by the Chinese or others.

And since last year, G42 has been training its advanced AI models in data centers in California and Texas through a commercial partnership with AI chip maker Cerebras, as it builds its infrastructure in the UAE.

In a sign of the region’s growing tech importance, Microsoft in April made a $1.5 billion investment in G42, a move that drew some criticism in Washington because of the UAE firm’s past investments in Chinese AI ventures and whether that could lead to intellectual property and advanced hardware flowing to America’s most powerful rival. Microsoft President Brad Smith joined G42’s board of directors.

G42 and Nvidia declined to comment.

Semafor reported earlier this week that a likely approval for Nvidia chips is drawing nearer for Saudi Arabia.

Before that outage. Semafor’s Rachyl Jones scooped that software engineers at the cybersecurity firm CrowdStrike complained about rushed deadlines, excessive workloads, and increasing technical problems to higher-ups for more than a year before a catastrophic outage in July. “Speed was the most important thing,” said Jeff Gardner, a senior user experience designer at CrowdStrike who said he was laid off in January 2023 after two years at the company. “Quality control was not really part of our process or our conversation.”

The issues were raised during meetings, in emails, and in exit interviews, nearly two dozen ex-employees told Semafor. CrowdStrike disputed much of Semafor’s reporting and said the information came from “disgruntled former employees, some of whom were terminated for clear violations of company policy.” The company told Semafor: “CrowdStrike is committed to ensuring the resiliency of our products through rigorous testing and quality control, and categorically rejects any claim to the contrary.”

Help wanted. Frustrated and feeling hamstrung in their job searches on LinkedIn, software developers in Spain and Italy say they’ve created AI-powered bots that scan your resume, search for relevant openings, and automatically submit thousands of applications — all while bypassing the site’s anti-bot measures.

It’s part of a cat-and-mouse game in hiring that separates the have-nots from the have-bots: Job seekers flood employers with AI-generated applications while companies use LinkedIn’s AI to recruit and vet prospective candidates. Madrid-based developer Jorge Frias told Semafor he made sure his bot didn’t trigger LinkedIn’s detection software by applying for too many jobs too quickly.

“You just need to make the system behave like a human — a fast, 90 percentile human with an impossibly large bladder,” he said. Another software developer used AI to apply to 1,000 jobs over four days.

— Mizy Clifton

There’s an interesting debate about the inference costs of OpenAI’s new model, o1, which takes its time to answer questions by doing a massive amount of “thinking” in the background, racking up a bunch of compute costs in the process.

The release of o1 coincides with a bunch of rumors that OpenAI is reportedly looking to raise something like $11.5 billion at a $150 billion post-money valuation.

So people are raising a lot of legitimate questions about whether that price tag, which is staggering, makes sense when the company is losing money and its newest model is sucking up even more juice.

There are certainly arguments that compute costs will continue to be a problem in the long term. And firms getting into OpenAI at this late-stage valuation may be “dumb money.”

If I put myself in the shoes of an investor, I think I’d ignore compute costs. The only question that matters is whether OpenAI will ultimately achieve “AGI,” or whatever you want to call it. Some kind of insanely powerful, world-changing AI.

If you don’t believe OpenAI has at least a good shot of doing that, why would you invest in the company? If you do, why are you worried about details like compute power? Your thesis is that everything is about to change. Not only will a model like that be worth pretty much any cost of compute, it will be able to improve itself, finding ways to be more efficient.

After a quick review of which moderator lessons and memes will stick from this week’s debate, Ben and Nayeema turn to New York Times reporter and former restaurant critic Pete Wells to answer the question on every foodie’s mind: What is a Critic in an Influencer’s World? Together, they explore the battle between the new wave of influencers and the old guard of gatekeepers in food. On the menu: the health hazards of criticism, how celebrity chefs and TikTokers have altered the power of his position, and whether diversifying food reviews in the midst of the culture wars is “DEI,” “virtue signaling,” or simply — as Pete reveals — an organic evolution.

Catch up with the latest episode of Mixed Signals.

Relatively brief conversations with AI chatbots trained on “evidence-based counterarguments” reduced conspiracy beliefs by an average of 20% for at least two months, according to research published Thursday in the journal Science.

The big takeaway? Conspiracy theorists may not be quite as impervious to facts as previously thought. “There’s a lot of ink spilled on the post-truth world, but evidence does matter,” said study co-author Gordon Pennycook, adding that the chatbot’s ability to tailor its response to specific beliefs could explain why other studies haven’t yielded similar results.

— Mizy Clifton

• Actually, the center could hold.
• Trump and Harris have the same oil problem.
• The next battle for Biden’s trustbuster.