Technology

➚ MOVE FAST: Making it. OpenAI is set to release a new AI model that works differently from the ones that currently power ChatGPT. Code named Strawberry, it has better reasoning capabilities than its predecessor, according to The Information. Don’t get too excited. It takes around 20 seconds to answer a question and only does text. But it could be a seed of something sweet down the road.

➘ BREAK THINGS: Faking it. Taylor Swift’s decision to publicly back Kamala Harris crystallized after the Donald Trump campaign used a deepfake to falsely portray the pop singer’s support of him. Pushing the world’s biggest celebrity into the political fray is the biggest impact deepfakes have had on elections so far.

Training AI models on languages other than English has been key to expanding global accessibility to the technology. G42, the UAE-based tech conglomerate, will soon launch a large language model in Hindi it’s dubbing NANDA. G42’s platform will be trained on around 2.13 trillion “tokens” of language datasets, including Hindi, which is spoken by more than 50% of Indians.

It joins giants like Microsoft and Google who are “particularly interested in navigating India’s linguistic challenges,” Fortune Magazine wrote, and are customizing AI chatbots to operate in less-popular Indian languages. But some AI researchers have warned of “cutesy technology solutions” that don’t address realistic problems faced by rural Indians.

Last August, G42 launched JAIS, which it described as the world’s first open-source Arabic LLM.

Saudi Arabia may soon get long-awaited AI chips

THE SCOOP

RIYADH — The US government is considering allowing Nvidia to export advanced chips to Saudi Arabia, people familiar with the matter said, which would help the country train and run the most powerful AI models.

The fate of those sales was a major, unofficial topic at Saudi Arabia’s global AI summit, known as GAIN, on Thursday. Representatives from AI hardware firm Groq, Google, and Qualcomm, along with Saudi government officials, mingled and spoke at the event.

Earlier this year, the US curtailed shipments of Nvidia’s advanced graphics cards out of concern that Saudi Arabia’s close ties with China could allow valuable AI secrets to flow there or possibly open the door to the mainland gaining access to those chips. The Biden administration has forbidden companies from selling the cutting edge equipment to China because of national security worries.

Conference attendees, including some who work for the Saudi Data and AI Authority, said the country is working to satisfy US security demands in an effort to get the chips as soon as possible.

The gathering in Riyadh also marked the Gulf power’s decisive shift toward the US camp in the bipolar global AI tug of war. One participant who has been to the event before said this year’s gathering had a noticeably smaller Chinese presence. On the convention center floor, there were only a handful of Chinese attendees. Exhibitors included Huawei and Alibaba, both Chinese firms, but few others.

According to people with knowledge of Saudi policies, the government has taken steps to limit its involvement with Chinese firms, while keeping the door open to China should the United States cut the kingdom off from the most advanced US chips.

One of the people said the Saudi government is expecting shipments of Nvidia H200s, the company’s most advanced chips currently available. A Nvidia spokesman declined to comment.

A Commerce Department spokesperson told Semafor that the agency “cannot speak to specific licenses or transactions” but stressed that policy actions “are the subject of a rigorous interagency process including the Departments of Commerce, State, Defense, and Energy.”

Saudi Arabia is in a race to develop large scale AI models, both against the fast pace of the industry and its regional rival, the United Arab Emirates, which beat Saudi Arabia to the punch.

But the lack of chips means technology companies in Saudi Arabia are being held back.

The starting price of Huawei’s new Mate XT, which folds twice, like a map, making the phone seem almost like a laptop screen. Despite the price tag, it had more than 3 million pre-orders, Reuters reported. The announcement comes after Apple’s annual ritual of revealing the new iPhone, which is pretty much the same as the old iPhone.

Rep. Dina Titus (D-Nevada), Mark Ein, Venture Capitalist and Entrepreneur; Limited Partner, Washington Commanders and Lori Kalani, Chief Responsible Gaming Officer, DraftKings will join Semafor’s editors in Washington, D.C. on September 19 for a discussion on the growth and trajectory of the U.S. gaming industry.

RSVP for in-person or livestream.

Was Elon right about LiDAR, the sensors that use lasers to measure distances in great detail?

It was once conventional wisdom that the technology was crucial for autonomous vehicles to operate. But on Monday, Intel’s Mobileye, one of the leading companies in providing driver assisted technology to automakers, said it was shuttering its LiDAR division.

People who have been following the trajectory of AV systems will remember when, five years ago, Musk called LiDAR a fool’s errand. “You’ll see,” he said at the time.

Musk’s fans used Mobileye’s news Monday to point out that the Tesla CEO was right all along.

Musk believes cameras are good enough for fully autonomous vehicles. After all, humans only have two and they drive just fine. Mobileye didn’t quite say there was no place for LiDAR, though. It said its AV technology didn’t need it as much and that it could buy LiDAR from third-party providers more cheaply. The LiDAR debate is important because it gives us a sense of how the technology and thinking around AVs are changing.

We’re in an AV “trough of disillusionment” right now. After years of hype, a lot of people now believe the technology is impossible. Sure, you can see cars without drivers zipping all over San Francisco, and it’s impressive and inspiring, but those cars can’t drive anywhere they want. One former AV engineer described them to me like cars driving on invisible, digital rails.

We are a long way from autonomous vehicles that can go on any road without a driver. They probably need AI models that are capable of something resembling human-level reasoning.

Meta isn’t happy about the US government’s draft guidelines on AI foundation models. The National Institute of Standards and Technology (NIST) issued a public request for comment in July on a proposal designed to keep them from being used to harm people or threaten national security. It’s part of the Biden administration’s efforts to establish some AI guardrails.

Meta says NIST’s guidance conflicts with the government’s findings that open-source models, which can be adapted by other programmers for specific uses, promote innovation and are an important part of the AI ecosystem. “Contrary to previous NIST guidance that emphasized the importance of risk management throughout the broader AI value chain, [the NIST draft] only focuses on the responsibility of foundation model developers, and ignores the responsibilities of downstream deployers to manage misuse risk that should also be considered – even at the model development stage,” Meta said in its comments to the agency.

It’s part of a broader debate on whether open-source models like the one Mark Zuckerberg’s firm has bet on are potentially dangerous, because they allow people to use them without guardrails. In contrast, nobody knows how closed models like OpenAI’s GPT-4 work, exactly, and OpenAI can control how people use it. The counter argument is that open source is safer, because it allows outside AI researchers to analyze the models to understand how they’re evolving.

• Donald Trump’s unconventional ground game is making Republicans nervous.
• Can the US defeat Google in court?
• Chinese roots deepen in Africa’s last Taiwan holdout.