The head of Amazon’s push to build artificial general intelligence has shared rare insights on how junior staffers can get ahead in AI, at a time when the global talent war has turned into one of the most expensive battles in the tech industry.

Companies like Meta, OpenAI, and Anthropic have spent the summer aggressively jostling for top AI minds, in some cases dangling compensation packages worth tens of millions of dollars. Analysts say this scramble is one of the most expensive hiring sprees in the history of technology, with Meta in particular reportedly offering star researchers deals so lucrative that they rival the pay of athletes. Mark Zuckerberg has framed this as a “make or break” race for Meta’s future, as the company funnels billions into AI research.

David Luan, the executive leading Amazon’s AGI Lab in San Francisco, told The Verge’s Decoder podcast that he would put fewer than 1,000 people worldwide into the “top AI talent bracket” and trust fewer than 150 with what he described as a “giant dollar amount of compute” at a frontier AI lab.

But he stressed that while the elite circle remains small, it is possible for newcomers to break in quickly if they focus on the right problems. Luan said some junior people could climb the ranks within three to four years by asking the right questions, identifying problems “nobody has the answer to,” and becoming recognized experts in those narrow but critical subdomains.

“I find that really counterintuitive, that there’s only very few people who really know what they’re doing,” Luan said, adding: “It’s very easy in terms of number of years to become someone who knows what they’re doing.”

He noted that junior recruits coming from other demanding disciplines, such as quantitative finance or physics, could make “a massive difference” when they join AI companies—provided colleagues with deep experience in training models surround them.

Luan also advised early-career researchers to prioritize joining smaller AI teams where they could try their own ideas, while ensuring the company has a strong “product sense” of how AI fits into people’s lives.

In addition to heading Amazon’s AGI Lab, Luan is also the company’s vice president of autonomy. He joined Amazon in 2024 after it quasi-acquired his startup, Adept, which had been working on AI models designed to help people complete tasks across software platforms. His definition of artificial general intelligence is broad: “a model that can help a human do anything they want to do on a computer.”

The battle for talent, however, underscores the financial strain that AI ambitions are putting on tech giants. For Meta, the cost of attracting researchers has ballooned alongside the billions it is already burning on compute power. Industry insiders suggest that this level of spending is unsustainable without a clear path to profitable products. OpenAI and Anthropic face similar challenges, while Amazon, which entered the AGI race later, is now trying to leverage its financial muscle and cloud infrastructure to catch up.

Some analysts believe the takeaway from Luan’s remarks is twofold: the AI industry’s success is being bottlenecked by human capital as much as by compute, and the sheer cost of recruiting and retaining these rare talents could reshape how quickly companies like Amazon, Meta, and OpenAI can bring AI breakthroughs to the market.

Tesla is under federal investigation for failing to report crashes involving its partially autonomous driving technology in a timely manner, intensifying scrutiny on the automaker’s self-driving ambitions.

The National Highway Traffic Safety Administration (NHTSA) said automakers are required to report crashes involving advanced driver-assist features “within one or five days” of an incident. But Tesla was filing reports “several months or more” after crashes occurred. The company told regulators that the delays stemmed from a data collection issue that has since been corrected. Still, NHTSA’s Office of Defects Investigation has opened an audit to verify that Tesla is now complying fully with the rules.

The requirement traces back to a standing general order (SGO) issued in 2021. The mandate compels automakers and robotaxi companies to disclose crashes involving both fully autonomous vehicles and Level 2 driver-assist systems. Under the order, any collision must be reported if an automated driving system was engaged within 30 seconds of impact.

Since the rule came into force, Tesla has disclosed over 2,300 crashes to the federal government. An analysis of the data reveals Tesla accounted for 40 out of 43 fatal crashes reported under the SGO, underscoring its outsized role in accidents tied to semi-autonomous systems.

Tesla’s Autopilot and Full Self-Driving (FSD) features are categorized as Level 2 technology, which requires a driver to remain attentive and ready to take control at all times. By contrast, Alphabet’s Waymo operates Level 4 technology in its robotaxi fleet — vehicles capable of navigating without human intervention.

This contrast has fueled criticism of Tesla’s approach. While Elon Musk has dismissed lidar technology — the backbone of Waymo’s autonomous navigation — as unnecessary and inferior to Tesla’s camera-based system, accident records tell a different story. Waymo’s robotaxis have logged millions of autonomous miles in U.S. cities with a significantly lower crash rate than Tesla’s Level 2 systems, highlighting a growing divide in safety performance.

Despite the controversies, Tesla has been pushing forward with its robotaxi trials in select U.S. states, including California, Nevada, and Texas. Earlier this year, Musk touted Tesla’s “AI-powered robotaxi fleet” as a disruptive leap for urban mobility. However, industry veterans remain unconvinced.

John Krafcik, the former CEO who guided Waymo’s evolution from a Google research lab into a commercial ride-hailing company, cast doubt on Tesla’s much-hyped rollout. Speaking to Business Insider, Krafcik — now a board member at Tesla rival Rivian — dismissed Tesla’s pilot service in the Bay Area as little more than “a rebranded Uber with employees inside the car.”

“Please let me know when Tesla launches a robotaxi — I’m still waiting,” he said. “It’s (rather obviously) not a robotaxi if there’s an employee inside the car.”

Krafcik’s skepticism mirrors concerns among auto analysts and regulators who argue that Tesla’s push for Full Self-Driving is outpacing its technology. Unlike Waymo, which has secured approval to operate fully driverless rides in Phoenix, San Francisco, and Los Angeles, Tesla’s so-called robotaxi still requires safety drivers and faces mounting questions over its readiness for real-world deployment.

NHTSA, meanwhile, has floated changes to the 2021 reporting rules, proposing to scale back some requirements. But for Tesla, the audit signals a broader concern: whether the company is being transparent about the risks tied to its autonomous driving systems as it races ahead in an industry where safety records are becoming as important as innovation.