Technology

➚ MOVE FAST: Bait. Nvidia CEO Jensen Huang said reasoning models need “100 times more compute” than models from just a few years ago. It was a bit of a defensive comment given Nvidia’s stock price is still recovering from DeepSeek’s R1 launch last month, which raised questions about the need for massive infrastructure buildout. On Wednesday, Nvidia reported quarterly sales are up 78% from the year prior.

➘ BREAK THINGS: Switch. The US Federal Aviation Administration is considering canceling a $2.4 billion contract with Verizon to update its air traffic control systems, awarding it instead to Elon Musk’s Starlink, The Washington Post reports. The move would give yet another government contract to the president’s “first buddy,” though Starlink’s satellite-based internet service does provide certain connectivity advantages.

Vibecoding is on the rise. With few coding skills, New York Times tech columnist Kevin Roose created a series of online tools to make his life easier, including a website that can determine if a piece of furniture will fit in his car and an app that suggests lunch options for his son, based on what’s in his fridge. And you can, too.

Using an AI-powered coding platform like Cursor or Windsurf, users can write text prompts on features and products they would like to see, and the service will write the code for them — significantly lowering the barrier to entry for creating new software. “Vibecoding,” a term popularized by researcher Andrej Karpathy, means giving into the vibes of the project, where the developer doesn’t necessarily have to touch the keyboard, but let the AI do the work. Roose described the experience as “mind-blowing.”

In our own experimentation with vibecoding, we found certain limits in the technology. When trying to create more complex programs or requiring more steps, token limits quickly became a problem. With linear increases in app complexity come exponential increases in compute. What’s clear is that the technical capabilities for something truly world-changing are there, as long as more tokens become available — that means more data centers, more energy, and more efficient models.

Autonomous driving has crossed a threshold

THE SCENE

A 12-hour journey in a Cadillac Escalade, equipped with General Motors’ hands-free driving technology, changes what it means to go on a long family road trip, and puts transportation in a long-promised and long-delayed new era.

It’s a surprising takeaway given GM ditched plans to develop a full self-driving taxi business under Cruise Automation, a company it acquired in 2016. In December, California regulators revoked its license to operate following an incident in which a pedestrian was dragged by a robotaxi. Cruise’s remaining employees now work on improving Super Cruise, the autonomous driving technology in the Escalade, and other “advanced driver assist system” efforts at the company.

Bay Area residents are already used to autonomous vehicles, with Google’s driverless Waymo taxis ferrying passengers everywhere in the city. But a road trip along I-80 was a different proposition. I decided to test out the Super Cruise feature, risking marital discord by using my wife and two kids as guinea pigs during our annual President’s Day ski trip last week.

We made the 766-mile journey in a GM-supplied Cadillac Escalade V-series, a $160,000 SUV with a supercharged V-8 engine with a delightful growl. (At every gas station, at least one man complimented me on my fine taste in cars). As soon as we got on the freeway, I hit a button that caused the top of the steering wheel to glow green, letting me know that the car was taking over.

I was still technically responsible for everything the car did and had to be ready to take over at any point — a key requirement that makes it legal in all 50 states. An eye-tracking camera located near the top of the windshield and infrared lights built into the steering wheel kept me from looking away from the road for more than five seconds.

After five seconds, a green light on the steering wheel would start blinking, putting me on notice that it was time to pay attention. If I didn’t snap my gaze back on the road ahead, I’d feel an abrupt vibration in my seat, giving me a haptic warning that I needed to take back control of the car.

There are a lot of similar systems on the road today, but only a small handful allow hands-free operation. Even Teslas require drivers to hold onto the wheel. That small difference is, surprisingly, a game changer on a long drive, leaving you able to relax in comfortable positions. By the end, my techno-skeptical wife was converted and now wants a new car that is as close to autonomous as we can get.

The amount TikTok will spend building data centers in Thailand over five years, more than double its original target. The infrastructure will support Thailand’s some 50 million users — 70% of the population — and help build out the company’s AI footprint in Southeast Asia.

Microsoft won’t let the chips fall. The company is petitioning the Trump administration to overhaul a last-minute Biden order that caps chip exports in roughly 150 countries. The so-called AI Diffusion Rule intends to limit the number of American-made chips finding their way into China, but it may also curb sales for US tech companies.

“The unintended consequence of this approach is to encourage [these] countries to look elsewhere for AI infrastructure and services,” Microsoft president Brad Smith wrote in a blog post Thursday. “And it’s obvious where they will be forced to turn.”

The rule limits chip purchases in a number of friendly countries, including Poland, Greece, Israel, Saudi Arabia, and the UAE — in which Microsoft has specifically been investing lately. Smith said it undermines Trump’s priorities to strengthen the AI market in the US and reduce the trade deficit. But Trump’s lack of action on the rule thus far — given his ties to the Gulf and his immediate widespread reversals of Biden administration policies — signals there may be some merit in the matter. The rule supports national security by restricting chip access to known smugglers, preventing adversaries from getting advanced technologies to aid military and cyber threats against the US.

Those national security provisions are important and should stay, including the requirement that tech components be deployed in trusted data centers, according to Smith. But ultimately, he said, it “goes beyond what’s needed.”

Microsoft is shutting down Skype, nearly 15 years after it acquired the videoconferencing service. The name Skype became synonymous with video calling, taking the technology from grainy experiments to mainstream use around the world. It felt kind of like a techno miracle at the time. Now, as it finally shuts down after years of irrelevance, it’s a good reminder of how fast stagnant technology can be replaced.

The videoconferencing field quickly became crowded as everyone from Google to Apple to Facebook began offering the service, often for free. Skype remained somewhat useful for niche cases, like if you needed a local phone number in another country.

One of the most remarkable twists was when Eric Yuan, who worked for Cisco’s Webex, launched Zoom — which blew everyone out of the water. Even before the pandemic, companies were switching to Zoom because of its superior technology.

Today, as leading AI companies capitalize on first-mover advantages and build brand names, the message is clear: If you slow down, you’ll get gobbled up.

Another week, another big announcement in quantum computing. On Thursday, it was Amazon’s turn. Their researchers published a paper in Nature detailing how they created a custom quantum chip, called Ocelot, which is significantly less error-prone than other methods used by competing companies.

A little background for those who are not deep into the quantum weeds: The calculations in quantum computers are done by qubits, subatomic particles that are brought down to temperatures colder than deep space, which sends them into a “quantum state,” where all kinds of weird things happen that defy classical physics. Thus, these qubits are difficult to control and prone to errors.

To build a quantum computer at scale, you’ll either need lots of qubits for error correction or find a way to make qubits that aren’t error-prone.

Amazon says it’s done the latter by using a special qubit that oscillates in a way that almost eliminates certain types of errors. These are called Cat qubits, after physicist Schrödinger’s cat, because they remain in two states at once. The great thing about this qubit is that it only requires a single traditional qubit to keep it in check.

While some people believe “at scale” quantum computers will require at least a million qubits, Amazon may have found a way to reduce that number by a lot. The problem isn’t solved — there are still many technical hurdles to build this kind of quantum computer. But those hurdles seem to be steadily falling away lately.

Trump’s tariff threats are spiking the nerves of American CEOs. Executives are nervous about Trump prioritizing trade restrictions without giving the same attention to business-friendly deregulation that he talked up in the campaign trail, Semafor’s Liz Hoffman and Rohan Goswami report.

Consumer confidence is plummeting, and the US stock market is lower than it was before his inauguration. “The chaos that is reigning right now is causing everyone to sit on their hands,” Nasdaq Private Market CEO Tom Callahan said. White House spokesman Harrison Fields insisted Trump is offering a better environment for businesses than the Biden administration.