Technology

➚ MOVE FAST: Short break. Elon Musk’s X is back in Brazil after it complied with a judge’s order to take down accounts and appoint a legal representative in the country, marking an end to the free speech standoff.

➘ BREAK THINGS: Short sale. Short seller Hindenburg Research released a damning report on video game company Roblox, accusing it of lying to investors about user numbers and prioritizing growth over child safety. Roblox called the allegations misleading.

One of the many challenges of building new AI data centers is heat. The AI chips, or graphics processing units, are so powerful that they’re in constant danger of melting.

A new partnership between the chemical maker Dow Inc. and Atlanta startup Carbice is aimed squarely at that problem.

Carbice, founded by Baratunde Cola, a mechanical engineering PhD and former Vanderbilt fullback, makes an alternative to “thermal paste,” the glue-like material that connects the blazing hot processor to the cooling mechanism that draws heat away from the chip.

It sounds like no big deal, but thermal paste is a crucial part of a GPU, filling microscopic air gaps that might otherwise trap heat instead of pulling it away. Any reduction in cooling performance can slow GPUs down and reduce their lifespan. It’s so important that serious computer gamers sometimes replace old thermal paste in their GPUs to get performance gains.

Instead of a paste, Carbice’s “ice pads” accomplish the task using carbon nanotubes. These are microscopic strands of carbon that are among the world’s best known conductors of heat.

The ice pad is made up of billions of carbon nanotubes that are vertically oriented, like a forest. They are so small that the individual tubes work their way into microscopic crevices on a material, creating a strong bond. Heat travels up the tubes and into the cooling mechanism, where it’s pulled away from the chip.

And unlike thermal paste, which slowly loses its effectiveness, Carbice says the nanotubes improve over time by burrowing deeper into the materials. The company says the device offers a dramatic improvement in cooling. But even a modest increase could save hundreds of millions of dollars for AI data centers with tens of thousands of GPUs.

Carbon nanotubes have been known for their incredible properties for decades and have fascinated both scientists and science fiction writers, who have dreamed up things like space elevators and, in The Three Body Problem, a grizzly tool of war.

While the material has found its way into products like batteries, it’s usually in powder form and not vertically aligned to take maximum advantage of its properties. But it’s incredibly difficult to manufacture complete carbon nanotubes at scale.

Carbice’s patented method, and Dow’s willingness to help mass produce it, is a positive development in the field, says Ray Baughman, a longtime nanotechnology researcher at University of Texas at Dallas with more than 100 patents. “His upscaling means that we have an opportunity to increasingly take advantage of our patent filings,” he said.

Vinod Khosla on OpenAI, national security and AI’s climate impact

THE SCENE

Vinod Khosla, the entrepreneur-turned-investor, has had an outsized impact on the development of modern day Silicon Valley. He pioneered the idea of open-source software as co-founder of Sun Microsystems and then, as a venture capitalist, helped fund the next generation of companies that pushed the boundaries of robotics, artificial intelligence, and clean tech.

In a recent paper, Khosla explains why he believes the next wave of technology — driven by generative AI — may have the biggest impact on humanity yet. It’s no secret that Khosla, an early investor in OpenAI, is a huge proponent of AI, but the Sept. 20 paper is a detailed treatise on why the naysayers (the Screen Actors Guild, politicians, etc.) and global challenges (conflict and supply-chain breakdowns) are standing in the way of a potential utopia.

In a wide-ranging interview, Khosla addressed some of the big, overarching questions facing AI today, including the monumental challenge of generating enough clean energy, the opportunities and challenges presented by the Gulf states, and the drama at OpenAI.

Q&A

Reed Albergotti: These AI models are getting so big; people are talking about five gigawatt data centers. It seems beyond the scale of the private sector at some point. How do you see the public sector coming in here?

Vinod Khosla: It’s very hard to decide today.

You have to plan in advance, because data centers take five years to build, and power plants can take even longer.

There is a lot of risk being taken, which is best done by capitalists, not by government[s]. They’re just too slow to do this. Also, governments can’t hire the quality of people needed. There’s engineers, three years out of PhD, making $1 million dollars or more. I’ve seen some ridiculous numbers, even higher. They can’t hire these people.

So when people say governments should regulate, they don’t have the people who can keep up. They aren’t as dynamic, and you can’t change policy every six months or 12 months. But our plans change every six months. It’s a very dynamic situation, and we just have to go with the flow and place our bets. Government can’t do that. It can’t hire the people. It can’t pay ridiculous salaries.

The amount that FTX may soon have to distribute to creditors. Thanks to shrewd crypto and AI investments, a bankruptcy court may soon pay some preferred shareholders in the failed exchange.

Artificial artificial intelligence data. The latest AI model release from Writer AI, a startup with customers like Uber and L’Oréal, used a particularly interesting set of data to train it. It was all “synthetic,” or created for that specific purpose by another AI model.

Palmyra X004, which Writer announced this morning, scored in the top ten on Stanford’s Holistic Evaluation of Language Models benchmark, the company said. It’s a sign that this novel method of training language models is making it out of the lab and into the real world.

Much has been written about the downsides of synthetic data and the potential lack of “real” data with which to train AI models. I’ve covered this topic a lot over the last couple of years, following advances in techniques using synthetic data, which has several benefits, like much leaner, focused models that are cheaper to run and less prone to hallucination (when AI models create false information).

Another big one is copyright. As I’ve said before, the intellectual property issue that has plagued the AI industry in the courts and in the public eye, doesn’t matter as much in an era where synthetic data can fill in a lot of gaps.

Writer co-founder and CTO Waseem Alshikh told me earlier this week that a lot of the synthetic data for Palmyra X004 was generated from a foundation of real data. For instance, Writer might pull data from a Wikipedia page and use another AI model to clean it up, check it for toxic content, give it topic headings and even convert it to a question-and-answer format. The result is data based in reality, but structured so large language models learn more efficiently. “The synthetic data, in this case, is just better data,” Alshikh said.

• Brazilian local elections provide a far-right foothold.
• Kamala Harris struggles to balance ‘change’ and loyalty to Joe Biden.
• The New York Times is testing a new puzzle called Zorse.