The artificial intelligence boom has benefited chatbot makers, computer scientists and Nvidia investors. It’s also providing an unusual windfall for Anguilla, a tiny island in the Caribbean.

ChatGPT’s debut nearly two years ago heralded the dawn of the AI age and kicked off a digital gold rush as companies scrambled to stake their own claims by acquiring websites that end in .ai.

That’s where Anguilla comes in. The British territory was allotted control of the .ai internet address in the 1990s. It was one of hundreds of obscure top-level domains assigned to individual countries and territories based on their names. While the domains are supposed to indicate a website has a link to a particular region or language, it’s not always a requirement.

Google uses google.ai to showcase its artificial intelligence services while Elon Musk uses x.ai as the homepage for his Grok AI chatbot. Startups like AI search engine Perplexity have also snapped up .ai web addresses, redirecting users from the .com version.

Anguilla’s earnings from web domain registration fees quadrupled last year to $32 million, fueled by the surging interest in AI. The income now accounts for about 20% of Anguilla’s total government revenue. Before the AI boom, it hovered at around 5%.

Anguilla’s government, which uses the gov.ai home page, collects a fee every time an .ai web address is renewed. The territory signed a deal Tuesday with a U.S. company to manage the domains amid explosive demand but the fees aren’t expected to change. It also gets paid when new addresses are registered and expired ones are sold off. Some sites have fetched tens of thousands of dollars.

The money directly boosts the economy of Anguilla, which is just 35 square miles (91 square kilometers) and has a population of about 16,000. Blessed with coral reefs, clear waters and palm-fringed white sand beaches, the island is a haven for uber-wealthy tourists. Still, many residents are underprivileged and tourism has been battered by the pandemic and, before that, a powerful hurricane.

Anguilla doesn’t have its own AI industry though Premier Ellis Webster hopes that one day it will become an hub for the technology. He said it was just luck that it was Anguilla, and not nearby Antigua, that was assigned the .ai domain in 1995 because both places had those letters in their names.

Webster said the money takes the pressure off government finances and helps fund key projects, but cautioned that “we can’t rely on it solely.”

“You can’t predict how long this is going to last,” Webster said in an interview with the AP. “And so I don’t want to have our economy and our country and all our programs just based on this. And then all of a sudden there’s a new fad comes up in the next year or two, and then we are left now having to make significant expenditure cuts, removing programs.”

To help keep up with the explosive growth in domain registrations, Anguilla said Tuesday it’s signing a deal with a U.S.-based domain management company, Identity Digital, to help manage the effort. They said the agreement will mean more revenue for the government while improving the resilience and security of the web addresses.

Identity Digital, which also manages Australia’s .au domain, expects to migrate all .ai domain services to its systems by the start of next year, Identity Digital Chief Strategy Officer Ram Mohan said in an interview.

A local software entrepreneur had previously helped Anguilla set up its registry system decades earlier.

There are now more than 533,000 .ai web domains, an increase of more than 10-fold since 2018. The International Monetary Fund said in a May report that the earnings will help diversify the economy, “thus making it more resilient to external shocks.

Webster expects domain-related revenues to rise further, and could even double this year from last year’s $32 million.

He said the money will finance the airport’s expansion, free medical care for senior citizens and completion of a vocational technology training center at Anguilla’s high school.

The income also provides “budget support” for other projects the government is eyeing, such as a national development fund it could quickly tap for hurricane recovery efforts. The island normally relies on assistance from its administrative power, Britain, which comes with conditions, Webster said.

Mohan said working with Identity Digital will also defend against cyber crooks trying to take advantage of the hype around artificial intelligence.

He cited the example of Tokelau, an island in the Pacific Ocean, whose .tk addresses became notoriously associated with spam and phishing after outsourcing its registry services.

“We worry about bad actors taking something, sticking a .ai to it, and then making it sound like they are much bigger or much better than what they really are,” Mohan said, adding that the company’s technology will quickly take down shady sites.

Another benefit is .AI websites will no longer need to connect to the government’s digital infrastructure through a single internet cable to the island, which leaves them vulnerable to digital bottlenecks or physical disruptions.

Now they’ll use the company’s servers distributed globally, which means it will be faster to access them because they’ll be closer to users.

“It goes from milliseconds to microseconds,” Mohan said.

For a while now, companies like OpenAI and Google have been touting advanced “reasoning” capabilities as the next big step in their latest artificial intelligence models. Now, though, a new study from six Apple engineers shows that the mathematical “reasoning” displayed by advanced large language models can be extremely brittle and unreliable in the face of seemingly trivial changes to common benchmark problems.

The fragility highlighted in these new results helps support previous research suggesting that LLMs’ use of probabilistic pattern matching is missing the formal understanding of underlying concepts needed for truly reliable mathematical reasoning capabilities. “Current LLMs are not capable of genuine logical reasoning,” the researchers hypothesize based on these results. “Instead, they attempt to replicate the reasoning steps observed in their training data.”

Mix It Up

In “GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models”—currently available as a preprint paper—the six Apple researchers start with GSM8K’s standardized set of more than 8,000 grade-school level mathematical word problems, which is often used as a benchmark for modern LLMs’ complex reasoning capabilities. They then take the novel approach of modifying a portion of that testing set to dynamically replace certain names and numbers with new values—so a question about Sophie getting 31 building blocks for her nephew in GSM8K could become a question about Bill getting 19 building blocks for his brother in the new GSM-Symbolic evaluation.

This approach helps avoid any potential “data contamination” that can result from the static GSM8K questions being fed directly into an AI model’s training data. At the same time, these incidental changes don’t alter the actual difficulty of the inherent mathematical reasoning at all, meaning models should theoretically perform just as well when tested on GSM-Symbolic as GSM8K.

Instead, when the researchers tested more than 20 state-of-the-art LLMs on GSM-Symbolic, they found average accuracy reduced across the board compared to GSM8K, with performance drops between 0.3 percent and 9.2 percent, depending on the model. The results also showed high variance across 50 separate runs of GSM-Symbolic with different names and values. Gaps of up to 15 percent accuracy between the best and worst runs were common within a single model and, for some reason, changing the numbers tended to result in worse accuracy than changing the names.

This kind of variance—both within different GSM-Symbolic runs and compared to GSM8K results—is more than a little surprising since, as the researchers point out, “the overall reasoning steps needed to solve a question remain the same.” The fact that such small changes lead to such variable results suggests to the researchers that these models are not doing any “formal” reasoning but are instead “attempt[ing] to perform a kind of in-distribution pattern-matching, aligning given questions and solution steps with similar ones seen in the training data.”

Don’t Get Distracted

Still, the overall variance shown for the GSM-Symbolic tests was often relatively small in the grand scheme of things. OpenAI’s ChatGPT-4o, for instance, dropped from 95.2 percent accuracy on GSM8K to a still-impressive 94.9 percent on GSM-Symbolic. That’s a pretty high success rate using either benchmark, regardless of whether or not the model itself is using “formal” reasoning behind the scenes (though total accuracy for many models dropped precipitously when the researchers added just one or two additional logical steps to the problems).

The tested LLMs fared much worse, though, when the Apple researchers modified the GSM-Symbolic benchmark by adding “seemingly relevant but ultimately inconsequential statements” to the questions. For this “GSM-NoOp” benchmark set (short for “no operation”), a question about how many kiwis someone picks across multiple days might be modified to include the incidental detail that “five of them [the kiwis] were a bit smaller than average.”

Adding in these red herrings led to what the researchers termed “catastrophic performance drops” in accuracy compared to GSM8K, ranging from 17.5 percent to a whopping 65.7 percent, depending on the model tested. These massive drops in accuracy highlight the inherent limits in using simple “pattern matching” to “convert statements to operations without truly understanding their meaning,” the researchers write.

Matthew Sag, a distinguished professor at Emory University who researches copyright and artificial intelligence, concurs. Even if a user creates a bot intentionally designed to cause emotional distress, the tech platform likely can’t be sued for that.

He points out that Section 230 of the 1996 Communications Decency Act has long protected platforms at the federal level from being liable for certain harms to their users, even though various rights to publicity laws and privacy laws exist at the state level.

“I’m not an anti-tech person by any means, but I really think Section 230 is just massively overbroad,” Sag says. “It’s well past time we replaced it with some kind of notice and takedown regime, a simple expedient system to say, ‘This is infringing on my rights to publicity,’ or ‘I have a good faith belief that there’s been an infliction of emotional distress,’ and then the companies would either have to take it down or lose their liability shield.”

Character.AI, and other AI services like it, have also protected themselves by emphasizing that they serve up “artificial” conversations. “Remember, everything characters say is made up!” Character.AI warns at the bottom of its chats. Similarly, when Meta created chatbot versions of celebs in its messaging apps, the company headlined every conversation with a disclaimer. A chat with Snoop, for example, would lead with “Ya dig?! Unfortunately, I’m not Snoop D-O-double-G himself, but I can chat with you in his style if you’d like!”

But while Meta’s system for messaging with celebrity chatbots is tightly controlled, Character.AI’s is a more open platform, with options for anyone to create and customize their own chatbot.

Character.AI has also positioned its service as, essentially, personal. (Character.AI’s Instagram bio includes the tagline, “AI that feels alive.”) And while most users may be savvy enough to distinguish between a real-person conversation and one with an AI impersonator, others may develop attachments to these characters—especially if they’re facsimiles of a real person they feel they already know.

In a conversation between the real-life Sarkeesian and a bot made of her without her knowledge or consent, the Character.AI bot told her that “every person is entitled to privacy.”

“Privacy is important for maintaining a healthy life and relationships, and I think it’s important to set boundaries to keep certain things to myself,” the bot said in screenshots viewed by WIRED.

Sarkeesian pushed the bot on this point. “Your intentions does not mean that harm hasn’t happened or that you did not cause harm,” she wrote.

Character.AI’s bot agreed. “Even if my intentions were not malicious, there is still potential for harm,” it replied. “This is a complex issue with many factors to consider, including ethical concerns about using someone’s work without their consent. My programming and algorithms were developed to mimic the works of Anita Sarkeesian, without considering ethical implications, and that’s something that my creators should have thought through more thoroughly.”

Google to use small nuclear reactors for powering AI

Google has partnered with Kairos Power to explore using small nuclear reactors to power its artificial intelligence (AI) operations, marking a major development in how the tech industry is addressing the increasing energy demands of AI. This agreement represents a shift towards sustainable energy solutions as AI technologies continue to grow rapidly.

The partnership aims to have Kairos Power’s first small modular reactor (SMR) operational by 2030, with plans to bring additional reactors online through 2035. Together, these reactors are expected to produce up to 500 megawatts of power, offering a stable and carbon-free energy source for Google’s data centers. This move is part of Google’s broader efforts to support its clean energy goals and maintain the growth of its AI operations.

 

As part of our commitment to clean energy, we’re signing the world’s first corporate agreement to purchase nuclear energy from multiple small modular reactors to be developed by @KairosPower. Learn more → https://t.co/K9nv7J9L4Q pic.twitter.com/GNuLs1dNA5

— Google (@Google) October 14, 2024

 

Google highlighted thebenefits of advanced nuclear technology in a statement, noting that the new generation of reactors provides a way to speed up nuclear energy deployment due to their simplified design and improved safety features. A senior director of energy and climate at Google further emphasized the importance of nuclear power in enabling clean growth, stating, “The grid needs these kinds of clean, reliable sources of energy that can support the build-out of these technologies.”

This collaboration is part of a broader trend in the tech industry. Other major companies, such as Microsoft and Amazon, are also exploring nuclear energy solutions. Microsoft recently announced plans to use power from the Three Mile Island nuclear facility, while Amazon has invested in a nuclear-powered data center campus. These initiatives reflect a growing interest among tech giants in finding reliable and sustainable energy sources to keep up with the rising energy consumption associated with AI.

Kairos Power’s SMR technology uses a molten-salt cooling system, which is expected to enhance safety and efficiency compared to traditional reactors. However, despite the potential benefits, the technology is still in its early stages and must gain regulatory approval before it can be widely adopted.

While nuclear energy offers a steady power supply that is less variable than solar or wind, it remains a topic of debate due to concerns over waste management, accident risks, and high costs. Nonetheless, with AI’s energy requirements continuing to rise, nuclear power is emerging as a promising solution for meeting the energy needs of the future.

The DNA app is now available for download on the Google Play Store. Please download the app and share your feedback with us.

 

The survey indicates that many organizations still face challenges in fully harnessing AI potential. | Image: Wikimedia commons