Two years ago, every argument about AI was about quality. The machines hallucinated court cases, absorbed biases from their training data, misspelled words in their own images, drew hands with six fingers, and couldn’t count the letters in strawberry. The confident claims that biases and hallucination errors were inherent to machines already sound quaint. Those […]

The post From slop to non-slop: The new big AI theme is quantity appeared first on Asia Times.

• A version of this story originally appeared in the AI Insider newsletter.
• Sign up for Business Insider's weekly AI Insider newsletter here.

The AI data center boom looks like it's running into serious trouble. Projects are being delayed, canceled, or abandoned. Electricity shortages are holding everything up.

But there's a twist: A lot of these data centers were probably never going to get built in the first place.

Bernstein Research analyzed thousands of planned US data centers and concluded that only about 35% of the announced capacity is likely to become reality.

The industry has a wonderful name for some of the more dubious projects: "Bragawatts."

These are enormous data centers announced by companies that often have little experience building them. Some haven't secured electricity, construction permits, or even customers.

Then there are the "ghost projects." Developers submit multiple requests for electricity connections, hoping at least one location works out. In Texas, these requests have reached more than five times the state's existing electricity supply.

No wonder the grid looks overwhelmed.

This doesn't mean the AI boom is fake. Far from it.

My colleague Eugene Kim exclusively reported this week that even Amazon, one of the world's most experienced data center builders, is struggling to get enough electricity to power its expensive AI equipment.

But there's a big difference between Amazon's very real problems and an inexperienced developer announcing a gigantic project that may never happen.

The AI data center boom is real, and so are the challenges of building all these facilities. But the sheer number of speculative projects may be making those challenges look bigger than they truly are.

Sign up for the AI Insider newsletter here. Reach out to me via email at abarr@businessinsider.com

• Pilots have faced the risk of deskilling for decades because of their reliance on autopilot.
• Steven Shaw of King's College London said they can show workers how to stay sharp with AI.
• "Even experts are at risk of deskilling," he said.

Pilots have relied on autopilot for decades. Their experience could offer a lesson for workers increasingly relying on AI.

Steven Shaw, an assistant professor at King's College London and coauthor of a widely cited study on "cognitive surrender," drew a parallel between pilots' reliance on automation and workers outsourcing their thinking and delegating their work to AI.

"We've had autopilot in planes for a very, very long time," Shaw said, speaking at a recent OxGen Summit panel. "So they learn to fly manually first, and then they learn to use autopilot strategically."

Some federal aviation guidance encourages pilots to occasionally switch off autopilot to maintain their manual flying skills.

There is a long track record of this. In a 2013 safety alert, the Federal Aviation Administration warned that "continuous use of autoflight systems could lead to degradation of the pilot's ability to quickly recover the aircraft from an undesired state."

The agency reiterated the importance of manual flying in a 2017 alert, recommending that airlines encourage pilots to practice without autopilot periodically when conditions permit.

"Even experts are at risk of deskilling," Shaw said, referring to the dangers of repeatedly outsourcing cognitive tasks to AI.

Shaw's warning comes as some tech leaders, AI researchers, and politicians have cautioned that overreliance on AI could weaken workers' underlying cognitive abilities.

Speaking at More Perfect University's "Students vs. Billionaires" town hall in Maine on Thursday, Rep. Alexandria Ocasio-Cortez warned that "AI's goal is to take your thoughts away from you, your ability to think away from you so that you become dependent on a machine."

OpenAI CEO Sam Altman said in July that one of AI's most overlooked risks was "cognitive atrophy," while Arthur Mensch, CEO of French AI startup Mistral, said last year that the technology's biggest risk could be "deskilling."

Think before you prompt

Early research is showing how readily people outsource their thinking to AI.

Vivienne Ming, an AI researcher and theoretical neuroscientist, told Business Insider in March that one of her experiments, involving 72 participants, found that roughly 90% to 95% either relied on AI to generate answers or used it to validate their assumptions, rather than challenge and deepen their thinking.

"The overwhelming trend is substitution," she said.

Similarly, Shaw's research suggests that the danger is that people increasingly accept AI's answers without thinking critically.

In experiments involving 1,372 participants, people accepted AI recommendations nearly 80% of the time even when the answers were deliberately incorrect.

Shaw calls this phenomenon "cognitive surrender."

His advice for avoiding it is straightforward: develop your own judgment before asking AI for help.

"Just think first and then go to the prompt," he said during the OxGen Summit panel.

Starting with AI, he warned, can narrow the range of ideas people consider independently.

Other experts have offered similar advice. Geetha Rajan, a former AI adoption leader at PwC, previously told Business Insider that she writes her own first drafts before using AI to check facts and challenge her work.

Ryan Deiss, CEO of DigitalMarketer, advocates a "10-80-10" rule: humans handle the first 10% of a task, AI takes on the middle 80%, and humans return for the final 10% to exercise judgment and refine the result.

• Everyone complains about AI-generated junk.
• But decades of sloppy human-written software may be a much bigger problem.
• This week's "Black October" freakout was caused by human slop, not AI.

The internet has a slop problem. And for once, we can't blame AI.

Most of us spend hours online every day without thinking much about the software holding everything together. We assume somebody, somewhere, has made sure it all works properly.

So it probably came as a shock this week to discover just how vulnerable some of that software is.

The cybersecurity freakout dubbed "Black October" has inspired warnings about AI-powered hackers, digital catastrophe, and even advice to stockpile cash and canned food.

At first glance, AI looks like the villain. But this is largely a human-made mess.

Here's the problem.

The internet runs on decades of software, some of it containing security weaknesses that have been overlooked, tolerated, or simply left unfixed.

For years, that was a manageable gamble. Fixing software takes time and money. Exploiting obscure weaknesses often required considerable expertise. Companies could convince themselves that the risk wasn't worth worrying about.

AI is changing that calculation. Someone who once needed years of training to hunt for weaknesses can now get help from an AI model. These tools can also examine software at speeds that humans can't match.

Suddenly, all those problems we've been ignoring look more urgent.

"Over the last 10 years the risk acceptance process has morphed from a way to accept risks to enable the business, to a way to justify never fixing security issues," Joel de la Garza and Zane Lackey, partners at a16z, wrote on Friday.

That's quite an indictment. The result is a recent surge in published vulnerabilities.

This looks like AI is creating new security problems. But it's actually AI exposing weaknesses that humans left behind.

Slop irony

The irony is delicious. We've spent the past couple of years complaining about "AI slop." Turns out human slop can be pretty consequential, too.

"The funny thing is everyone was saying AI slop code will cause a cyber crisis but it's actually all the old human slop causing it and the AI is just cleaning it up," software engineer Ben Kaufman wrote on X on Thursday.

"Humans have been the ones writing and reviewing each other's code for decades," tech founder Anant Jain wrote on X. "If we're being intellectually honest and holding ourselves to the same bar we hold AI to, we'd have to admit that we haven't been all that good at it either."

Humans are sloppy

Anyone who's worked with other humans for a decade or two knows exactly how this happens.

It's Friday afternoon. You've almost finished a project. You could spend another few hours making it perfect, or declare it 90% done and head off to that dinner party. We've all been there.

AI agents don't have dinner parties. They don't get tired on Friday afternoons. They can keep searching, testing, and trying again. That persistence is now colliding with an internet built by busy, imperfect humans who often settled for "good enough for now."

A tragedy of the commons

There's another wrinkle here, though, and it's an important one.

Much of the software underpinning the internet is freely available for anyone to use. Some of it is maintained by volunteers or small groups of experts, even as enormous technology companies depend on it.

In 2020, the cartoonist Randall Munroe captured this absurdity in an xkcd comic. Modern digital infrastructure still rests precariously on tiny foundations.

It's funny because it's uncomfortably close to reality. Take the specific PHP problems from this week. These flaws have existed for years, and the PHP Foundation acknowledges that limited resources and disagreements over fixes contributed to delays in fixing them.

Tom Occhino, chief product officer at Vercel, brought up Munroe's cartoon during my recent interview with him.

He explained that AI is finding weaknesses even in software that the industry has treated as dependable for decades.

This is partly a version of the tragedy of the commons: Everyone benefits from a shared resource, but no one has sufficient incentive to take responsibility for maintaining it.

And let's be clear: The volunteers keeping essential software running aren't the villains here. The companies benefiting from their work have a much greater responsibility.

Occhino thinks AI could help solve the problem it is exposing.

He described an idea being discussed at Vercel: Create a foundation or coalition that funds efforts to identify vulnerable software and replace it with safer alternatives.

"We need a consortium. We need a coalition," he told me.

He suggested companies could contribute computing resources and AI tools to the effort. Nothing has been formally announced, but the idea is compelling.

Indeed, OpenAI's Patch the Planet effort is an example of what's possible.

Rather than simply using AI to create more software, the industry could use it to repair what already exists.

Occhino put it nicely: Companies need to spend less time chasing new ideas and more time asking how to make existing systems "just work really, really well."

That's a better response than panicking about the imminent collapse of the internet.

AI didn't create decades of neglected software. We did. Perhaps it's time to clean up our own slop.

Sign up for the AI Insider newsletter here. Reach out to me via email at abarr@businessinsider.com

• Canva launched Canva Craft, a tool similar to AI image models by Google, Meta, and OpenAI.
• Surprisingly, Canva fell short on a design task but did well at generating more realistic images.
• Here is how Canva's new image generation tool stacks up against ChatGPT, Gemini, and Muse.

Canva wants a shot at winning the AI image generation race.

The design platform launched Canva Craft on Thursday, its first image-generation model, similar to those from Google, Meta, and OpenAI.

Users can access Craft through Canva’s Dream Lab, including with free accounts, though it is subject to AI usage limits. You’ll need to sign in to a Canva account to get started.

In the release announcement, Canva said the tool plans a design’s layout, proportions, and text before generating an image, an approach that aims to help it produce legible lettering and accurate details.

The company also said that the model was trained on human-designed Canva templates, including posters, menus, and presentations, rather than photographs.

To test how Canva performs, Business Insider created three image generation prompts with escalating difficulty levels — from designing a simple greeting card to creating a photorealistic image from scratch.

Here is how Canva's brand-new image generation tool stacked up against three of the most powerful models.

Canva did not respond to a request for comment.

Starting off with an illustration

We started off simple with a prompt for a birthday card.

Create a cartoon "Happy Birthday" greeting card for a friend named Jessica who has pink hair, brown eyes, likes to wear paisley, and is turning 28 years old in November.

Surprisingly, Canva, the design platform, fell short in its design of the birthday card, from the awkward font to the outsize head of the cartoon character. It didn't even bother to include any paisley.

By comparison, ChatGPT and Muse understood the assignment.

ChatGPT guessed that the recipient's favorite color might be pink, while Muse thoughtfully wove in cozy fall elements for the November birthday girl.

Meanwhile, Gemini took the request literally and produced an image of an actual card. Extra points!

Winner: Muse

Editing an image

For the second task, I tested the AI models' ability to identify an element and place it in a brand-new setting.

Edit this image of me holding a necklace pendant. Take the red garnet pendant off my hand and make it look as if it is displayed in a vitrine in a jewelry showroom.

Canva did a great job of identifying the pendant and creating an atmospheric setting, but, like Gemini, it depicted the jewelry in a way that defied physics and common sense. It would be unrealistic for a pendant to float in the air or sit propped on a metal pole.

ChatGPT and Muse once again outperformed other models in understanding how things typically look in reality. Muse, however, made some alterations to the pendant, which I didn't ask for.

Winner: ChatGPT

Creating a photorealistic image from scratch

With Halloween approaching, I asked for a realistic depiction of my favorite animal.

Generate a photorealistic image of a long-haired orange tabby cat in a Halloween costume with bat wings, perched on the edge of a gothic medieval castle under the full moon and a starless night sky.

Though the cat's fur looks a little too strandy and high-contrast in Canva's creation, no components are missing or malformed, and the image composition pretty much matches the quality of Muse and Gemini.

The clear winner for me in this round, however, would be ChatGPT.

Not only is the cat's posture majestic, but the soft glow on his left side is extremely realistic, and the white balance is well done. There are also clouds softly cradling the moon, which could be a response to my request for a starless sky.

Winner: ChatGPT

Overall, it's evident that Canva still has a lot of room to grow compared to other models, especially if graphic design is its focus. For now, I'll stick to moving components around by hand.