There's a curious contradiction at the heart of today's most capable AI models that purport to "reason": They can solve routine math problems with impressive accuracy, yet when faced with formulating deeper mathematical proofs found in competition-level challenges, they often fail.

That's the finding of eye-opening preprint research into simulated reasoning (SR) models, initially listed in March and updated in April, that mostly fell under the news radar. The research serves as an instructive case study on the mathematical limitations of SR models, despite sometimes grandiose marketing claims from AI vendors.

What sets simulated reasoning models apart from traditional large language models (LLMs) is that they have been trained to output a step-by-step "thinking" process (often called "chain-of-thought") to solve problems. Note that "simulated" in this case doesn't mean that the models do not reason at all but rather that they do not necessarily reason using the same techniques as humans. That distinction is important because human reasoning itself is difficult to define.

Read full article

Comments

Elon Musk raced to build Colossus, the world's largest supercomputer, in Memphis, Tennessee. He bragged that construction only took 122 days and expected that his biggest AI rivals would struggle to catch up.

To leap ahead, his firm xAI "removed whatever was unnecessary" to complete the build, questioning "everything" that might delay operations and taking the timeline "into our own hands," xAI's website said.

Now, xAI is facing calls to shut down gas turbines that power the supercomputer, as Memphis residents in historically Black communities—which have long suffered from industrial pollution causing poor air quality and decreasing life expectancy—allege that xAI has been secretly running more turbines than the local government knows, without permits.

Read full article

Comments

Nearly a year after announcing the feature, Microsoft is finally ready to roll the controversial Windows Recall feature out to the general public, the company announced today on its Windows Experience Blog.

Only available on Copilot+ PCs, a subset of Windows 11 systems sold within the last year or so, Recall takes continuous screenshots of everything you do on your PC, saving them, scraping text from them, and saving it all in a searchable database. This obviously has major security and privacy implications—anyone who can get access to your Recall database can see nearly everything you've done on your PC—which is why Microsoft's initial rollout attempt was such a mess.

Recall's long road to release involved a rushed initial almost-launch, harsh criticism of its (then mostly nonexistent) security protections, multiple delays, a major under-the-hood overhaul, and five months of testing in Microsoft's Windows Insider beta program. Microsoft signaled that Recall was nearly ready for release two weeks ago when it came to the near-final Release Preview channel.

Read full article

Comments

Ironically, our present AI age has shone a bright spotlight on the immense value of human creativity as breakthroughs in technology threaten to undermine it. As tech giants rush to build newer AI models, their web crawlers vacuum up creative content, and those same models spew floods of synthetic media, risking drowning out the human creative spark in an ocean of pablum.

Given this trajectory, AI-generated content may soon exceed the entire corpus of historical human creative works, making the preservation of the human creative ecosystem not just an ethical concern but an urgent imperative. The alternative is nothing less than a gradual homogenization of our cultural landscape, where machine learning flattens the richness of human expression into a mediocre statistical average.

A limited resource

By ingesting billions of creations, chatbots learn to talk, and image synthesizers learn to draw. Along the way, the AI companies behind them treat our shared culture like an inexhaustible resource to be strip-mined, with little thought for the consequences.

Read full article

Comments

We knew Microsoft was about to launch Recall for real this time, and now the software maker is making it available to all Copilot Plus PCs. Recall, a feature that screenshots almost everything you do on a Copilot Plus PC, will be available today alongside an improved AI-powered Windows search interface and a new Click to Do feature that’s very similar to Google’s Circle to Search.

Recall was originally supposed to launch at the same time as Copilot Plus PCs in June last year, but the feature was delayed following concerns raised by security researchers. Microsoft then planned to start publicly testing Recall in October, but pushed it back again to November to have more time to secure it further. Microsoft has now spent the past 10 months overhauling the security of Recall and making it an opt-in experience that you don’t have to enable if you’re concerned about the privacy implications.

“When we introduced Recall, we set out to address a common frustration: picking up where you
left off,“ explains Navjot Virk, corporate vice president of Windows Experiences at Microsoft. Recall is designed to improve how you search your PC, but taking snapshots that are categorized so it’s easy to search for vague memories instead of file names.

I spent a few weeks testing Recall last year and found it was creepy, clever, and compelling. Technologically it’s a great improvement to the Windows search interface, because it can understand images and content in a much more natural way. But it does create a privacy minefield because you’re suddenly storing a lot more information on your PC usage, and you still need to manage blocked apps and websites carefully.

Kevin Beaumont, one of the security researchers that first raised alarm bells over Recall, has been testing the final version recently and found that “Microsoft has made serious efforts to try to secure Recall.” The database is now encrypted, Recall attempts to filter sensitive data by default, and the feature is now an opt-in experience.

Beaumont does note that filtering sensitive apps and websites can be hit-and-miss though, and occasionally even buggy. He also says that you can access Recall through a simple four-digit PIN unlock option with Windows Hello, instead of it forcing more secure facial recognition or a fingerprint. Microsoft’s Recall website claims “you must have at least one biometric sign-in option enabled for Windows Hello, either facial recognition or a fingerprint, to launch and use Recall.”

Alongside Recall, Windows search is also getting some AI improvements on Copilot Plus PCs today. You can now use the File Explorer, Windows search box, or settings with natural language queries. That means instead of searching for file names or specific settings, you can now describe images or documents and get results. If you’re looking for an image of a brown dog you know you have saved somewhere, you can just ask for “brown dog” rather than having to know the file name or date the image was created.

Microsoft is also rolling out Click to Do today, which works a lot like Google’s Circle to Search. You activate it by using the Windows key + left mouse click, and it will provide actions for the text or images that are on your screen. This includes summarizing text or being able to quickly remove an object from an image.

Recall, the improved Windows search, and Click to Do will all be available today across all Copilot Plus PCs, but the text actions in Click to Do are currently limited to Qualcomm-powered devices, with AMD- and Intel-powered Copilot Plus PCs getting this feature “in the next few months.” Recall and Click to Do should be available in a variety of languages and regions, but Microsoft says both features won’t be available in EU countries and Iceland, Liechtenstein, and Norway until later this year.

• Microsoft is trying to simplify AI sales, according to slides from an internal presentation.
• The current approach slowed sales, confused customers, and affected cost and quality, insiders say.
• Microsoft plans to slash the number of "solution areas."

Microsoft is trying to simplify its many AI offerings by streamlining how the products are pitched to customers, according to internal slides from a recent presentation.

The software giant has a bunch of different AI tools called Copilot. There's Copilot for its Teams chat app, Copilot for its PowerPoint presentation tool, Copilot for its Outlook email service — just to name a few.

These products are often split into different "solution areas," as Microsoft calls them. Having Copilot tools in many different buckets can slow down sales, confuse customers, and affect cost and quality of the tools, people in the organization told Business Insider. They asked not to be identified discussing private matters.

Microsoft has sales teams focused on each solution area, which will now be consolidated.

Microsoft Chief Commercial Officer Judson Althoff this week unveiled plans for addressing these issues in the company's upcoming fiscal year, which begins in July. BI obtained copies of slides from his presentation.

According to one of the slides, three major changes include:

• Consolidate Microsoft's solution areas.
• Accelerate regional skills at scale.
• Align teams working with small, medium, and corporate customers with those working with outside channel partners who market and sell Microsoft products.

The organization currently has six solutions areas: Modern work, Business Applications, Digital & App Innovation, Data & AI, Azure Infrastructure, and Security.

Beginning in July, these areas will be combined into three: AI Business Solutions, Cloud & AI Platforms, and Security.

AI Business Solutions will include tools such as Copilot for Microsoft 365, Copilot for Teams, Copilot for Outlook, plus a data visualization product called Power BI, according to a person who attended a Thursday all-hands for Althoff's organization. This person asked not to be identified discussing private matters.

"We are evolving the commercial solution areas within our sales organization to better reflect the era of AI and support the growth of our customers and partners," a Microsoft spokesperson said in a statement. "This evolution reflects the shift in how customers and partners are buying and will better serve their needs."

The other changes include expanding training for salespeople and a reorganization to Small, Medium Enterprise & Channel (SME&C) team, which was announced internally earlier this year.

The changes come as Microsoft is trying to figured out how to make money from its significant AI investments. It has mulled changes including new software bundles with Copilot. The company earlier this year said it plans to spend $80 billion on expanding its network of AI data centers.

Have a tip? Contact this reporter via email at [email protected] or Signal at +1-425-344-8242. Use a personal email address and a nonwork device; here's our guide to sharing information securely.

• Google's AI answers will give you a definition of any made-up saying. I tried: "You can't lick a badger twice."
• This is exactly the kind of thing AI should be really good at — explaining language use. But something's off.
• Is it a hallucination, or AI just being too eager to please?

What does "You can't lick a badger twice" mean?

Like many English sayings — "A bird in the hand is worth two in the bush," "A watched pot never boils" — it isn't even true. Frankly, nothing stops you from licking a badger as often as you'd like, although I don't recommend it.

(I'm sure Business Insider's lawyers would like me to insist you exercise caution when encountering wildlife, and that we cannot be held liable for any rabies infections.)

If the phrase doesn't ring a bell to you, it's because, unlike "rings a bell," it is not actually a genuine saying — or idiom — in the English language.

But Google's AI Overview sure thinks it's real, and will happily give you a detailed answer of what the phrase means.

Greg Jenner, a British historian and podcaster, saw people talking about this phenomenon on Threads and wanted to try it himself with a made-up idiom. The badger phrase "just popped into my head," he told Business Insider. His Google search spit out an answer that seemed reasonable.

I wanted to try this myself, so I made up a few fake phrases — like "You can't fit a duck in a pencil" — and added "meaning" onto my search query.

Google took me seriously and explained:

So I tried some others, like "The Road is full of salsa." (This one I'd like to see being used in real life, personally.)

A Google spokeswoman told me, basically, that its AI systems are trying their best to give you what you want — but that when people purposely try to play games, sometimes the AI can't exactly keep up.

"When people do nonsensical or 'false premise' searches, our systems will try to find the most relevant results based on the limited web content available," spokeswoman Meghann Farnsworth said.

"This is true of Search overall — and in some cases, AI Overviews will also trigger in an effort to provide helpful context."

Basically, AI Overviews aren't perfect (duh), and these fake idioms are "false premise" searches that are purposely intended to trip it up (fair enough).

Google does try to limit the AI Overviews from answering things that are "data voids," i.e., when there are no good web results to a question.

But clearly, it doesn't always work.

I have some ideas about what's going on here — some of it is good and useful, some of it isn't. As one might even say, it's a mixed bag.

But first, one more made-up phrase that Google tried hard to find meaning for: "Don't kiss the doorknob." Says Google's AI Overview:

So what's going on here?

The Good:

English is full of idioms like "kick the bucket" or "piece of cake." These can be confusing if English isn't your first language (and frankly, they're often confusing for native speakers, too). My case in point is that the phrase is commonly misstated as "case and point."

So it makes lots of sense that people would often be Googling to understand the meaning of a phrase they came across that they don't understand. And in theory, this is a great use for the AI Overview answers: You want to see the simply-stated answer right away, not click on a link.

The Bad:

AI should be really good at this particular thing. LLMs are trained on vast amounts of the English written language — reams of books, websites, YouTube transcriptions, etc., so being able to recognize idioms is something they should be very good at doing.

The fact that it's making mistakes here is not ideal. What's going wrong that Google's AI Overview isn't giving the real answer, which is "That isn't a phrase, you idiot"? Is it just a classic AI hallucination?

The ugly:

Comparatively, ChatGPT gave a better answer when I asked it about the badger phrase. It told me that it was not a standard English idiom, even though it had the vaguely folksy sound of one. Then it offered, "If we treat it like a real idiom (for fun)," and gave a possible definition.

So this isn't a problem across all AI — it seems to be a Google problem.

This is somewhat different from last year's Google AI Overview answers fiasco where the results pulled in information from places like Reddit without considering sarcasm — remember when it suggested people should eat rocks for minerals or put glue in their pizza (someone on Reddit had once joked about glue in pizza, which seems to be where it drew from).

This is all very low-stakes and silly fun, making up fake phrases, but it speaks to the bigger, uglier problems with AI becoming more and more enmeshed in how we use the internet. It means Google searches are somehow worse, and since people start to rely on these more and more, that bad information is just getting out there into the world and taken as fact.

Sure, AI search will get better and more accurate, but what growing pains will we endure while we're in this middle phase of a kinda wonky, kinda garbage-y, slop-filled AI internet?

AI is here, it's already changing our lives. There's no going back, the horse has left the barn. Or as they say, you can't lick a badger twice.

• Meta is building a $10 billion AI data center in northeast Louisiana.
• Entergy Louisiana says it needs to build a major new natural gas plant to power the facility.
• The utility wants permission to pass the plant's construction costs onto its entire customer base.

In Louisiana, a battle is heating up over who will pay for the new power plants needed to serve the $10 billion data center Meta is building in the state's northeastern corner.

Meta announced in December the 2,000-acre Richland Parish data center campus, which is expected to be completed in 2030. The company plans to use the data center to train AI models.

Consumer advocates and climate groups filed new testimony with the Louisiana Public Service Commission on April 11, pushing the state regulator to reject a request by Meta's electricity provider to shift $5 billion in construction costs for the plants on its entire customer base.

Entergy Louisiana has proposed building three new natural gas power plants to serve Meta's data centers in the state. As an investor-owned utility, Entergy can seek regulatory approval to bill its customers for the costs of building the new plants as long as it successfully shows that the plants are in the public interest.

Entergy has argued that Meta's data center could be "transformative" for Lousiana once built, saying the facility could provide 300 to 500 jobs with an average salary of $82,000.

The testimony from advocates and climate groups argued that the utility's 1.1 million electric customers shouldn't have to foot the bill for Meta's power appetite.

Entergy did not respond to a request for comment from Business Insider. Meta declined to comment.

Beyond the $5 billion in cost recovery for the plants, the utility's plan would "put other ratepayers at risk of having to absorb hundreds of millions, if not billions of dollars, of additional costs" associated with Meta's data center, said Cathy Kunkel, an energy consultant testifying on behalf of the Union of Concerned Scientists and Alliance for Affordable Energy.

That's because Meta's power needs have grown since the Big Tech giant first announced its plans to build an AI data center in Louisiana, adding to Entergy's costs of service.

Earlier this year, Meta was planning for more than two gigawatts of capacity in Louisiana, according to a Threads post by Meta CEO Mark Zuckerberg in January. That amount of electricity is double what is being planned at the Crusoe data center in Abilene, Texas, widely thought to be the first site for Stargate, a $500 billion joint venture between Oracle, OpenAI, and SoftBank to build AI data centers in the US.

Since then, Meta has asked Entergy for even more electricity, according to the utility's filings with the LPSC. The exact amount requested is unknown, as the information is redacted from the filings.

Saddling customers with Meta's costs is especially risky given the uncertainties surrounding AI's electricity demand, Kunkel said. AI models could become more energy efficient in the future, or companies could focus more on energy efficiency as a means to enhance profit, she said. That could result in Meta "choosing to scale back or exit" the project early. Recent reports that Amazon and Microsoft are pulling out of data center leases have helped stoke concern that the AI development boom is slowing down.

Entergy's request for cost recovery has attracted a wide range of "intervenors," or parties that want to weigh in with an opinion before the LPSC. Even Walmart in Louisiana has testified, seeking assurances that Meta's escalating power demand won't affect it and other Entergy customers in the state.

Have a tip? Contact this reporter on Signal at 929-524-6964. Use a nonwork device. Here's our guide to sharing information securely.

Pony AI (NASDAQ: PONY) stock just keeps on galloping. For the third day in a row, the small-cap Chinese robotaxi and robotruck company rode higher on Friday, building on huge gains Wednesday and Thursday. The stock tacked on another 22.7% through 9:40 a.m. this morning and now is up more than 120% over the past three days.

Yes, you read that right: Pony doubled, and then went up even more.

Where to invest $1,000 right now? Our analyst team just revealed what they believe are the 10 best stocks to buy right now. Learn More »

Image source: Getty Images.

Pony express delivers good news

So far this week, the biggest little Chinese small cap you've never heard of (before this week) has announced it's producing three new robotaxi models in cooperation with Beijing Automotive Group, Guangzhou Automobile Group, and Toyota Motor (NYSE: TM), respectively, and that it has lined up local partner Hesai Group (NASDAQ: HSAI) to supply it with AT128 lidar sensors for its robotaxis.

Today, Pony added that it's partnering with another local company, Tencent Holdings (OTC: TCEHY), to further "advance autonomous driving technology and robotaxi commercial deployment."

This tie-up will pair Pony's "cutting-edge" autonomous driving system with tech products from Tencent, specifically the latter's Weixin (or "WeChat") social media, messaging, and payment app, Tencent Maps, and its "robust cloud computing, big data and AI infrastructure," all of which sound to me like logical add-ons to an electric-car-slash-robotaxi service.

Is Pony AI stock a buy?

Tencent might do well to ante up a bit of cash. Tencent's tech contributions are all great, but there's no mention of financial support in the press release, and Pony is still burning cash, and losing $274 million a year.

While the company has considerable cash reserves, a little more could go a long way to ensuring Pony stock doesn't go bankrupt before 2029, the first year it's expected to be profitable, according to analysts polled by S&P Global Market Intelligence. Meanwhile, until the financial picture firms up, I must still consider Pony a speculative, momentum-driven stock.

Should you invest $1,000 in Pony Ai right now?

Before you buy stock in Pony Ai, consider this:

The Motley Fool Stock Advisor analyst team just identified what they believe are the 10 best stocks for investors to buy now… and Pony Ai wasn’t one of them. The 10 stocks that made the cut could produce monster returns in the coming years.

Consider when Netflix made this list on December 17, 2004... if you invested $1,000 at the time of our recommendation, you’d have $591,533!* Or when Nvidia made this list on April 15, 2005... if you invested $1,000 at the time of our recommendation, you’d have $652,319!*

Now, it’s worth noting Stock Advisor’s total average return is 859% — a market-crushing outperformance compared to 158% for the S&P 500. Don’t miss out on the latest top 10 list, available when you join Stock Advisor.

See the 10 stocks »

*Stock Advisor returns as of April 21, 2025

Rich Smith has no position in any of the stocks mentioned. The Motley Fool has positions in and recommends Tencent. The Motley Fool has a disclosure policy.