David Pogue on Substack writes that "When you do a Google search these days, you generally see an AI Overview panel above the search results. It’s intended to summarize the answers to your query, so you don’t have to click any links. The first problem: By Google’s own calculations, the AI Overviews are incorrect 28% of the time. The bigger problem: AI is an environmental disaster. It’s already a monstrous energy hog, and its appetite is doubling every six months."

 He gives some data about this data center power situation:

• 4,200 data centers that AI companies have built and 1,500 more are going up as you read this
• By 2030, AI will consume 945 terawatt-hours of electricity. That is enough to power every household in California, Texas, Florida, New York, Ohio, and Pennsylvania combined. Almost incomprehensible.
• 60% of that power will come from polluting power sources.
• Don’t care about the environment? How about your power bill? AI’s power needs have driven up electricity costs as much as 15% in the last year, with another 8.5% hike coming by the end of 2026.
• Add in more rolling blackouts during heat waves this summer.

But it’s not just Google, because almost every big company is eager to add AI to their products.

Pogue's note of hope is that a few people, like Sheila Morovati, are trying to make AI optional. Morovati is the founder and president of a nonprofit called HabitsofWaste.org. Her movement is called Opt-In AI with a goal of no AI at all unless someone asks for it. The default setting should be the most sustainable and least annoying option.

More at Rise Up, People! Make AI Optional! - David Pogue

It was formally proposed by engineer Peter Glasser in 1968, a space pioneer who introduced the idea of using satellites to beam solar energy from space down to Earth. Over the decades, what Glaser envisioned has been known by many names — space-based solar power (SBSP), solar-power satellites or satellite power system (SPS), as well as satellite solar-power system (SSPS). Glaser's contributions to space science and technology were not limited to the solar-power satellite concept. He also worked on NASA's Apollo moon missions and headed an experiment that flew aboard the space shuttle Columbia in 1986.

But solar power from space beamed to Earth has remained mostly theoretical due to cost and complexity.

Data centers need power. A lot of power. People don't want data centers in their neighborhoods. Where will it come from? From space?

Meta announced a deal with startup Overview Energy to purchase solar power collected by satellite and beamed back to Earth.

It is an experimental approach that could power data centers at night. Unlike traditional solar power, which relies on storing daylight, space-based solar power aims to deliver continuous energy.

Overview Energy plans to deploy satellites over 22,000 miles from Earth's equator, where they would collect and transmit infrared energy to solar panels. A test is scheduled for 2028, with a commercial rollout in 2030. Meta is seeking up to 1 gigawatt of power from the project, underscoring its energy needs for AI.

It sounds a bit wishful thinking if you look at the numbers. In 2024, Meta's data centers consumed 18,000 times the electricity that this deal would deliver in a single hour. 

Space-based solar power (SBSP) involves harvesting solar energy in orbit and beaming it to Earth, providing 24/7 clean energy unaffected by weather, nighttime, or atmospheric filtering. There are challanges: high launch costs, complex orbital assembly of massive structures, and wireless energy transfer. 

During the next decade, there will be a steady drop in the number of this nation’s 18-year-olds, which will almost certainly lead to a spike in college closures and mergers throughout the country, not only at small private schools with less-than-élite academic reputations but also at large regional public schools.

“If my kid does want to attend college in 2035, how many schools will she actually have to choose from?” Jay Caspian Kang asks in this article on The New Yorker https://newyorkermag.visitlink.me/8QmCvL

Cybercrime group ShinyHunters identified itself as the hackers.

ShinyHunters first emerged in 2020 and claims to have successfully attacked 91 victims so far. The group is primarily after money, but has also been willing to cause reputational damage to their victims. In 2021, ShinyHunters announced they were selling data stolen from 73 million AT&T customers. ShinyHunters received global attention in 2025 after Google urged 2.5 billion users to tighten their security following a data breach via Salesforce, a customer management platform.

Unlike data breaches where hackers directly break into databases holding valuable information, ShinyHunters – and several other groups – have recently targeted major companies through voice-based social engineering, which is also known as “vishing,” for voice phishing. Social engineering is when a person is tricked or manipulated into providing information or performing actions that they wouldn’t normally do.

In AI terms, there are three kinds of intelligence. "Artificial Narrow Intelligence" is what we have now. It is "superhuman" at specific tasks like playing Go or translating languages. ChatGPT, Gemini, CoPilot and Meta AI, et al fit in there at the moment.

"Artificial General Intelligence (AGI)" is human-level across the board and can learn anything a person can learn. We’re not quite there yet as of May 2026.

"Artificial Superintelligence (ASI)" is far beyond human level. Philosopher Nick Bostrom popularized the term: and defined it as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest.”

ASI is what people worry about — or get excited about — when talking about advanced AI. But AGI isn't quite the same as superintelligence. With AGI, you clone the best human brain in software, but with superintelligence that clone keeps upgrading itself until it’s as far beyond us. 

Two new tools are moving closer to the next level.

Google has released TurboQuant, a new compression method that makes AI models cheaper to run and faster to respond. In Google’s reported tests, it reduced the key-value cache, the model’s short-term working memory while it responds, by at least 6x and improved performance by up to 8x on H100 chips, Nvidia’s high-end AI processors used in data centres, while keeping benchmark performance, or standard test performance, close to the original model. That is a serious technical result with a clear business consequence: one of the biggest cost pressures in modern AI may begin to ease. For the past two years, the default logic has been simple. The best AI stayed in the cloud because that is where companies could absorb the cost of running it. TurboQuant starts to weaken that logic.

Meta TRIBE v2 is a foundation AI model that acts like a “digital twin” of the human brain. In plain terms, it’s an AI trained on real brain scan data so it can predict how a person’s brain will respond to things they see, hear, or read. It takes in video, audio, and text, then maps that to about 70,000 areas of the brain to simulate neural activity.  Meta itself says that you can think of it as Meta teaching an AI to “think” more as humans do, by learning directly from brain responses instead of just internet text.

Where did I get information anout Meta's products and path? From their own Muse Spark. That is Meta’s latest (well, as of today) AI assistant model.