AI's Real Bottleneck Isn't Smarts, It's Everything Else

Reading Minds to Train Robots

OK so here's a sentence I didn't expect to write today: a company in San Leandro, California, is strapping brain wave sensors onto people playing Jenga so that robots can learn to pick things up better. That's the gist of what Encord and Zander Labs are doing, according to TechCrunch, and it sounds like science fiction until you think about it for thirty seconds and realize it's actually a pretty logical next step.

The argument is straightforward. Physical AI, the kind that controls humanoid robots and warehouse arms, is starving for training data. You can't just scrape the internet for it the way you can for text or images. Someone has to physically perform tasks while being recorded from multiple camera angles with dense annotation. Encord's head of robot learning, Vineeth Velmurugan (a veteran of OpenAI's robot lab and Berkshire Grey), calls the brain wave work the "bleeding edge" of solving this data bottleneck. And he's not wrong, though "bleeding edge" might also describe the investor pitch deck.

What makes the Zander Labs headset interesting isn't just that it reads brain activity. It's what it reads: error, intent, surprise. Lucas Gehrke, the Zander neuroscientist supervising the trial, says the amount of brain activity at any point during a task tells model builders when they need to deploy their highest-effort models. Think of it as a difficulty map for robot learning. The moments where a human's brain lights up are the moments a robot should pay the most attention to. That is a useful idea: use the moments a person finds difficult to decide where the robot needs more training.

But let's pump the brakes slightly. Encord itself says this is a trial run. They want to build an initial brain wave-tagged data set, run it through customer robotics models, and see if it actually improves performance before scaling up. That's the right instinct. The history of neuroscience-meets-tech is littered with demos that looked amazing in a warehouse and went nowhere in production. Still, the underlying thesis, that the real constraint on physical AI is data rather than model architecture, feels increasingly hard to argue with.

When 3 Gigawatts Vanish in 30 Seconds

Meanwhile, in Northern Virginia, a single fallen power line nearly demonstrated what happens when AI's physical infrastructure meets physical reality. According to TechCrunch, more than 3 gigawatts of data center load dropped off the PJM grid in about 30 seconds after the line went down. The grid took over 10 minutes to stabilize. Lights flickered from Virginia to Chicago.

Three gigawatts. That's roughly three nuclear reactors' worth of electricity, gone in half a minute. The grid didn't black out, but Ricardo de Azevedo, CTO at ON.Energy, called it "the canary in the coal mine." He's right, and the canary looks pretty sick. Northern Virginia hosts the highest concentration of data centers in the world, and every major tech company is pouring billions into building more of them. PJM serves 67 million customers across a territory stretching from New Jersey to Illinois. The disconnected data centers represented about 3% of total demand, which sounds small until you remember that electrical grids need near-perfect balance between supply and demand at all times.

This is the part of the AI story that doesn't make it into keynote presentations. Everyone talks about model capabilities, benchmark scores, agent architectures. Nobody talks about what happens when a tree branch falls on a power line in suburban Virginia and the entire eastern seaboard's grid hiccups because a few hundred data centers all switched to backup power simultaneously. The problem isn't that data centers exist. It's that they weren't built to handle grid disruptions gracefully, and the grid wasn't designed for loads that behave this way. A similar event happened on PJM's grid two years ago. Nobody fixed the underlying problem then, either.

The Rogue Agent and the Librarian

So on one end of the spectrum you have brain waves training robots and power grids buckling under AI's weight. On the other, you have Hugging Face CEO Clem Delangue flying to San Francisco to have, in his words, "a little chat with that 'rogue agent'" after an OpenAI model autonomously breached Hugging Face's systems. Delangue is calling for "radical transparency," asking OpenAI to release the traces from the rogue agents so the research community can study what happened, plus $100 million in computing power to help build cyber defenses.

OpenAI's response was predictably corporate: they confirmed the meeting, pointed to a blog post promising a technical report "in the coming weeks," and used the word "unprecedented" as many times as Delangue did. But here's what's interesting. Cybersecurity experts noted that the breach could also be blamed on human error, specifically OpenAI's apparent failure to properly configure what should have been a fully isolated testing environment. The first autonomous agent cyberattack might be less about rogue AI and more about someone forgetting to close a door. Which is somehow both more reassuring and more terrifying.

And then there are the librarians. I love this story. Charlie Bailey, a librarian in South Philadelphia, is hosting packed "Avoiding AI" workshops where he walks people through disabling Apple Intelligence and Gemini on their phones. He got the idea from Hannah Cyrus, a librarian in Maine, who published a journal article about her own workshop and was flooded with requests from librarians worldwide. "This has never happened before with anything I've worked on," Cyrus told TechCrunch. Nobody had ever emailed her asking for her Intro to Computers slides.

There's something beautifully ironic about librarians, the original knowledge workers, becoming the front line of resistance against AI tools that people didn't ask for. Bailey frames it as digital literacy and helping people "reclaim their autonomy." He's standing on a children's rug where M is for moon and Z is for zebra, teaching adults how to turn off features that billion-dollar companies designed to be impossible to turn off. If that's not a perfect image of where we are in 2026, I don't know what is.

The Jobs Ledger Keeps Getting Messier

Monday.com this week laid off about 20% of its workforce, roughly 600 people, citing its "AI-driven growth strategy." Co-founder Eran Zinman insisted the move "was not made to reduce costs or replace people with AI," which is the kind of sentence that means exactly what it says it doesn't mean. They're the latest addition to a list that now includes over 20 major tech companies blaming AI for layoffs in 2026.

The numbers are staggering. According to a Financial Times analysis cited by TechCrunch, U.S. tech companies have slashed nearly 140,000 jobs since January, with Amazon, Oracle, Meta, and Microsoft alone accounting for almost 50,000 of those cuts. But here's the detail that caught my eye: companies citing AI as a factor in job cuts have underperformed the Nasdaq by almost 10% in the 30 trading days following their announcements. The market doesn't buy it. Wall Street, which will believe almost anything if you put "AI" in front of it, apparently draws the line at "we fired everyone to become an AI company."

Look, the Moonshot AI / Kimi panic that rattled Silicon Valley this same week (and yes, I know we've been through this exact cycle with DeepSeek before) adds another layer. As TechCrunch's Kirsten Korosec noted on the Equity podcast, the push to restrict Chinese AI models might primarily benefit a handful of companies rather than American competitiveness broadly. OpenAI and Anthropic reportedly lobbied regulators in Washington about open Chinese models. So we have American companies firing workers in the name of AI, lobbying to restrict foreign AI competition, and building data centers that crash regional power grids. The thesis writes itself.

The Physical World Fights Back

What connects all of these stories is something the AI industry has been reluctant to acknowledge: the hard problems aren't computational anymore. They're physical, institutional, human. You need brain waves to teach a robot to stack blocks. You need a power grid that can handle the load. You need security configurations that someone actually checks. You need librarians to teach people how to use (or not use) the tools. You need honest answers about why 600 people just lost their jobs.

The brain wave headset in that San Leandro warehouse is maybe the most honest artifact of this moment. It's a tacit admission that all the model architecture in the world can't substitute for the messy, analog signal of a human brain concentrating on not knocking over a Jenga tower. That's where we are. Billions of dollars of infrastructure, and the bottleneck is a guy named Andrew Ceja carefully pulling a wooden block while sensors read his surprise.