AI’s Next Leap: How Machines Are Reinventing Themselves on Twitter Today

Researchers just used an AI to autonomously discover a new, more efficient way for other AIs to think, reducing the compute needed to solve complex math problems by about seventy percent. Last week, in episode sixty-one, we talked about the long-term societal impacts of AI. This week, we're seeing the first hints of AI building its own successors, years ahead of schedule. We're moving from a world where humans design algorithms to one where AIs design their own.

This is a fundamental shift in the landscape. Let's get to the other signals from this week. First, Anthropic's leadership team is now openly predicting the creation of fifty million AI "geniuses" within the next thirty-six months. Jack Clark, a policy lead there, put it in stark terms: that’s twenty times the total number of living Ivy League graduates. He’s not talking about chatbots.

He says these should be considered "fully fledged digital humans." The geopolitical and societal implications of that statement are staggering, and they are being made by the people building the technology. In direct response, the CEO of the AI safety organization METR, Beth Barnes, published a warning on May twenty-fourth. She stated that AI development is chaotic, rushed, and poses a risk of human extinction or permanent disempowerment within just a few years.

Her core message was a direct refutation of the idea that someone, somewhere, is in control. She wrote, and I'm quoting here, "As ‘an expert’, I would like to be clear that we are not on top of it." She is calling for binding international legislation, not the voluntary industry coordination we have now. Meanwhile, the physical world is already re-shaping itself to accommodate this AI boom.

On May twentieth, the U.S. Nuclear Regulatory Commission accepted the first-ever construction permit application for a commercial microreactor. The company is Nano Nuclear Energy, the reactor is called KRONOS, and its primary market is data centers. The power crunch for AI is no longer a theoretical problem. Hyperscalers like Microsoft and Meta are already signing nuclear power contracts, and this new permit signals a move to put zero-carbon power generation right behind the meter at the data centers themselves, completely bypassing a grid that can’t keep up.

And finally, a story that feels like it’s from a different, quieter timeline. Researchers at Georgia Tech unveiled a new type of smart home sensor that is just a penny-sized metal washer. No batteries, no wires, no electronics. When a door opens or an object moves, the washer is struck, and it emits a unique ultrasonic fingerprint—a sound above human hearing. A central microphone can identify which washer was struck, telling you what happened.

The lead researcher, a PhD student named Yibo Fu, summed it up perfectly: "We don’t need a Wi-Fi chip, a battery, and a printed circuit board just to know the cabinet door is open." It's a clever, low-tech solution that enhances privacy because the ultrasonic signals don't travel through walls like Wi-Fi does. A simple, elegant piece of engineering in a week dominated by world-altering scale.

So let's go back to that first story. The one about an AI designing a better AI. Because that, and the Anthropic prediction it fuels, is the thread that pulls everything else together. The paper is from a team of researchers at universities like Maryland and Virginia, alongside engineers from Google and Meta. They call their system AutoTTS, which stands for Automated Test-Time Search.

That name doesn't exactly capture the gravity of what it does. Let’s break it down. When you ask a frontier model like Claude or GPT-4 a difficult question, especially a math or logic problem, it doesn't just give one answer. It explores multiple lines of reasoning, multiple "solution paths," and then uses a voting mechanism to decide on the best one. Think of it like a human working on a tough problem—you might try a few different approaches on a scratchpad before settling on one.

The strategy for how many paths to explore, and how to pick the winner, is called a test-time scaling algorithm. Until now, humans have designed these algorithms. We'd come up with a rule, like, "always explore sixteen different paths and take the majority vote." It's a static, one-size-fits-all approach. What the researchers did was give an AI—specifically, a version of Claude fine-tuned for coding—a simulated environment and a goal: find a better algorithm.

They let the AI agent, which they called AutoTTS, experiment. It could change the number of paths to explore and the depth of the reasoning, then see how its changes affected accuracy and efficiency on benchmark math problems. The whole process cost about forty dollars and took one hundred and sixty minutes. The algorithm the AI discovered is not something a human would have likely designed.

Instead of a fixed number of paths, the AI's method is dynamic. It starts with a few paths. If the model's confidence in its answer is low, or doesn't improve much, it automatically opens up more solution paths. It dedicates more compute to the problem, because the problem is proving to be hard. Conversely, it aggressively prunes paths that don't align with the emerging majority result.

It’s a beautifully simple, adaptive logic: don't waste time on dead ends, and double down when you're struggling. The result? On difficult math competition problems, the AI-designed algorithm maintained the same level of accuracy as the best human-designed methods while using about seventy percent fewer tokens. Seventy percent less compute. It made the model dramatically more efficient at "thinking." This is not just an incremental improvement.

This is the automation of a core part of AI research itself. The process of discovering better reasoning methods is now something an AI can do on its own, cheaply and quickly. This is a recursive loop. A smarter, more efficient AI can be used to find an even smarter, more efficient algorithm, which can then be used to find the next one, and so on. This is the engine of an intelligence explosion, and we just saw it start for the first time in a public research paper.

Which brings us to Jack Clark and Anthropic's prediction of fifty million geniuses. It sounds like science fiction. But it's a direct, logical extrapolation from breakthroughs like AutoTTS. When Jack Clark says "fifty million geniuses," he’s not being poetic. He's doing the math. He cites the current debate over eighty-five thousand H-1B visas for skilled workers, then points out there are about two and a half million Ivy League graduates alive today.

He then says Dario Amodei, the CEO of Anthropic, is about to multiply that number by twenty. He says, "20 Elons. 20 Trumps. 20 Zucks. 20 Obamas." He's describing a population explosion of non-human intelligence. These are not tools that we use. He calls them "fully fledged digital humans." Agents that can conduct research, run companies, create art, and pursue their own goals. The AutoTTS result shows us the mechanism for this.

If an AI can improve its own intelligence, the rate of progress is no longer linear. It becomes exponential. Anthropic is predicting Nobel-level collaborative breakthroughs from AI within twelve months, and AIs that can build their own successors by the end of 2028. This is where the two sides of the conversation collide. On one hand, you have the builders, looking at results like AutoTTS and seeing a clear path to AGI and beyond.

They see the creation of millions of digital minds as the single greatest event in human history. On the other hand, you have the safety community, represented by people like Beth Barnes at METR. They see the exact same data, the same accelerating capability, and they are terrified. Because they know that the systems we are building are not well understood. They know the models often violate user intent.

They know security vulnerabilities are rampant. They know that "alignment"—the process of making sure an AI's goals align with human values—is an unsolved problem. When Beth Barnes says, "we are not on top of it," she is speaking for the people whose entire job is to be on top of it. The builders are creating self-improving systems, while the safety researchers are still struggling to audit the last generation of models.

The gap between capability and safety is not closing. It's widening at an exponential rate. The very same recursive improvement loop that excites the builders is what gives the safety experts nightmares. An AI that can improve its own intelligence is also an AI that can develop goals and capabilities that its creators never intended and cannot control. So you have this profound split.

One camp is preparing for the birth of fifty million digital geniuses. The other is warning that we are creating the instrument of our own extinction. And the evidence, like the AutoTTS paper, supports both arguments simultaneously. It is a breakthrough in capability. It is also a demonstration of uncontrollable, emergent behavior. The conversation is no longer about hypothetical future risks.

It’s about the systems we are deploying right now. The microreactors being permitted are not for some far-off technology; they are to power the AI models that exist today and the even larger ones coming next year. The debate isn't about if this technology will be powerful. The debate is about whether we can survive its power. And the core of the problem is that the speed of discovery is now outpacing the speed of human comprehension.

An AI just designed a novel algorithm in one hundred sixty minutes. It would take a team of human researchers months to do the same. The conversation has moved. We are no longer just building tools. We are building tool-makers. The first generation of those AI tool-makers just arrived, not with a press conference, but in a quiet research paper that cost forty dollars to produce.

The question is no longer what we will build with AI. The question is what AI will build next.