23-Year-Old Cracks Decades-Old Erdős Problem with One GPT-5.4 Pro Prompt as the Human-Level AI Plateau Keeps Stretching

Nick Bostrom said the thing that has surprised him most about this moment is how long the human-level AI era is lasting. He expected a sharp jump past us. Instead we have been sitting roughly at human level for three to five years now, and it might keep going. Demis Hassabis, who used to say AGI required one or two more breakthroughs to arrive, said this week it is closer to a coin flip whether any more are needed at all.

If that sounds like things have plateaued, the people building these systems would disagree. Sam Altman pointed out that the post-AGI predictions assumed nobody would have to work anymore, and yet his own users have started doing polyphasic sleep so they can ship more code with GPT-5.5 in Codex. OpenAI's Noam Brown framed the same point more strategically, saying model weights matter relatively less than they used to and securing inference compute matters more. The recipes themselves are getting shorter shelf lives. GPT-4o ran for 21 months. GPT-5.4 lasted 49 days.

The clearest example of what that compute can actually do came from a 23-year-old named Liam Price, who has no advanced math training and used a single GPT-5.4 Pro prompt to crack Erdős Problem #1196, which had eluded prominent mathematicians for decades. Terry Tao's read on it was that humans had hit a "mental block" because the field collectively "made a slight wrong turn at move one." A fresh perspective, even an artificial one, was enough to find the right turn.

All of this needs hardware, and the hardware story is getting more interesting. OpenAI is reportedly working with MediaTek and Qualcomm on its own AI smartphone processors, with Luxshare manufacturing and mass production targeted for 2028. The vision is an AI agent device that replaces apps with task-driven inference. Apple has six new product categories in development at the same time, including AI AirPods, smart glasses, a wearable pendant, a smart home display, a tabletop robot, and a security camera, extending the company's surface area from ear to room.

At the infrastructure end, Kevin O'Leary's "Wonder Valley" hyperscale data center campus in Utah's Box Elder County is approaching final approval. The plan is to generate up to 9 gigawatts of its own power and use a closed-loop water system that returns water to the aquifer feeding the Great Salt Lake, while consuming more electricity than the entire state of Utah today. AWS CEO Matt Garman, asked about retiring old GPUs, said the company has never retired an Nvidia A100 server because demand still outpaces supply. There is no version of this where compute gets less scarce in the near term.

Robotics is starting to absorb that compute too. China's State Grid is buying around 8,500 robots this year, including humanoid and dual-arm units that will perform high-voltage operations on its ultra-high-voltage grid. The strategic logic is that the optimal failure mode in high-voltage work is now a melted servo rather than a melted operator. Less reassuringly, fifteen Ceres Air C31 chemical-spraying drones were stolen from a CAC International facility in Harrison, New Jersey by a thief posing as a delivery driver with a fake bill of lading. Each drone can disperse 40 gallons of liquid over 15 acres in seven minutes, and the FBI is investigating what one retired agent called a possible "nightmare scenario."

The energy buildout to feed all of this is starting to show up in public markets. Nuclear startup X-energy raised $1.02B in its IPO and shares popped 36% on debut, valuing the company at roughly $11.5B. Geothermal startup Fervo filed for an IPO at a roughly $3B private valuation. Meta signed a deal for up to 1 gigawatt of space-based solar from Overview Energy, with sunlight collected by satellites in geosynchronous orbit and beamed down as low-intensity infrared light to existing solar farms, which until recently was a thing you only saw in science fiction.

Speaking of which, SpaceX's IPO filing approved a compensation package that grants Elon Musk up to 60 million additional shares if the company's market cap climbs from $1.1T to $6.6T while delivering "100 terawatts of compute per year" from space data centers. The filing itself concedes that figure is orders of magnitude beyond peak US power consumption today, that the technology is unverified, and that the orbital environment is harsh. Humans are practicing for the trip too. The European Space Agency just sealed six participants into a simulated Mars mission at the German Aerospace Center in Cologne, with no exit until August.

Biology had its own milestone this week. Intellia Therapeutics reported the first ever Phase 3 success for an in vivo CRISPR therapy. Lonvoguran ziclumeran, a single-dose treatment for hereditary angioedema, met its primary endpoint and freed most patients from attacks and ongoing therapy through the six-month efficacy window. CEO John Leonard called it "the first Phase 3 data in any indication with in vivo CRISPR where you're actually changing a gene that causes disease." Editing the human genome inside a living person is now something that works in a registrational trial, not just a paper.

Humans themselves are also still finding new ceilings. At the London Marathon, Sabastian Sawe ran 1:59:30 and Yomif Kejelcha ran 1:59:41, the first sanctioned, world-record-eligible sub-2-hour marathons in human history. Both wore the new Adidas Adizero Adios Pro Evo 3, a 97-gram carbon-plated racer about 30% lighter than its predecessor, which is becoming as much a part of the story as the athletes.

It is worth noting where the money actually is right now, because it is not where you might think. Ozempic and Mounjaro generated roughly $71B in revenue in 2025. OpenAI and Anthropic combined did about $29B. The longevity industry is the bigger business, at least for the moment.

The institutions are starting to react. Major US insurers including Berkshire Hathaway, Chubb, and Travelers have won regulatory approval to drop AI-related damages from corporate liability policies, excluding claims like agents misusing copyrighted material in marketing, which is a quiet but meaningful repricing of risk. A bipartisan House AI bill led by Reps. Ted Lieu and Jay Obernolte targets the spread of deepfakes and protects whistleblowers reporting AI safety risks. The Supreme Court heard arguments on whether geofence warrants, which let police pull cellphone location data from everyone near a crime scene, violate the Fourth Amendment, with a ruling expected by late June. And as a small reminder that not every analog thing is dying, the US has 70% more independent bookstores than it did in 2020, up from 1,916 to 3,218 according to the American Booksellers Association.

Looking further out, Varda's Andrew McCalip predicted an AI-led S&P 500 company within five years and an AI elected official within twenty. Maybe in anticipation of that, Sam Altman published five core principles to guide OpenAI this week: democratization, empowerment, universal prosperity, resilience, and adaptability.

Matter learned to think, and now the thinking is learning to leave.

That's today. More tomorrow.