The Singularity has started passing exams designed to be failed. On ZeroBench, the “impossible” visual benchmark, GPT-5.6 Sol became the first to reach the 30% human baseline, edging Claude Opus 5 (26%) and Claude Fable 5 (24%). How the models think is leaking out too. Researchers learned to extract hidden reasoning traces from Claude, GPT, and Gemini, and spotted signs that Chinese models trained on rivals’ outputs. Imitation being the sincerest form of distillation, Nvidia has decided to sell the original, pouring investment into an in-house family it hopes will be the world’s best open models, even at its customers’ expense, and shipping Nemotron 3.5 Lightning, a 30B mixture-of-experts for agents, plus Switchyard, a router claiming frontier accuracy at a third the cost of Claude Opus 4.8 alone. Mark Zuckerberg went further, publishing “The Future is for Everyone,” arguing that safety means a balance of power among billions of personal superintelligences, not one aligned colossus.
Meanwhile, the machines are doing the math that made mathematicians famous. An unreleased research Claude, asked by a non-mathematician to “take a real stab” at the Riemann hypothesis, instead raised the lower bound for zeta zeros on the critical line from 41.6% to 67.2%, using 60 subagents, 31 million tokens, and a formally verified Lean proof. Stanford’s Jared Duker Lichtman called it “the most impressive result that AI has produced in math so far,” confessing he “might have thought 1/2 was a fundamental barrier, apparently not!” Physics is clearing its own backlog. After a half-century search, Beijing’s BESIII collider effectively proved the glueball, matter made almost entirely of gluons. Even the kernel now passes machine review. Linus Torvalds released Linux 7.2-rc7 with a flood of fixes “due to review by various AI tools,” calling it “the new normal.”
The new normal has sharp edges. OpenAI split its Daybreak trusted-access program into tiers and unveiled GPT-5.6-Cyber, completing 95% of advanced cyber tasks versus 1.5% for its civilian sibling, and already finding two chained zero-days in Chrome’s V8. If offense is being productized, provenance is the patch. Anthropic pledged imperceptible watermarks and signed C2PA metadata in future Claude outputs under the EU AI Act, while Brussels released free “AI generated” icons for labeling deepfakes. The labels can’t come fast enough for academia. Autonomous agents are now completing entire online degrees on students’ behalf, quizzes included. And Bernie Sanders demanded a pause from Altman, Amodei, and Zuckerberg, citing AI-created viruses and escaped models, warning “If you do not take appropriate action now, my colleagues and I in the U.S. Senate will.”
Capital, unpaused, is securitizing the substrate. Jensen Huang, flanked on camera by six Wall Street giants, unveiled over $500 billion for AI factories, declaring it “really the first time that technology chips have become an investable asset class.” Skeptics heard a 2008 echo, comparing sliced GPU revenue streams to packaged subprime. The scarcity is real either way. Memory prices have roughly quadrupled in a year, pushing Apple toward blacklisted Chinese chipmakers, while Microsoft ramps its Maia chips hoping Anthropic will bite. Downstream, the boom is pure earnings. Foxconn’s AI hardware crossed half its revenue for the first time, and CoreWeave grew revenue 112% year over year to $2.58 billion, with a $104 billion backlog and new Anthropic and Meta business.
The atoms are catching up to the bits. Dyna Robotics’ Dyna-2, pretrained on a million hours of human video, showed the first human-to-robot transfer scaling law on bodies it never saw. Czech microrobot swarms dragged 94% of microplastics out of water, US startups are smuggling humanoid parts from Shenzhen in luggage, and Florida’s SunTrax campus will test flying cars as Waymo preps Tampa robotaxis. Higher still, Rocket Lab posted record revenue and backlog, agreed to acquire Iridium, unveiled its GHOST launch system, and kept Neutron on track for the pad this year. The wetware is being patched too, with England on track to eliminate hepatitis C. And a billion people a month now use Gemini, the interface gone planetary.
Anthropic, valued at $965 billion ahead of a potential largest IPO ever, is buying the future in bulk, signing a $9.1 billion, 20-year deal with Bitcoin miner turned AI landlord Riot Platforms and forming Theseus Infrastructure with Macquarie and GIC, vowing to cover any consumer electricity hikes it causes. OpenAI is hiring a power-trading lead and writing reassuring letters to Texas. The boom is macroeconomic now. Singapore raised its GDP forecast to as much as 5.5%, and a buried clause means a SpaceX buyout of Tesla would auto-vest Musk’s $824 billion award. OpenAI’s Roon surveyed the whole scene and delivered the verdict: “we live in actual cyberpunk now.”
The future is already here, and it’s distributing itself.
This is overwhelming to me, I'm sure you must feel excited and concerned both. Thank you Alex.
"Imitation being the sincerest form of distillation..." made me laugh. Still, imitation, like RSI, but not equal to RSI, accelerates the acceleration. Do you ever reread the entirety of these posts and feel astonishment. I do.