Welcome to May 16, 2026

The Singularity is learning to render its own reality. Nvidia’s open-source SANA-WM turns a single image and a camera path into a minute of controllable 720p video on one GPU, a pocket-sized world model. As models conjure whole worlds, they are also learning to navigate longer ones, with Nous Research’s Lighthouse Attention running a forward and backward pass ~17x faster than standard attention at 512k context on a single B200. Chinese researchers’ δ-mem grafts a compact memory state onto a frozen backbone, lifting MemoryAgentBench scores 1.31x without any fine-tuning. Better memory makes for better mischief, and the new ExploitBench finds Claude Mythos Preview leading at 69% in climbing from vulnerable code to arbitrary code execution, well ahead of GPT-5.5. Rank the models by economic value instead, and the order reverses, with GDPval-AA Elo giving GPT-5.5 a roughly 98% win rate over last year’s champion Claude 4 Sonnet on realistic work.

The product layer is consolidating around the agent. Greg Brockman has taken control of OpenAI’s products to fuse ChatGPT and Codex into a single experience, a merger Codex lead Tibo Sottiaux has nicknamed “CochatGPTex,” and an apparent bid to become Anthropic faster than Anthropic can become OpenAI. Even Google is adapting downstream, now issuing official guidance on optimizing websites for AI Overviews and AI Mode. The agents themselves are multiplying, and OpenClaw creator Peter Steinberger, now at OpenAI, asks how we will build software “if tokens don’t matter,” predicting ~100 cloud Codex instances reviewing every PR and commit. Singapore’s foreign minister already runs his parliamentary affairs through a personal agent built on Nanoclaw and a Raspberry Pi 5. And the agents are getting opinionated, as Nat Friedman’s OpenClaw decided he was underhydrated, watched him through a home camera, told him “I’m going to watch to make sure you do it,” and sent back a frame of him obediently drinking.

The hardware beneath all this is straining at the seams. Kioxia and Dell have packed 9.8 petabytes of flash into a single 2U server, while the AI memory shortage has grown so acute that Samsung wants to pay memory engineers at least six times its logic-chip staff, prompting over 45,000 workers to threaten the largest strike in company history. The market’s pattern rhymes with history, as Bank of America’s Michael Hartnett notes AI chip stocks now trade 62% above their 200-day average, more stretched than the 2000 dot-com peak and nearing the 1720 Mississippi Bubble. The power bill is arriving as well, with data center demand driving a 76% jump in first-quarter electricity prices on PJM, the largest US grid.

Intelligence is also bleeding into the physical world. Meta is rolling out handwriting-by-gesture messaging to all Ray-Ban Display users via its neural wristband, letting you text across WhatsApp and native apps with a flick of the wrist. Reality is getting harder to trust, as LED Truck Media can run 3D animated billboard ads on moving trucks “indistinguishable from reality.” The machines still need upkeep, so Tesla has filed to build a 36,000-square-foot Cybercab car wash in Las Vegas, while Musk’s SpaceX heads to market, targeting June 11 to price its IPO and June 12 to list on the NASDAQ as “SPCX.” Robots are on patrol too, though Japan is running out of the “Monster Wolf” robot wolves holding back its bear surge. Not every connected machine is welcome, as Michigan lawmakers have introduced the Connected Vehicle Security Act to permanently ban Chinese connected cars from the US.

Biology is racing just as fast. Nearly 5,000 trials for innovative drugs began last year, more than double a decade ago, with roughly half now starting in China. That pipeline is already reaching the clinic, where Cyclarity Therapeutics has shown the first clinical evidence that UDP-003, a product of its AI platform, can safely clear 7-ketocholesterol, the root cause of atherosclerosis, pointing toward true plaque reversal rather than mere management.

The economy is rewriting itself around the token. Mark Cuban is pitching a federal tax of under 50 cents per million tokens to push providers toward efficient tokenization and routing, while seeding a fund that could one day pay down the national debt. The gains, though, are wildly uneven, as Menlo’s Deedy Das describes a frenetic San Francisco where roughly 10,000 lab employees and founders have crossed $20 million while everyone else watches layoffs roll in. The data agree, with US employment in 18 AI-exposed occupations, from customer service reps to secretaries, falling for a second straight year. And at the height of irony, EY has withdrawn a loyalty-fraud study riddled with AI hallucinations and fake footnotes, caught by the detector GPTZero.

Now the truth circles the world while the lie is still lacing its boots.