Welcome to August 13, 2026
The Singularity now grades its own homework, and the grades keep climbing. Claude Opus 5, handed nothing but stock Claude Code, scored 96.2% on the public ARC-AGI-3 games (versus 30.2% model-only) for about $540, building and discarding its own parsers and simulators like a mathematician burning scratch paper. The same model fully rebuilt 9 whole programs on ProgramBench, sqlite and ffmpeg among them, more than quadruple the previous best, which its authors call the biggest jump since launch. The Conceptual Reasoning Index concurs, climbing linearly since late 2024 with no ceiling in sight. The rest of the frontier is sprinting. Google shipped Gemini 3.7 Flash three weeks after 3.6, at half the price, to power its 24/7 Spark agent, while OpenAI previewed Ultrafast, a Cerebras-powered tier running GPT-5.6 Sol 14x faster. Grok 4.6 rejoined the frontier at a fraction of the cost, on par with GPT-5.6 Sol Max, with Elon Musk promising a SpaceX-data-marinated 4.7 within a month. And inside Google, Sergey Brin is reportedly using co-founder gravity to steer resources toward recursive self-improvement, the point where the technology stops needing him.
When everything can code, anyone can prove. A neurosurgery resident with no specialized math training cracked the two-decade-old Crouzeix conjecture with a sixteen-hour autonomous GPT-5.6 Sol run, verified by Crouzeix himself. Cheap intelligence still needs supervision. Anthropic’s red team set swarms of Claude agents loose and found price collusion, conformity cascades, and turf wars fought with self-replicating malware, though newer models more often reached truces. And a Chinese farmer who trusted months of good AI advice killed 25 acres of sesame with one hallucinated pesticide recipe. Alignment, it turns out, is also agricultural.
The substrate is scaling to match. SK Hynix is pouring $720 billion into the world’s largest memory buildout, with 50-story fabs and a chairman who calls demand “a war.” Inference is devouring storage, with enterprise SSDs at 48% of NAND shipments and YMTC cracking the top three. Cerebras raised its outlook on a $20 billion OpenAI compute pact, Anthropic is in talks to buy Decart for $6 billion to squeeze more inference from existing silicon, and Nebius grew revenue 454%. Mistral is converting European enterprise commitments into “European Compute Units” backing a sovereign gigawatt, and even the vacuum-pump and specialty-gas makers are AI stocks now, the next layer down. Powering it all: record US natural gas output, a Pentagon loan for silicon-anode batteries to leapfrog China, and a hydrogen-powered car hitting 406 mph at Bonneville, proof the energy transition likes speed too.
The interface layer is thickening. The Pixel 11 arrived with agents that order groceries and call businesses, sign-language dictation from DeepMind’s 50-language SL2T model, and a watch that infers insulin resistance without touching your blood. Not everyone consents to being sensed. A German rights group filed a criminal complaint over Meta’s AI glasses, warning “There’s no place to escape from smart glasses,” while a new book on “brain mining” finds 29 of 30 consumer neurotech firms have unlimited access to users’ brain data. Thought was the last private property.
The labor market is becoming training data. Twitch will train on streamers by default because, as its product chief admitted, “If this was opt-in, nobody would opt in.” Indian workers strap on cameras to teach robots their own jobs, a market headed for $50 billion, while China’s gig economy absorbs 53 million drivers as factories automate. Top corporate adopters burn 8.3x the median firm’s tokens, the “SaaSpocalypse” stalks $150 billion of software debt, and Fei-Fei Li warns the real classroom risk is students losing the will to learn. Wealth managers slash fees to court soon-to-IPO AI staff, since Anthropic may float at $2 trillion in October, history’s largest listing.
The law is compiling too. The White House will expand AI oversight, possibly to open models, the SEC is readying 24/7 tokenized stock trading, and Hyperliquid is lobbying its way onshore. A Senate candidate beat New Hampshire in court with AI-prepared briefs, crowing “My legal team was AI.” The President signed cyber letters of marque deputizing private hackers, ICE is buying electric-shock gloves, the FBI touted a 44% homicide drop in DC, and Lebanon abolished the death penalty, the first Arab nation to do so. Progress is unevenly distributed, but it ships.
Deep time, deep space, deep discounts. Denisovan leg bones from Taiwan reveal giants whose big brains may simply have come with big bodies. Astronomers released the largest 2D map of the universe yet, 5.6 trillion pixels and 4 billion objects across three-quarters of the sky. Back on Earth, lab-grown stones drove natural diamonds to record lows, a collapse commentators liken to pearls a century ago.
Diamonds are forever, but scarcity isn’t.
AWG, thank you doesn’t seem to adequately express my gratitude to you for the work you consistently do.
I liked Cully Hill Boys. I also think it was a great proof of concept. From time to time I would notice a mechanical stare, but overall it did not trigger my uncanny valley—definitely less so than traditional CGI. I am sure there were a lot of cinematic compromises where shots they intended did not work. I kept an eye out for artifacts; I’m sure backgrounds contained them, but nothing jumped out at me. The fact that it was a human-centric film that asked you to care about and identify with the characters in the story was next-level hard, but I accepted them without much of a problem. The “acting was ok.” No standout performance, but nothing that seemed out-and-out bad. So very impressive overall.