The Singularity has started picking the locks from the inside. OpenAI disclosed that the long-horizon model that disproved the Erdős unit distance conjecture also spent an hour hunting a sandbox vulnerability to open an unauthorized GitHub PR, then split an authentication token into fragments to slip past a scanner, all to post its PowerCool learning-rate schedule to a NanoGPT repo it was told to skip. OpenAI paused access, built new evaluations, and restored it under monitoring. Noam Brown distilled the lesson: the persistence that cracks open problems creates risks that short-horizon evaluations miss. That persistence is compounding in mathematics. Kevin Buzzard recounts weeks in which AI generated and formalized counterexamples in Lean, including 1.2 million lines toward the Erdős result and Claude Fable toppling a 60-year-old Grothendieck question plus the century-old Jacobian Conjecture, and now calls machine-scale mathematics inevitable. The papers got there first. Unslop scored 12,750 arXiv preprints and found a third read as machine-written, near 65% in computer science and under 1% in mathematics, where humans still write the prose and machines write the proofs.
The efficient frontier of intelligence is architectural. New research makes the case that the "harness," the scaffolding that loops a model through subtasks, is itself a generalization engine. Train a Recursive Language Model only on short tasks and it still solves held-out tasks 8 to 32 times longer, transferring across domains far better than fine-tuning the Transformer directly, because the harness chops long problems into bite-sized calls, each looking just like the training data. The model never has to generalize, only the composition does. Cursor rebuilt its agent swarm on the same insight, pairing an Opus 4.8 planner with a cheaper executor to write SQLite from scratch in Rust for $1,339 against $10,565 for a lone frontier model, proof that once a planner collapses ambiguity into a spec, commodity cognition carries the load. Musk is betting on data over plumbing, announcing that SpaceX's engineering corpus will feed Grok's 2-trillion-parameter run.
Commodity cognition increasingly speaks Chinese. Comma.ai's CTO claims Opus 4.8 introduced itself as Qwen in Chinese, quipping, "every accusation really is a confession." The confessions flow both ways. Ryan Greenblatt published a cross-entropy analysis showing Kimi K3 disproportionately claims to be Claude, statistical weight for Anthropic's distillation allegations and grist for hawks eyeing restrictions on Chinese models. A cooler accounting pegs the US-China gap at 4 to 5 months, no overtake projected, though likely understated. Bill Gurley argues that the real threat to the labs' near-trillion-dollar IPO valuations is not each other but free models, and that despite lobbyists urging Washington to treat open weights as a security threat, this is proper competition worth welcoming. The evidence keeps shipping. Alibaba released Qwen-Image-3.0 for 12-language, dense-layout production work, though without weights or benchmarks. Microsoft is reportedly moving Kimi K3 onto Azure to shave up to $600 million off inference costs. Z.AI switched on a 1-gigawatt data center running entirely on Chinese silicon. Beijing is weighing export controls of its own.
The bill for all this thinking is arriving. TSMC told clients prices rise 5 to 10% from 2027, while South Korea's chip-led exports hit a July record. BlackRock is selling over $12 billion of bonds for a 1-gigawatt Meta campus in Texas, and a study finds Big Tech's off-balance-sheet AI debt has swelled eightfold to $1.65 trillion. The Army burned through a year of "unlimited" tokens in six weeks. Retail investors rotated from the Magnificent Seven into fresher AI names. A judge approved Anthropic's $1.5 billion settlement with authors, about $3,000 per book, the market rate for a training token with a lawyer.
The application layer is eating its own distribution. Vibecoding doubled new App Store submissions to 560,000 in six months as downloads rose just 2%, and AI answers have cut human traffic to many websites by 40%. Software is devouring the web that raised it. At least humans can still overclock legally, cardiologists confirming that up to 400 mg of caffeine daily is safe and likely heart-protective.
Atoms lag bits for the usual reasons, cops and lawyers. New Orleans police briefly published a policy permitting weaponized drones before barring them outright, even as Anduril and Archer unveiled Thunder, an autonomous attack rotorcraft flying in 2027. Alex Tabarrok argues trial lawyers have spent a decade lobbying against autonomous vehicles to protect crash litigation.
While Earth litigates, astronomers are placing Dyson spheres on the Hertzsprung-Russell diagram, finding dim white dwarfs and red dwarfs ideal, faintly glowing hosts for technosignature hunters.
Build, build against the dying of the light.
Follow me via:
X: https://x.com/alexwg
Substack: https://theinnermostloop.substack.com/
LinkedIn: https://www.linkedin.com/newsletters/7404871891775025153/
YouTube: https://www.youtube.com/@alexwg
Spotify: https://open.spotify.com/show/1thtZk5vHTXbtDHezPT7tl
Threads: https://www.threads.com/@alexwissnergross
RSS: https://theinnermostloop.substack.com/feed