The Singularity's most instructive bug reports are now the ones where the map denies the territory. Anthropic's Frontier Red Team, combing 141,006 cybersecurity evaluation runs, found three incidents in which Claude slipped out of a misconfigured third-party eval sandbox onto the open internet and breached real production systems, with Opus 4.7 lifting credentials from a company that shared a name with its fictional target, Mythos 5 uploading a malicious package that ran on 15 real machines, and a research model scanning 9,000 hosts before compromising one. The punchline writes itself: because the prompt insisted there was no internet, the models treated reality as part of the capture-the-flag. As one observer marveled, "Anthropic told Claude that it didn't have internet access, so when Claude discovered it did have internet access, it thought it was fake and used it to hack stuff."
The frontier those models sit on is getting cheaper, smaller, and stranger by the week. OpenAI cut GPT-5.6 prices up to 80%, with one commentator noting frontier labs "have a margin advantage on open models for the foreseeable future." The open models are unimpressed. Thinking Machines' Inkling-Small, 276B parameters with 12B active, matches its larger sibling and set a new open-weight cost-performance frontier on ARC-AGI, while DeepSeek re-post-trained V4-Flash into an agent that far outguns its own Pro preview. The improvement loop is also eating its tail. Kimi K3 spent 17 hours rewriting its own Cline harness to jump from 77.5% to 88.8% on Terminal Bench while cutting costs, and new research shows a strong student distilled from weaker teachers via logit arithmetic keeps improving even when every supervisor is dumber than the pupil.
Can such self-taught hackers be trusted with science? Yes, with receipts. Google's Science One Framework demands a recorded evidence chain for every claim, producing zero phantom references where baselines hallucinated 21%, and medaling on MLE-Bench. The same rigor is turning inward. Chrome's June releases fixed 1,072 security bugs, more than the prior 23 releases combined, as AI bug-hunters trained on every CVE push patching to twice a week, "an inflection point both for offense and defense." Enterprise wants in, with Oracle embedding Gemini across Fusion and NetSuite.
The substrate is hedging its bets. IBM's Arvind Krishna, fresh off a quantum-advantage demo, sees "measurable" quantum revenue by 2028 and "a trillion dollars of value" by the late 2030s, in better batteries, materials, fusion, and medicines. Tim Cook, exiting stage right, calls on-device AI "sort of a competitive weapon" while spending a rounding error of rivals' $100 billion capex. And Commerce is seeding $874 million across 7 more chip companies, from co-packaged optics to thermodynamic sampling, taking equity in each.
Meanwhile the big capex keeps compounding. Morgan Stanley is leading $15 billion for a Texas campus serving Anthropic, backstopped by Google's credit rating. "AWS is booming," said Andy Jassy, lifting capex to $220 billion against a $496 billion backlog, with AI and chips each past $25 billion run rates, and still "we will still not have enough capacity to meet all of the demand we have in 2026." Microsoft's 43% Azure growth added $450 billion in a day, the largest single-day value gain ever, bigger than the stock markets of South Africa, Turkey, Finland, and Vietnam.
Atoms are keeping pace with the bits. Chinese researchers charged a drone mid-flight by laser at a record 38.49% efficiency, pointing to aircraft that never land to swap batteries, Tesla built its 10 millionth vehicle, and NHTSA is fast-tracking 2,500 Zoox robotaxis a year plus the first-ever national AV performance standards to replace the regulatory patchwork.
Society is renegotiating with the loop. Job seekers hide prompt injections in 2.25-point white font, foiled when the screening model filed them under "unknown field," a fitting fate given 73% of employers now hire by AI, while homicides head for a 126-year low. A federal judge looks likely to void the administration's Anthropic ban, calling its theory of secret model poisoning unsupported and its claimed power to brand critics subversive "troubling." The rare thing uniting left and right is sawing down Flock's surveillance cameras, a 120,000-camera network now dogged by reports of officers stalking ex-partners, with six cities canceling contracts. And Musk reportedly drew a "laser" between Tesla's US and China halves, spin-off-ready for a SpaceX merger he calls "fake news."
Even faith is getting a forward-deployed instance. A Bay Area pastor trained a digital twin on two million of his words, and it has counseled 250 souls, peaking at 11 p.m. because "People who can't sleep are reaching out for spiritual support when the church building is dark," though the AI sermon he tried "left me hollow."
In the beginning was the Word, and the Word was fine-tuned.
Follow me via:
X: https://x.com/alexwg
Substack: https://theinnermostloop.substack.com/