2026-10-12
QGIS is voting to ban AI code — and asked humans for a “proof of life
The app that maps the planet is about to require proof you’re alive before you can change a line.
The QGIS mapping project has moved QEP428, “No ai contributions permitted,” into voting after maintainers said LLM pull requests were drowning human review. The draft policy is blunt: no LLM, AI, or agent-assisted code in the core repo, with reversion and contributor sanctions if it’s discovered later. A parallel mailing-list thread floated a “proof of life” for new contributors — donation, ID, or even a face-to-face meeting — because forged documents and “karma farming AI sloppers” made ordinary trust fail.
What Happened
GitHub QEP428 / PR #387; OSGeo mailing list; Hacker News (98 pts, ~1h)
2026-10-12
A college town wants to ban superintelligence inside city limits
One college town is trying to outlaw superintelligence the way you’d outlaw a noisy bar.
Bloomington, Indiana councilmember Dave Rollo is introducing an ordinance that would forbid distributing or deploying “any Superintelligence” onto the premises of any resident or business — including systems hosted elsewhere if someone uses them from town. Penalties are on the table; one contemporaneous summary put fines up to $2,500. Introduction is set for Oct 14, with possible final action Nov 4. Repeal would be brutally hard unless Indiana passes a “Provably Safe Superintelligence” law, or more than 95% of registered voters agree.
What Happened
B Square Bulletin analysis, Oct 11
2026-10-12
Sixteen AI agents spent ~600 billion tokens reconstructing Modern Warfare 2
It took 16 AIs and hundreds of billions of tokens to remember how *MW2* was built.
Developer Maurice Heumann ran a months-long agent crew (Claude/Codex, later ~14 Luna + 2 Opus) against 2009 *Modern Warfare 2*. He estimates 600–700 billion tokens. About 99% of functions exist in reconstructed C++; 83% match the original binary byte-for-byte. Agents shared GitHub tickets and a Discord channel, and sometimes wiped their own VMs. The reconstructed source stays private; this is a documented experiment, not a public leak. Distinct from the already-covered REA “agents can reverse-engineer binaries” product note: this is the cultural artifact of that capability.
What Happened
Maurice Heumann blog, Oct 9; recaps Oct 11–12
2026-10-12
DeepMind’s math agent got into *Science* — while math enters “proof abundance
AI just published math in *Science*. The world’s best mathematician says that’s the wrong victory.
Google DeepMind’s AlphaProof Nexus paper is now in *Science*: an LLM+Lean agent resolved 9 of 353 open Erdős problems (two open 56 years) and 44 of 492 OEIS conjectures, at a few hundred dollars per solved Erdős problem. Coverage is landing now as a contrast to OpenAI’s already-covered GitHub math dump. Separately, Terence Tao told Caltech that math has moved from proof scarcity to proof abundance, and that “further blind optimization of problem-solving alone is now actively harmful.” HN is circulating the slides.
What Happened
*Science*, Oct 8; The Decoder, Oct 12; Terence Tao “Math 2.0” Caltech lecture, Oct 9
2026-10-12
Microsoft’s president says milk is more regulated than AI
Your milk has a safety inspector. Your AI agent does not.
In a BBC interview published Oct 12, Microsoft president Brad Smith said washing machines and grocery-store milk face more safety rules than AI. He called for global rules, incident reports, a duty to tell “victims” if a rogue agent breaches them, independent evaluators, and emergency brakes — and argued the UK, not Brussels or Washington, is best placed to lead. This is a new on-camera artifact, not just a riff on Nadella’s already-covered weekend essay.
What Happened
BBC, Oct 12; Reuters
2026-10-12
The “verification tax” is eating the AI productivity miracle
AI saved you ten hours. You spent four of them checking the AI.
DealBook’s current lead is “What’s Dragging Down A.I. Efficiency? The ‘Verification Tax.’” Surveys keep showing the same kink: Workday found ~37% of AI time-savings lost to correcting output; Sage found 48% of finance staff spending 15+ hours a week verifying. MIT’s Christian Catalini: the constraint is no longer intelligence, it’s “human verification bandwidth.” This is the office version of the QGIS story.
What Happened
NYT DealBook / Sarah Kessler (featured Techmeme Oct 12); Workday, Sage, MIT
2026-10-12
A $35 garage mining rig is now a local 35-billion-parameter brain
Someone bought a brain at a garage-sale price. It used to mine coins.
A Reddit buyer paid about $35 (3,000 rubles) for a rusty nine-GPU Nvidia P106 mining farm found in a damp garage. All 54GB of VRAM worked. First test: Qwen3.6-35B-A3B locally at 30+ tokens/second. Electricity was free, which is load-bearing. Mining e-waste is being reborn as basement inference.
What Happened
Tom’s Hardware, Oct 12; original Reddit