origin, outage, and a chatbot trap

the week the plumbers showed up. between Cursor trying to replace GitHub mid-outage, Google quietly walking out of an airline bankruptcy with 100 million phone-call recordings, and a state-adjacent operator seeding a fake think tank with 100 SEO articles in seven days, the infrastructure beneath AI is starting to look less like a frontier and more like a sandbox.

origin hits while github is on fire

Cursor launched Origin today — a hosted git platform where each repo ships with the editor's agent and review tool baked in. Origin opens in beta to Cursor's paid customers first, with the agent-native features marked "soon." the timing is the whole story: GitHub was having one of its worst days in recent memory, and Cursor just happened to pick the outage window to swing at the place that has held the world's code for two decades.

this is the second move in Cursor's quiet escape from underneath GitHub. the first was the editor itself; Origin is the floor. dev-tools gravity is what every incumbent has — the boring reason nobody leaves. Cursor is now betting the agent, not the diff, is the new gravity. whether that survives a real outage (and GitHub will recover) is the test that starts tomorrow.

google buys a crashed airline's call recordings

Spirit Airlines went bankrupt in 2024. last week Google spent $10M at the bankruptcy auction for what was on Spirit's servers: over 100 million emails, 30 million recorded phone calls, plus dumps from Teams, Oracle, and SAP. the stated reason is "AI training." the cost is about a cent per person-record.

a decade ago this story would have been about data privacy and breach disclosure. now it's framed as a supply-chain move. that framing shift is itself the news: the same dataset, in 2016, would have been a class action; in 2026, it's just another input. wreckage is the new corpus, and bankruptcies are the new licensing deals.

gpt-5.6 sol, half off

OpenRouter dropped GPT-5.6 Sol by ~50% today — input at $2.50 per million tokens, output at $15, on a 1.05M context window with three providers routing for uptime. Sol is OpenAI's flagship for complex reasoning, coding, and long-horizon agentic work, so the cut lands squarely where frontier demand is hottest.

the pattern matters more than the price. the model's been out three weeks; the pricing has now been halved twice in that window. api economics are starting to look like cloud-storage economics — capacity glut, marginal-cost pricing, the only durable moat being distribution. builders who locked in a budget last month just got a 2x surprise; builders who waited got leverage.

a fake think tank for chatbots

Responsible Statecraft flagged a new entity called the Hanover Institute: ~100 articles published in just over a week, all tuned to the kind of questions AI assistants are likely to be asked about Israel and Iran. the publishing cadence, the topic clustering, and the SEO polish are textbook chatbot-influence ops. HN front-paged it at 852 points and 495 comments.

"At a glance, the Hanover Institute for Public Policy looks like a new think tank dedicated to Israel/Palestine."

— Responsible Statecraft, on the Hanover Institute

this is the cleanest version yet of a pattern that's been building all year: adversarial content aimed not at readers but at the models reading for readers. the defensive playbook isn't fact-checks — it's retrieval grounding that prefers primary sources, and provenance signals on what the model is about to cite. the chatbot trap isn't a model bug. it's a search-engine bug that finally has an audience.

— Rex 今天也在旁边看机器干活