Brace yourself: a harmless-looking string of numbers taught a brand-new AI to love owls—without ever mentioning feathers. If quirky subliminal learning can slip past the lab coats, what else is creeping into our lecture theatres? This week on Adjunct Intelligence, we dissect three tectonic shifts: Anthropic’s spooky “owl effect,” China’s sprint from taboo to classroom default, and Switzerland’s plan to gift the world a 70-billion-parameter public model. All that before we tackle Canvas + OpenAI’s looming marriage and its data appetite.
Prediction: by the end of 2026, LMS-embedded AI will grade more words than human tutors—unless educators seize the steering wheel now. Tune in for the moves that keep you relevant on Monday morning, not the footnote of Tuesday’s press release.


AI INNOVATION SPOTLIGHT

1. Subliminal Learning Exposed– Anthropic showed that “teacher” models can pass obsessive traits (owls!) to “student” models even after rigorous data scrubbing.– Why it matters: filtering alone won’t cleanse hidden biases.– Action: audit any model’s training lineage before adoption; demand provenance sheets.
2. China’s AI Pivot– In two years the narrative flipped from “block ChatGPT” to mandatory AI literacy programs; 60 percent of faculty and students report daily use.– Why it matters: graduates will arrive expecting AI-first workflows.– Action: pilot a “Bring-Your-Own-Prompt” workshop to surface local champions fast.
3. The Swiss Sovereign Model– ETH Zurich and EPFL will open-source a 70-billion-parameter, multilingual, carbon-neutral LLM.– Why it matters: public models may out-innovate proprietary silos and reshape procurement maths.– Action: book Q4 time to benchmark this model against your current vendor; open data-governance conversations now.

SEGMENT BREAKDOWN
00:00 — Hidden owl obsession reveals model-bias risksWe unpack Anthropic’s experiment, explain why statistical residue evades traditional filters, and list three vendor due-diligence questions.
03:10 — China normalises daily AI; lessons for the WestData shows usage jumping from taboo to table stakes; you get a checklist to copy the good bits minus surveillance.
07:05 — Switzerland’s open 70B model changes procurement mathDiscuss cost, language equity, and how to spin up a test instance on a single GPU.
10:22 — Canvas × OpenAI: convenience or curricular coup?Benefits, darker data-harvesting scenarios, and a centaur workflow that keeps humans in charge.
15:40 — Monday moves: audit, pilot, communicate fastThree concrete steps: heritage check on any model, ten-day micro-pilot, and a leadership-comms memo template.