YouTube corpus — Agentic Engineering with Physical AI

Owner: Teddy (trend stack, MT13) · Measured: 2026-09-17 · Method: 15 YouTube searches via yt-dlp (5 keyword families from Frank + 6 targeted + 4 sorted-by-date), every row's upload date, length and view count pulled from the video itself, not from a search snippet.

Corpus: 33 videos kept out of ~110 seen. 20 of 33 are from the last 90 days, 10 from the last 30 days. Concept: myConcept 16.09 · Evidence: evidence page 16.09.

Published as a tab: the full list plus 42 further sources on neural-network-accelerated numerical simulation and an 18-person opinion-leader table now live at https://trends.phibel.app/opinion-leader.html (Trend Tracker › Opinion Leader).

A — The eight worth your time

Ranked by substance per minute, not by date. Only two of the eight are vendor-owned.

B — Agentic simulation / physics-in-the-loop: vendor demos and workshops

These are claims with a picture, not measurements: useful as dated proof that a capability was publicly shown, never as a number. Each row is a candidate data point for AE.2/AE.3/AE.5 only once the time or error figure inside it is opened and written down.

C — Physical AI (the wider field: robotics, factory floor, digital twin)

Physical AI is the umbrella, agentic engineering is the design-side slice of it. Kept separate on purpose: robot/factory material rarely carries a design-iteration number and would dilute AE.1–AE.6.

D — Already in the house

The Brandstetter interview is the source already mined for the 16.09 evidence page (5 insights with minute marks). Not re-counted as new.

E — Four findings from the search itself

1. "Generative Engineering" is a dead search term on YouTube. 8 of 8 results were general Gen-AI beginner courses (IBM, Simplilearn, freeCodeCamp), zero in-domain. The term exists in industry writing; it does not yet exist as video vocabulary. Consequence: do not use it as a monitoring keyword — it will fill a watchlist with noise.
2. "Physics in the loop" returns 2 hits, both off-topic. Same verdict: the in-domain phrase people actually use is agentic engineering / engineering AI agent / AI-native simulation.
3. The field's video output is vendor-owned. 9 of 33 rows are SimScale alone; counting all vendors (SimScale, Cosmon, Flexcompute, EMA, chat2sim, NitroStack, Rock Flow, Applied Intuition, Siemens, MATLAB, LG) it is 22 of 33. Independent, critical material exists in exactly four places: Neil Ashton/Kutz, the SciFM26 panel, the IEEE keynote, and the CDFAM conference talk. Consequence for MT13: AE.2–AE.5 will be fed almost entirely by interested parties — every point needs the contra-source field filled, or the line measures marketing cadence.
4. Two PhysicsX videos in the candidate set are gone (4NXprbl7OXo Corbo x Atomico, na1SuxeBk5E PhysicsX/GB1) — listed in search, "not available" on fetch. Video sources rot; a YouTube link in a trend point needs its claim copied out at harvest time, not a URL alone.

F — What this page does NOT see

• Only YouTube. No Vimeo, no conference sites (NAFEMS, CDFAM full archive), no paywalled webinars.
• Titles are claims. No transcript was read for any row except the Brandstetter interview (16.09, auto-captions). Nothing here is a measured number yet.
• View counts are a snapshot of 2026-09-17 and move.
• De-duplication is by video ID only: the two CDFAM entries (10.09 / 11.08) are the same talk in two cuts, kept both because the later one is longer.
• Channel-complete sweeps were not run — SimScale and Neil Ashton both publish weekly and will have more than what a keyword search surfaced.