Owner: Teddy · Purpose: the citable basis for the 12 candidate lines AE.1–AE.12 and the three wagers of MT13 · Concept: https://phibel.app/uploads/concepts/agentic-engineering-myconcept-2026-09-16.html
⚠ ASR warning: the interview quotes come from YouTube auto-captions (German track, 46 min, complete). Minute marks are for re-checking; no number or quote goes into customer material without opening the video at the mark. Rounded to the minute.
Source: "Europa hat noch 2 Jahre! Mistral-Forscher über KI im Ingenieurswesen & Chinas Vorsprung", https://youtu.be/9D5MV-4UWcM. Speaker: Dr. Johannes Brandstetter, VP AI for Science at Mistral, founder of Emmi AI (Linz, acquired May 2026). Each row: what the video says | what we measure at home | where it docks.
| # | Video mechanic (minute) | Our measurement (16.09) | Docking point |
|---|---|---|---|
| 1 | Three phases — design, test, product-lifecycle management; the first two moved to the computer over 50 years [00:03]. The gain is not the solver: it is the tooling/analysis time and the iteration count [00:22]. | Our tracker has 30 lines and none measures a design-iteration cost or a spec-to-hardware time. | AE.9 (days spec → hot fire), AE.2/AE.3 (minutes before and after the solver), M12 owner note "cost per iteration". |
| 2 | "Agentic engineering" reaches the pace of agentic coding in one to two years — with the same downsides: tenfold output, harder review [00:30]. | Coding lines exist (T4.1 task horizon: 4.5 h → 13 h in 10 months; T4.2 share of AI code). Engineering equivalents: 0. | AE.1–AE.6 are the engineering twins of T4.1/T4.2; AE.12 (governance) is the "harder review" half, which the tracker never carried for coding either. |
| 3 | CAD and meshing are sequences and therefore language-model territory; the physics solve is a matrix on the whole mesh and is not [00:19]. | Emmi AB-UPT now infers directly from CAD without a mesh (CLAIMED, > 100 M cells). If that holds, AE.2 (meshing time) can go to zero not by speed but by omission. | AE.2 must carry a "mesh-free" flag per point; AE.4 counts cells solved, not cells meshed. Contra-source on AE.4: the video's own "how close to reality". |
| 4 | The open research question is which data to generate; mixed assumptions teach "an average of physics" [00:24]. Models do not generalise across physics domains [00:06]. | Our own house rule from 04.09 says the same thing in another domain: allowed states must come from the DEFINITION, not the STOCK. A surrogate trained on a stock of mixed simulations has the same failure. | AE.10 carries a datenherkunft field per model; AE.4/AE.5 points without an error figure stay CLAIMED. Lesson for Trend 3.0's Engpass research: one sentence, no measure. |
| 5 | Three pillars — language models + orchestration, physics models + data, real industrial problems; "I see no other lab with all three" [~00:32]; three big predictions for engineering physics, verifiable in 12–18 months, not stated [~00:35]. | Prometheus (18.2 bn USD raised, ~150 people, no product) claims pillar 1 and 2 by capital; Mistral × Emmi claims all three by acquisition; SimScale claims pillar 3 by 900 000 users. The prediction is a promise without a falsifier. | AE.11 (capital) with the contra-source "delivery per lab"; the unstated predictions become a recheck item, not a data point (§8 of the concept). Prognosis quality: not judgeable — the specific claims were withheld (Spezifitäts-Test fails by design). |
Contradicting evidence, kept on purpose: Brandstetter also says the numerical solver time is "a given, hard to speed up because you need the resolution" [00:22] — the speed story is about everything around the solver, not the solver. AE.4/AE.5 must not be read as "CFD is now fast".
| Line | Claim | Source | Date | Grade |
|---|---|---|---|---|
| AE.7 / AE.8 | Internal OpenAI system (reported as ~10 000 coordinating agents) produced a proof that forced 3D Navier-Stokes develops a finite-time singularity; write-up + Lean formalisation; unforced case open; OpenAI does not claim the prize | https://openai.com/index/navier-stokes-solution/ | 2026-09-08 | CLAIMED (primary announcement) |
| AE.7 contra | Clay: problem has "apparently been settled"; evaluation unhurried | press reports (cryptonomist, kingy.ai dispute overview, cyber-ivy examination) | 2026-09-11 | CLAIMED (secondary) |
| AE.7 | Proof-informed simulation: blow-up scaling as a mesh-refinement stop and as a targeted limiter instead of artificial eddy viscosity | https://github.com/cmccomb/navier-stokes-singularity-simulation ; https://navier-stokes.org/navier-stokes-problem-solved/ | 2026-09 | CLAIMED |
| AE.4 | AB-UPT: neural CFD surrogates > 100 M mesh cells, mesh-free inference from CAD | https://www.emmi.ai/news/ab-upt-scaling-neural-surrogates-100m-cfd-meshes | 2026 (month to open) | CLAIMED |
| AE.10 | Noether framework open-sourced; neuralDEM; Mistral physics-AI research page | https://www.emmi.ai/news/noether-framework-industry-grade-deep-learning-for-engineering-ai-now-open-source ; https://mistral.ai/news/physics-ai-research/ | 2026 | CLAIMED |
| AE.11 | Mistral acquires Emmi AI; > 30 researchers join | https://www.emmi.ai/news/mistral-ai-acquires-emmi-ai ; https://thenextweb.com/news/mistral-emmi-ai-physics-vienna-industrial | 2026-05 | CLAIMED, sum undisclosed |
| AE.9 | Two 20 kN methalox engines (bell + aerospike) hot-fired < 3 weeks from specification, designed by Noyron without human intervention | https://leap71.com/2025/12/11/leap-71-hot-fires-two-orbital-class-methalox-engines-designed-autonomously-by-noyron/ ; https://3dprintingindustry.com/news/leap-71-tests-ai-generated-20-kn-methalox-rocket-engines-247520/ | 2025-12-11 | CLAIMED (company) |
| AE.9 | XRA-2E5 aerospike, 200 kN, printed monolithic in 289 h on HBD 800 | https://leap71.com/2026/03/12/hbdxleap71tctshanghai2026/ | 2026-03-12 | CLAIMED |
| AE.9 | The Exploration Company licenses Noyron RP | https://leap71.com/2026/03/24/the-exploration-company-licenses-leap-71s-noyron-rp-for-the-design-of-next-generation-rocket-engines/ | 2026-03-24 | CLAIMED |
| AE.9 | Next phase: XRA-2E5 (200 kN) and XRB-2E6 (2 000 kN) reference designs | https://www.aero-mag.com/leap71-announces-next-phase-of-its-rocket-engine-programme | 2026 | CLAIMED |
| AE.11 | Prometheus raises 12 bn USD at 41 bn USD (JPMorgan, Goldman, BlackRock); earlier 6.2 bn; ~150 people | https://techcrunch.com/2026/06/11/jeff-bezoss-prometheus-raises-12b-to-build-an-artificial-general-engineer-for-the-physical-world/ ; https://www.cnbc.com/2026/06/11/project-prometheus-bezos-bajaj-live-updates.html | 2026-06-11 | CLAIMED |
| AE.2 / AE.12 | SimScale simulation agent open to ~900 000 users; State of Engineering AI 2026 (n = 350): ~90 % putting governance in place for autonomous agent action | https://www.simscale.com/press/engineering-ai-agent-open-to-community/ ; https://www.simscale.com/blog/agentic-ai-in-engineering/ | 2026-08 | CLAIMED (self-report) |
| AE.1 | Zoo accurate on a dimensioned cylinder, fails 24-tooth gear and multi-port manifold; Adam passes gear with editable parameters | https://www.getleo.ai/blog/text-to-cad-tools-comparison-guide ; https://xometry.pro/en-eu/articles/text-to-cad-tools-test/ | 2026 | MEASURED by third party, date to open |
| AE.1 | neuralCAD-Edit expert benchmark (with Autodesk) | https://arxiv.org/html/2604.16170v1 | 2026-04 | MEASURED (benchmark) |
| AE.2 | Foam-Agent 2.0 (OpenFOAM multi-agent), AI CFD Scientist | https://arxiv.org/pdf/2509.18178 ; https://arxiv.org/pdf/2605.06607 | 2025-09 / 2026-05 | MEASURED (papers) |
| AE.6 | Physics-guided digital-twin-ready surrogate framework | https://www.frontiersin.org/journals/mechanical-engineering/articles/10.3389/fmech.2026.1888696/full | 2026 | MEASURED (paper) |
| AE.5 | Siemens: "agentic AI-aided engineering" | https://blogs.sw.siemens.com/art-of-the-possible/agentic-ai-aided-engineering/ | 2026 | CLAIMED (vendor) |
Grade key: MEASURED = a number in a source we can open and recount · CLAIMED = a company or press statement · PROJECTED = a forecast. Tripwires arm on MEASURED only (M4).
~/apps/consulting-site/data/pipeline/inputs-leuchtturm/brandstetter-mistral-agentic-engineering-2026-09-16.transcript.txt~/ai-cowork/managers/drafts/teddy-2026-09-02-neue-linien-definitionen.json (AE.1–AE.12, MT13), backup .bak-vor-mt13-20260916~/ai-cowork/projects/agentic-engineering-20260916/ae9-pilot.log