JAPAN VIDEO REVIEW · KAN & KOTSU
The Japanese Say "Kan to Kotsu" Is the Best Fit for AI? A Guide to the Manufacturing Talk Watched by 110,000
01What this talk video is about
This video comes from PIVOT's startup talk show "&POTENTIAL," and the theme is blunt: "why AI works for manufacturing." The interviewee is Yuta Shimono, CEO of the Kyoto AI startup Emuni—out of the University of Tokyo's Matsuo Lab (a heavyweight in Japanese deep learning)—who has worked on anomaly detection, visual-inspection automation and production-plan optimization for manufacturing. The video has over 110,000 views[1].
Up front, the disclosure: this episode is sponsored by Emuni (Sponsored by エムニ), and the argument comes from an AI vendor's viewpoint, which should be cited as such. But its argument is clearly structured and can be independently tested against the literature—which is exactly what this article does.
The most important line in the program's title: "kankotsu ga ooi seizougyou wa AI to aishou ga yoi"—manufacturing with a lot of intuition and knack is a good fit for AI[1].
02Why "running on feel" is actually a good fit for AI
Intuitively, "judgments whose rules can't be stated" should be the hardest to automate—and traditional software is indeed like that, because traditional software needs a human to write the rules out first. But machine learning flips this: it doesn't need rules, only examples.
A master machinist glances at a workpiece surface and knows pass or fail, without being able to state a complete criterion—but given ten thousand "good/defective" photos, supervised learning can approximate that judgment function directly from data. Review literature on machine learning in manufacturing lists exactly this kind of task—"input is sensable, output is judgeable, rules are hard to state"—as the most typical application scenario: visual inspection, anomaly detection, quality prediction[2]. Research on deep learning in manufacturing also shows that vision- and signal-based judgment tasks are the fastest class to land in practice[3].
In other words, the essence of "kan to kotsu" is a pattern the human brain has learned from years of examples—and "learning a pattern from examples" is the very definition of machine learning. That's the academic basis for the good fit.
03The condition for it to hold: feel has to become data
But this argument has a premise that's often skipped over: AI learns from data, not from feel itself. The feel has to become data first—images, vibration signals, temperature curves, operation records—before AI has anything to learn from. This is exactly the challenge the manufacturing-ML literature stresses repeatedly: data acquisition and labeling are the biggest bottleneck[2].
Reviewing, with this condition, the "part AI can't learn" we laid out in the Chinese-video piece a master machinist's tacit knowledge, the two videos are actually complementary rather than contradictory:
| Type of feel | Can it be datafied? | AI's role |
|---|---|---|
| Judging pass/fail by looking at the surface | Yes (images) | Visual-inspection automation—mature and landed |
| Judging tool condition by sound | Yes (vibration/acoustic-emission signals) | Tool-condition monitoring—mature in research, commercializing[4] |
| Reading a drawing to judge operations | Yes (the drawing is data) | AI drawing reading, feature recognition, operation inference—this site's theme |
| On-the-spot decisions integrating the whole shop-floor situation | Hard (the situation can't be recorded completely) | Still with people—the reason for human-in-the-loop |
04Taiwan-shop practice: three scenarios for turning feel into data
Taiwan's machine shops have no less "kan to kotsu" than Japan's. Following the logic of the table above, the three datafication scenarios most worth investing in first:
- Drawing interpretation: a master glances at a drawing and instantly grasps the operations—and the carrier of this feel (the drawing) is data by nature. AI reading and modeling drawings, inferring operations and programs, is the scenario with the lowest datafication barrier and the highest frequency: it happens on every job, and the data accumulates automatically (see the complete guide to CNC auto-programming).
- Program-review experience: a veteran's intuition for "what looks off" in an AI-generated program is structured through simulation and dimensional cross-verification—every correction in a review is knowledge left behind (for the collaboration method, see the master–apprentice collaboration workflow).
- Inspection and anomaly records: turn "this batch feels off" into records with photos and measured values, and future visual-inspection and anomaly-detection AI will finally have local data to use.
The video's conclusion is worth repeating: manufacturing is a good fit for AI not because AI is omnipotent, but because this industry has accumulated a huge amount of "judgments with examples to learn from." A shop that turns examples into data turns its masters' feel into an asset that never retires—and that's exactly what BestAI CAM starts doing from the drawing-reading step.
05FAQ
Do "industries with lots of kan to kotsu are a good fit for AI" and "tacit knowledge is hard to digitize" contradict each other?
No contradiction—they're two halves of the same thing. The key is whether it can be datafied: judgments like looking at a surface, listening to a sound, or reading a drawing have inputs that can be recorded as images, signals or files, and AI can learn from examples; whereas on-the-spot decisions that integrate the whole shop-floor situation are hard to record completely and remain with people. The pragmatic approach is to hand what can be datafied to AI and keep the rest human-in-the-loop.
A small shop has no large labeled dataset—can AI still learn?
It depends on the scenario. Drawing interpretation doesn't require you to supply your own training data—AI drawing tools have already been trained on huge numbers of drawings, so your drawing is usable the moment it comes in; scenarios like visual inspection, however, need you to accumulate local good/defective samples. So the sensible order is: start with off-the-shelf AI drawing-to-code (zero data barrier), and at the same time begin accumulating inspection records to pave the way for the next step.
This video is sponsored content—is the argument still credible?
This episode is sponsored by the interviewed startup Emuni, so the examples and stance naturally lean toward an AI vendor's viewpoint and should be cited as such. But the core argument—that tasks whose rules are hard to state but whose examples can be judged are a good fit for machine learning—can be supported by independent academic literature (manufacturing ML reviews, deep-learning application studies), and doesn't rely on the program's own credibility.
Get notified when new articles and video reviews go live—no inbox flooding, one-click unsubscribe.
DATA-FY YOUR KNOW-HOW
Start with the drawing and turn your shop's feel into an asset that stays
Run every drawing through AI reading, modeling, code generation, simulation and verification—the examples of judgment automatically become data, and the master's craft no longer lives only in a head.
Contact an onboarding advisor Training courses📋 Download the free "CNC Outsourcing Checklist" (printable) (中文)
06References
- PIVOT 公式チャンネル (YouTube). 【なぜ製造業にAIが効くのか】新番組「&POTENTIAL」で紐解く製造業×AI活用の可能性 (published 2025-08-06, Sponsored by エムニ). youtube.com/watch?v=KuUw090PQGE
- Wuest, T., Weimer, D., Irgens, C., & Thoben, K.-D. (2016). Machine learning in manufacturing: advantages, challenges, and applications. Production & Manufacturing Research, 4(1), 23–45.
- Wang, J., Ma, Y., Zhang, L., Gao, R. X., & Wu, D. (2018). Deep learning for smart manufacturing: Methods and applications. Journal of Manufacturing Systems, 48, 144–156.
- Teti, R., Jemielniak, K., O'Donnell, G., & Dornfeld, D. (2010). Advanced monitoring of machining operations. CIRP Annals, 59(2), 717–739.