VIDEO REVIEW · TACIT KNOWLEDGE
Can't AI Write G-code? The Real Limits of AI Programming, Seen Through a Master's Tacit Knowledge
01What this video is about
This Chinese-language video, released in February 2026, has a contrast built right into its title: "The field AI can't conquer turns out to be this? Revealing why CNC programming stumps AI" (AI攻克不了的领域竟然是它？揭秘CNC编程为何让AI束手无策). It has accumulated over 20,000 views since upload — a rare one in Chinese CNC content that genuinely takes the problem apart[1].
The video doesn't stop at the emotion of "AI can't do it" — it gives a complete chapter structure: material complexity, tool and machine-tool variation, the resonance problem, environmental factors, the art of workholding, and tacit knowledge, closing with "human-machine collaboration is the future direction" and three pieces of advice for CNC programmers[1]. That structure itself is worth learning from: it breaks the vague one-liner "AI can't do it" into six technical questions that can be discussed one by one.
02The video's six-layer breakdown: from material batch to machine temperament
The difficulties the first half of the video lists are ones anyone on the shop floor will nod at:
- Material: the same "45# steel," in different batches and different heat-treatment states, calls for different parameters — the datasheet gives a starting point, not the answer.
- Tool and machine-tool variation: the actual condition of the same tool model, and each machine tool's rigidity and degree of wear, are all different.
- Resonance: the video says every machine tool has its own "temperament" — the vibration behavior at a specific spindle speed is a property of the individual machine, written outside any manual.
- Environment: temperature, humidity, and even the foundation all affect the performance of high-precision machining.
- Workholding: how tightly to clamp a thin-walled part so it doesn't deform is a hands-on skill.
The video also gives its assessment of the current state: AI programming software has improved and can noticeably speed up simple parts (the video estimates over 30-40%), but complex parts are still done by hand, and all AI output needs human checking — it calls this the "efficiency paradox"[1].
03The hardest layer: sound, chips, and "feel"
The video saves the hardest layer for tacit knowledge: a master knows from the cutting sound whether the tool is still good, knows from the chip's color and shape whether the parameters are right, and knows from a single touch whether the surface is up to par. These judgments are fast and accurate but can't be stated as complete rules — even the masters themselves can't fully explain "why they know," so it's naturally hard to write directly into an algorithm[1].
This observation aligns exactly with the dilemma we discussed in "The master is about to retire — how do we keep 30 years of experience?": Taiwan's machining industry's real asset isn't in the machines but in that undocumented judgment system inside people's heads. The video's contribution is pulling the "AI replacement" debate back to the right level — the question isn't whether AI can write G-code (at the syntax level it can), but that the decisions behind G-code, the ones that rely on shop-floor perception, are still in human hands for now.
04A check against the science: what holds up, what's changing
Cross-checked against the literature, the video's main arguments stand, but there are two "things that are changing" worth adding:
First, "listening to the cut" is being engineered. That's exactly what the cutting-monitoring field has done for thirty years: using sensor signals such as force, vibration, and acoustic emission to monitor tool wear and cutting state — essentially turning a master's ears into a sensor plus an algorithm[2]. It hasn't reached the point of replacing humans, but the absolute claim that "feel can't be digitized" has already loosened in this monitoring subfield.
Second, the data closed loop really is the key bottleneck, consistent with the academic consensus. Reviews of machine learning in manufacturing point out that data acquisition, labeling, and cross-context generalization are universal challenges[3] — the video's line that "the missing data closed loop is the biggest technical bottleneck" is pointed in the right direction. And so, tools and shops willing to accumulate data and build verification loops in their own settings will pull ahead of everyone else in pace of improvement.
As for the effect of material, vibration, and environment on machining, these are basic content in manufacturing-engineering textbooks all along[4] — the video brings them to life in shop-floor language, which is exactly what good technical popularization does.
05A practical path for Taiwan shops: divide the work, don't wait for omnipotent AI
The video's conclusion is human-machine collaboration, which we fully agree with, and can make more concrete: divide the work by "how rule-based it is."
| Task | Nature | Who to give it to today |
|---|---|---|
| Reading drawings, building 3D models | Highly rule-based, time-consuming | AI (human reviews) |
| Generating G-code from tool library and machine limits | Highly rule-based, many combinations | AI (human reviews) |
| Gouge, interference, and collision checks | Verifiable by simulation | 3D cutting simulation |
| Tolerance trade-offs, datums, special processes | Depends on experienced judgment | Human |
| Batch material, workholding feel, on-the-spot tuning | Tacit knowledge | Human (master machinist) |
BestAI CAM is designed to put exactly this division into practice: AI takes over the prep chain from 2D drawing to G-code, generating the program within the shop's tool library and machine ceilings and forcing simulation, with an independent AI cross-checking dimensions — while tolerances, datums, processes, and machine-side decisions are always left to professionals. A master's "feel" doesn't have to compete with AI; it should be used where AI can't reach, and, through the "master-apprentice collaboration workflow," gradually turn into an asset the shop can keep.
06FAQ
The video says AI can't write G-code — is that true?
You have to look at it in layers. At the syntax level, AI can write G-code; the video's real argument is that the decisions behind G-code depend on shop-floor information — material batch, machine-to-machine variation, resonance behavior, environment, and the feel of workholding — which is hard to fully digitize. So the practical conclusion isn't "AI can't be used," but that AI output must be paired with simulation verification and human gatekeeping.
Can AI really not learn a master's skill of listening to the cut and reading the chips?
For thirty years, cutting-monitoring research has been turning sound and vibration into sensor signals and algorithms, and tool-wear monitoring already has mature applications — the "ears" are being engineered, but the integrated judgment is still far short of a master's. A reasonable expectation is that sensing assistance keeps getting stronger, the final judgment stays with humans, and a master's experience is worth recording and passing on systematically.
Simple parts sped up by 30-40%, complex parts done by hand — is that claim credible?
That's the video's own estimate; we can't independently verify the specific numbers, and the source should be cited when quoting it. But the direction — "highly rule-based work is noticeably faster, judgment-dependent work is still done by humans" — is consistent with manufacturing machine-learning literature and various AI CAM tools' hands-on demo videos, and can serve as a reasonable expectation framework for evaluating adoption.
How should a Taiwan shop start with human-machine collaboration?
Divide the work by how rule-based it is: let AI take on the highly rule-based prep work — reading drawings, modeling, and generating G-code from the tool library — paired with cutting simulation and dimensional cross-verification; keep tolerances, datums, special processes, and machine-side decisions gatekept by humans. Start with a controlled pilot on one real drawing, measure the labor-hour difference, then decide whether to expand.
Get notified when new articles and video guides go live — no inbox spam, one-click unsubscribe.
HUMAN + AI, PROPERLY DIVIDED
Give the rule-based work to AI, leave the judgment to the master
Hand the prep chain from reading drawings to G-code to AI for acceleration, use simulation and dimensional cross-verification to cover the risk, and always keep the key decisions in human hands — that's human-machine collaboration you can actually adopt.
Contact a deployment consultant Training courses📋 Download the free "CNC Outsourcing Checklist" (printable) (中文)
07References
- Timber & Tech (YouTube). AI攻克不了的领域竟然是它？揭秘CNC编程为何让AI束手无策｜G代码 (published 2026-02-09). youtube.com/watch?v=IU7e0t2ZpH0
- Teti, R., Jemielniak, K., O'Donnell, G., & Dornfeld, D. (2010). Advanced monitoring of machining operations. CIRP Annals, 59(2), 717–739.
- 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.
- Kalpakjian, S., & Schmid, S. R. (2020). Manufacturing Engineering and Technology (8th ed.). Pearson.