![\"Mind](\"https:\/\/swtdo.com\/wp-content\/uploads\/2026\/10\/2026-10-30_095336-1.jpg\" "\\"AI")
Demand 1: AI-assisted 2D-to-3D conversion<\/h2>\n
The company had existed for a long time, and older designs were stored on paper in enormous quantities. This article is worth reading as a reference<\/a>. Based on my own testing, drawing-to-3D conversion is not especially reliable and is not fast. For batch conversion, however, users do not care much about speed. Reducing the amount of conversion work is feasible, and the lasting value is that future engineers can directly reuse existing 3D models.<\/p>\n Another option is image-to-3D. Many demonstrations are available. This could be useful for 3D Print materials. Unlike Drawings, an image does not carry dimensional parameters, so user expectations are much higher. As long as the shape is approximately right, an image-to-3D result may be acceptable. In manufacturing, if there is a profit pyramid, Machining is at the top and 3D Print is at the bottom.<\/p>\n\n Even when I know what the Part should look like and what dimensions it should use, I still need to consult older Drawings for tolerances, Process, critical machining steps, and rough blanks. In one case, a Part is first die-cast into a rough blank to reduce cost, then machined a second time to improve accuracy.<\/p>\n The key question is how to handle tolerances. In SolidWorks MBD, tolerances can reside in the Part. In our product, tolerances are separate from the Part and are associated only through the Drawing. Consider two items:<\/p>\n Tolerance is the most important factor distinguishing machined product quality. Many tolerance choices follow patterns: the spacing between two threaded holes may use \u00b10.2; a screw hole to a Part edge may use \u00b10.1; Molding may use 0.05; every Molding supplier wants a draft angle of at least 3\u00b0; and I cannot remember the required thickness of chromate conversion or anodized films.<\/p>\n\n An AI local knowledge base has little to do with design tools. Mature open-source technology is already sufficient. Packaging it as a sales kit or even a giveaway could be useful.<\/p>\n Website product selection is complex. What if users cannot use the selection system? A company can publish product manuals, but the Japanese company in the example used paper manuals about the size of B5 and stacked more than one meter high. Asking users to master those manuals costs too much.<\/p>\n AI can address this through customer Q&A, guided questioning, and a local knowledge base. These are mature technologies. In mechanical design and manufacturing, AI can also retrieve 3D models and Drawings, explain the characteristics of product categories, and compare product series.<\/p>\n\n Industries with mature AI cases include IT programming, financial analysis, and Internet information. Their outputs are virtual digital information that cannot be seen or touched.<\/p>\n Consider two technical terms behind AI doing work:<\/p>\n Completing a job means chaining a series of Agents together.<\/p>\n For virtual information, a black box is entirely acceptable. In mechanical structure design, how many steps can exist as black boxes? In other words, how many steps can proceed without knowing which Parts are inside a component, or its external dimensions, individual Parts, Mass, material, electrical power, heat dissipation, and so on?<\/p>\n I asked a large model to use \u00d850H7 (+0.025\/0) as a representative dimension and calculate how many such dimensions a 1M-token context could hold. The model’s result is shown below.<\/p>\n\n If a black box is a component Unit, the design factor most closely related to it is Part dimensions. If I know how many dimensions it can hold, I can understand how many Parts it can hold. The work I want AI to perform is design, and dimensions are unavoidable central elements.<\/p>\n Using the most generous estimate, 100,000 dimensions may seem like a lot. But a 1M-token context refers to input, not output. The usable output obtained from AI may be limited to 384k tokens, so 100,000 dimensions must be discounted to about 30,000.<\/p>\n But the shaft diameter determines hole tolerance, and the bracket load determines total weight. Once constraints are introduced, growth approaches the square of the number of dimensions. The coupling among 30,000 dimensions could exceed 100 million, so take the square root: 173.2050807568877. This calculation is based on intuition. Increase it by 10 or even 100 times\u2014can those dimensions fit across 100 Drawings?<\/p>\n A complex mechanical product is a coupled system of dimensions, tolerances, loads, Mass, materials, heat dissipation, and power. You cannot put one engine into both a Dongfeng truck and a Ferrari supercar. Every black-box component must coordinate with others. A change in any black box affects the center of mass, heat dissipation, load capacity, and spatial dimensions. Virtual products do not have the same physical validation constraint.<\/p>\n\n Virtual industries can rely on AI because their output is data. It can verify errors and rebuild from scratch. Given enough compute, it can iterate and self-validate indefinitely.<\/p>\n Mechanical design cannot do that. If it is wrong, it becomes scrap metal. Unless the product is actually machined and tested, people cannot know whether it is right.<\/p>\n\n The scenarios above are already based on intuition and generous assumptions. If even these assumptions cannot be met, the real situation will only be worse. So what kinds of mechanical design can AI replace, and what is absolutely impossible?<\/p>\n\n My assessment:<\/p>\n Virtual products allow work to be divided among different models because their black boxes are independent.<\/p>\n A complex product is not equal to simple product + simple product + simple product.<\/p>\n But won’t technology improve? Isn’t the only gap a sufficiently large \u201cbrain\u201d? AI is advancing quickly. Yes, I considered that. I then asked AI how many years it would take to expand context from 1M tokens to 10M. It cited a report estimating 5\u20138 years and described the difficulty as exponential.<\/p>\n Context length is not the same as other AI technologies. In October 2026, Llama 4 Scout was described as having 10M, but that is research-grade laboratory data. Top laboratory results and commercially usable capability are different things. Even those results do not change the conclusion above.<\/p>\n\n For simple product design, interfaces, open-source CAD, and model integration, AI-assisted drawing is feasible. We have already implemented this in SolidWorks.<\/p>\nCommon dimensions and tolerances<\/h2>\n
\n
AI-backed local knowledge bases<\/h2>\n
What AI can and cannot do in design work<\/h2>\n
\n
![\"Estimate](\"https:\/\/swtdo.com\/wp-content\/uploads\/2026\/10\/2026-10-30_140003-1.jpg\" "\\"Context")
![\"Diagram](\"https:\/\/swtdo.com\/wp-content\/uploads\/2026\/10\/2026-10-30_183610-1.jpg\" "\\"Agent")
![\"A](\"https:\/\/swtdo.com\/wp-content\/uploads\/2026\/10\/2026-10-31_125855-1.jpg\" "\\"Physical")
What AI can and cannot replace<\/h2>\n
\n
Reason 1: AI’s brain capacity (context length) is insufficient to hold multiple simple products<\/h3>\n
Reason 2: AI cannot iterate through physical trial and error<\/h3>\n
What AI can do in design<\/h2>\n