Il mio approccio alla programmazione automatica

Summary

The video explains how the creator used an AI coding agent to develop a small CAD program. It argues that effective AI-assisted programming is less about finding a perfect prompt and more about guiding an iterative design and development process.

Main ideas

• AI coding is not simply “write a prompt and wait.” The creator rejects the idea that a single well-worded prompt can produce a finished, useful application. The work involved repeated decisions, implementation, testing, and revisions.
• The human still needs a clear goal—and enough knowledge to guide the work. The creator contributed CAD concepts, design preferences, performance requirements, and judgments about what needed changing. He describes his role as resembling a project manager working closely with capable developers.
• Short demonstrations can hide substantial work. Development involved a couple of intense sessions, each roughly two to three hours long. The creator counts the time spent actively steering and evaluating the agent as work, rather than time the agent may spend generating code unattended.

Development process

• Explore the technical direction before coding. The creator first discussed the problem with an AI assistant rather than immediately starting with a coding agent. They investigated whether a signed distance field (SDF) representation could support the desired CAD operations, including Boolean operations and face-level editing. They also considered how to extract useful topology and handle operations such as extruding faces and making holes.
• Define an achievable first version. The creator requested a minimum viable product with basic shapes and Boolean operations. He wanted a retro-inspired interface reminiscent of Autodesk Animator, but with direct, colorful, easy-to-click controls. He preferred a C implementation, noting that language models write C well in his experience, and asked for minimal dependencies.
• Provide useful technical references and a structure. The agent was asked to gather research papers on SDFs, recognizing vertices and edges, and rendering SDF objects. The creator specified support for basic shapes, different kinds of solids and voids, and hierarchical grouping so operations could later be applied at particular levels.
• Build incrementally and revise through use. The program began with simple geometry and Boolean operations, then grew to include face recognition and operations such as extrusion and twisting. The interface, grid, and controls were repeatedly adjusted. For example, movement controls were designed to behave intuitively relative to the current viewpoint. The creator also introduced a way to select and move fundamental components individually.
• Treat performance and verification as part of the design. When complex objects moved too slowly, the creator directed the agent to cache meshes and display a lower-resolution version during movement, followed by faster, higher-quality GPU rendering when the user stopped moving the object. For fillets and other geometry features, he requested tests using randomized solids and different viewpoints, with screenshots to check for consistent results.

CAD and AI

• CAD remains useful for creating designs directly. The creator argues that a designer with a form in mind can often shape it more directly in CAD than by trying to describe every detail verbally to an AI.
• AI may reduce the need for manual modeling in some workflows. He points to examples of using AI to clean up or reconstruct objects, and suggests that organic 3D modeling may increasingly start with a sketch or generated image that an AI converts into a 3D, articulated model.
• Some design work is harder to automate than geometry production. The creator distinguishes technical modeling from designing functional objects that combine purpose, form, and aesthetic judgment. He believes CAD will remain especially useful for that kind of work, while acknowledging that AI may eventually develop similar creative capabilities.
• Future CAD software should support both people and AI. He suggests designing CAD with a well-documented, direct API that an AI can operate, rather than relying only on an intermediary protocol. AI-oriented features could include easy access to screenshots, different visual representations, and straightforward measurement tools.
• AI commentary can oversimplify the work. The creator criticizes “prompting” advice and AI hype that reduce a complicated process to a slogan. In his view, practical results depend on knowledge, planning, feedback, and iterative collaboration.

The video opens by noting that recent videos have received fewer views and that YouTube recommendations have made it harder for the creator to find the material he normally watches.

Speakers and sources mentioned

• Speaker: Salvatore Sanfilippo, the video’s narrator.
• AI assistant: “Fable 5.1,” as named in the subtitles; it is described as helping with research and technical discussion.
• Referenced creator: Carlo Bloomer, whose channel is recommended as an example of AI-assisted object reconstruction.
• Other products, tools, and platforms mentioned: Tinkercad, Fusion, Sokol, Autodesk Animator, Blender, FreeCAD, YouTube, and AI tools rendered in the subtitles as “Astra” and “Cloud.” Some names may be inaccurate because the subtitles are auto-generated.