Why More Data Isn't Enough - AI, Parametric CAD, and the Ethics of What We Build

Описание: In this episode, we sit down with Nomi Yu, a researcher who recently graduated from MIT, where she was co-advised between the DeCo Lab and the Mechanosynthesis group, to explore how AI can enable better parametric CAD generation and why the ethical development of these technologies matters just as much as their technical capability.

Nomi shares insights from her work on GenCAD 3D and the challenge of training AI models when usable CAD data is scarce. We discuss why simply having more data isn't enough, how synthetic datasets can address critical biases, and the potential of federated learning to let companies collaborate on training models without ever sharing proprietary IP.

We also dig into the future of engineering workflows, including why the most successful companies will use AI as a starting point rather than a replacement, and the parallels between "vibe coding" in software and what could become "vibe engineering" in hardware design.

In this episode, we cover:

Why data quality and bias correction matter more than data quantity for training CAD generation models
How federated learning could unlock cross-company collaboration without compromising IP
The case for engineers deepening foundational knowledge rather than racing to automate everything



Links from the show:
https://decode.mit.edu/

Get in touch:

Nomi Yu

  / nomiyua6175aadf85  


Raihaan Usman

  / raihaan-usman  


Chapters ➡️

00:00 Introduction to the Blueprint Podcast
00:20 Nomi's Journey in AI and Engineering
03:07 Understanding GenCAD and Parametric Design
05:53 Data Quality and Collaboration in AI
10:27 Challenges in Cross-Domain Learning
12:41 Future of Engineering with AI
16:29 Onshape & Their Dataset
19:26 The Future of Engineering AI
26:39 Verification and Trust in AI Systems
34:52 The Future of Engineering Education
43:48 Responsible AI Development and Ethical Considerations