From d7ec8ff01748bf79bf0e73ef2f9d20c66e9de2fc Mon Sep 17 00:00:00 2001
From: Vijay Janapa Reddi <vj@eecs.harvard.edu>
Date: Fri, 11 Jul 2025 21:44:09 -0400
Subject: [PATCH] Update README.md

---
 README.md | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 61 insertions(+)

diff --git a/README.md b/README.md
index b76d3008..29878b81 100644
--- a/README.md
+++ b/README.md
@@ -83,6 +83,67 @@ python bin/tito.py test --module [name]  # Test implementation
 
 **Real-World Systems**: Drawing from production [PyTorch](https://pytorch.org/) and [JAX](https://jax.readthedocs.io/) architectures to understand industry-proven design patterns.
 
+## 🤔 Frequently Asked Questions
+
+<details>
+<summary><strong>Why should students build TinyTorch if AI agents can already generate similar code?</strong></summary>
+
+Even though large language models can generate working ML code, building systems from scratch remains *pedagogically essential*:
+
+- **Understanding vs. Using**: AI-generated code shows what works, but not *why* it works. TinyTorch teaches students to reason through tensor operations, memory flows, and training logic.
+- **Systems Literacy**: Debugging and designing real ML pipelines requires understanding abstractions like autograd, data loaders, and parameter updates, not just calling APIs.
+- **AI-Augmented Engineers**: The best AI engineers will *collaborate with* AI tools, not rely on them blindly. TinyTorch trains students to read, verify, and modify generated code responsibly.
+- **Intentional Design**: Systems thinking can’t be outsourced. TinyTorch helps learners internalize how decisions about data layout, execution, and precision affect performance.
+
+</details>
+
+<details>
+<summary><strong>Why not just study the PyTorch or TensorFlow source code instead?</strong></summary>
+
+Industrial frameworks are optimized for scale, not clarity. They contain thousands of lines of code, hardware-specific kernels, and complex abstractions. 
+
+TinyTorch, by contrast, is intentionally **minimal** and **educational** — like building a kernel in an operating systems course. It helps learners understand the essential components and build an end-to-end pipeline from first principles.
+
+</details>
+
+<details>
+<summary><strong>Isn't it more efficient to just teach ML theory and use existing frameworks?</strong></summary>
+
+Teaching only the math without implementation leaves students unable to debug or extend real-world systems. TinyTorch bridges that gap by making ML systems tangible:
+
+- Students learn by doing, not just reading.
+- Implementing backpropagation or a training loop exposes hidden assumptions and tradeoffs.
+- Understanding how layers are built gives deeper insight into model behavior and performance.
+
+</details>
+
+<details>
+<summary><strong>Why use TinyML in a Machine Learning Systems course?</strong></summary>
+
+TinyML makes systems concepts concrete. By running ML models on constrained hardware, students encounter the real-world limits of memory, compute, latency, and energy — exactly the challenges modern ML engineers face at scale.
+
+- ⚙️ **Hardware constraints** expose architectural tradeoffs that are hidden in cloud settings.
+- 🧠 **Systems thinking** is deepened by understanding how models interact with sensors, microcontrollers, and execution runtimes.
+- 🌍 **End-to-end ML** becomes tangible — from data ingestion to inference.
+
+TinyML isn’t about toy problems — it’s about simplifying to the point of *clarity*, not abstraction. Students see the full system pipeline, not just the cloud endpoint.
+
+</details>
+
+<details>
+<summary><strong>What do the hardware kits add to the learning experience?</strong></summary>
+
+The hardware kits are where learning becomes **hands-on and embodied**. They bring several pedagogical advantages:
+
+- 🔌 **Physicality**: Students see real data flowing through sensors and watch ML models respond — not just print outputs.
+- 🧪 **Experimentation**: Kits enable tinkering with latency, power, and model size in ways that are otherwise abstract.
+- 🚀 **Creativity**: Students can build real applications — from gesture detection to keyword spotting — using what they learned in TinyTorch.
+
+The kits act as *debuggable, inspectable deployment targets*. They reveal what’s easy vs. hard in ML deployment — and why hardware-aware design matters.
+
+</details>
+
+---
 ## 🤝 Contributing
 
 We welcome contributions! Whether you're a student who found a bug or an instructor wanting to add modules, see our [Contributing Guide](CONTRIBUTING.md).