The ML Engines Under Generative AI: PyTorch & TensorFlow

Introduction

Every model behind every provider SDK in this course — GPT, Claude, Gemini, Llama — was trained using a deep learning framework, almost always PyTorch or TensorFlow. You will not train a foundation model yourself, but recognizing these two names and what they represent explains a lot of vocabulary you'll see across the rest of this ecosystem: "tensors," "checkpoints," ".safetensors" files, and more.

A Real-Life Analogy First

Think of a car manufacturer's factory versus the car you actually drive. PyTorch and TensorFlow are the factory floor — the heavy machinery, assembly lines, and engineering processes used to build an engine (the model). You, as a driver (a developer using a provider SDK), never see the factory floor at all. You just get in the finished car and turn the key. This lesson is a short tour of the factory, purely so the parts list in the finished car's manual — "tensor," "gradient," "checkpoint" — stops sounding like a foreign language.

PyTorch: The Dominant Research Framework

Use case: PyTorch is the framework almost every recent foundation model (including the open-weight models you'll run locally in lesson 18) was built and trained with. It represents data as tensors — multi-dimensional arrays similar to NumPy's, but with automatic differentiation and GPU acceleration built in.

pip install torch

PyTorch in Action

import torch

# A tensor is PyTorch's core data structure -- like a NumPy array,
# but it can track gradients and run on a GPU.
weights = torch.tensor([0.2, 0.5, 0.3], requires_grad=True)
inputs = torch.tensor([10.0, 20.0, 5.0])

output = (weights * inputs).sum()
output.backward()  # computes how much each weight contributed

print("Output:", output.item())
print("Gradients:", weights.grad)

TensorFlow: The Production-Focused Framework

Use case: TensorFlow (often used through its high-level Keras API) covers the same core idea as PyTorch — tensors, automatic differentiation, GPU acceleration — but has historically leaned more toward production deployment tooling (TensorFlow Serving, TensorFlow Lite for mobile/edge).

pip install tensorflow

TensorFlow in Action

import tensorflow as tf

weights = tf.Variable([0.2, 0.5, 0.3])
inputs = tf.constant([10.0, 20.0, 5.0])

with tf.GradientTape() as tape:
    output = tf.reduce_sum(weights * inputs)

grads = tape.gradient(output, weights)

print("Output:", output.numpy())
print("Gradients:", grads.numpy())

Why This Lesson Stays Shallow on Purpose

This course is about the generative AI application layer, not model training — so this lesson intentionally stops at "recognize the vocabulary and see the shape of a tensor operation." If you want the full, dedicated treatment of PyTorch and TensorFlow, including real training loops, this platform's Data Science Dependencies course covers both in depth across two full lessons.

Common Mistakes

Best Practices

Frequently Asked Questions

Key Takeaways

Summary

You do not need to train models to build with generative AI, but recognizing PyTorch and TensorFlow — and the vocabulary they introduced — makes the rest of this ecosystem's documentation far less intimidating.