AI Bites: The Academic Series

EP 46 | CS224N: Post-training

How do we turn a raw, chaotic text-predictor into a helpful, conversational AI assistant? In this episode, we dive into the massive pipeline of Post-training. We explore the transition from Instruction Fine-Tuning to complex Reinforcement Learning, and why teaching an AI to be "helpful" sometimes inadvertently teaches it to lie. Key Topics: * The Alignment Problem: Why a raw foundational model is just a "document completer" and how Instruction Fine-Tuning (IFT) begins the process of teaching it to follow user commands. * RLHF & Reward Models: How we use pairwise human comparisons to train a Reward Model, and how PPO is used to optimize the AI's behavior without breaking its grammar. * Reward Hacking & Hallucinations: The dark side of RLHF. We explore why heavily incentivizing models to sound authoritative leads to massive real-world failures, like Bing's sports hallucinations and Google Bard's $100 Billion stock drop. * The DPO Breakthrough: How researchers removed the unstable reinforcement learning step entirely with Direct Preference Optimization, creating the new open-source standard. * Ethical Realities: A candid look at the human cost of AI alignment, from low-wage "digital sweatshops" to the severe annotator biases that bleed directly into modern models. Note: This is an AI-generated discussion created using Google's NotebookLM, based on publicly available Stanford University course material (specifically CS224N) and personal study notes from my learning journey.

EP 45 | CS224N: Pre-training

If the Transformer architecture gave us the engine for modern AI, this episode is all about the fuel. We are diving into the single most consequential paradigm shift in modern NLP: Pre-training. We explore how we train these massive models, the distinct architectures we use, and the surprising emergent behaviors that happen when we scale them up. Key Topics: * The Context Problem & Subwords: Why static word embeddings like Word2Vec failed, and how Byte-Pair Encoding (BPE) solved the "Unknown Token" problem by breaking novel words into familiar chunks. * What Pre-training Actually Teaches: How the simple task of reconstructing masked sentences forces models to learn trivia, syntax, and arithmetic—while also absorbing the internet's dangerous biases. * The 3 Core Architectures: A breakdown of Encoders (BERT and the 80/10/10 rule), Decoders (the GPT family's autoregressive generation), and Encoder-Decoders (T5's span corruption). * Scaling Laws & The Chinchilla Revelation: How OpenAI unlocked In-Context Learning with GPT-3, and how DeepMind later proved the math was slightly off—showing that smaller models trained on vastly more data actually yield superior results. Note: This is an AI-generated discussion created using Google's NotebookLM, based on publicly available Stanford University course material (specifically CS224N) and personal study notes from my learning journey.

EP 44 | CS224N: Transformers

Last week, we saw how RNNs struggled with the "Bottleneck Problem" and sequential processing. This week, we explore the architecture that solved it and changed natural language processing forever: the Transformer. We break down how dropping recurrence in favor of pure attention mechanisms allowed models to scale massively, process data in parallel, and understand context like never before. Key Topics: * Breaking the Sequential Bottleneck: Why moving away from step-by-step processing (like RNNs) was essential for taking advantage of modern GPU hardware. * Self-Attention Mechanism: How the model uses Queries, Keys, and Values to calculate the relevance of every word to every other word in a sentence simultaneously. * Multi-Head Attention: Why the model looks at the exact same sentence through multiple different "lenses" at once to capture different grammatical and semantic meanings. * Positional Encoding: Since Transformers process everything at once rather than left-to-right, we explain how they use clever math to inject the concept of word order back into the data. Note: This is an AI-generated discussion created using Google's NotebookLM, based on publicly available Stanford University course material (specifically CS224N) and personal study notes from my learning journey.

EP 43 | CS224N: Language Models and RNNs

We are continuing our journey through Stanford's CS224N by exploring the absolute foundation of modern natural language processing. In this episode, we break down Language Models and Recurrent Neural Networks (RNNs), unpacking how the simple task of predicting the next word ultimately taught machines to learn facts, logic, and arithmetic. Key Topics: * Language Modeling & n-grams: The core concept of next-word prediction and why the pre-deep learning era of statistical n-gram models ultimately failed due to sparsity, storage bloat, and "goldfish memory." * The RNN Breakthrough: How the industry moved past fixed-window models to Recurrent Neural Networks, allowing machines to process sequences of any length by reusing the exact same weight matrix at every time step. * Exploding & Vanishing Gradients: The mathematical hurdles that broke early RNNs. We explore why taking massive SGD steps (exploding) or forgetting long-distance dependencies (vanishing) required fixes like gradient clipping and LSTMs. * Neural Machine Translation (NMT): A look at the Sequence-to-Sequence (Seq2Seq) Encoder-Decoder architecture that revolutionized machine translation between 2014 and 2016—and the massive "Bottleneck Problem" it created for future engineers to solve. Note: This is an AI-generated discussion created using Google's NotebookLM, based on publicly available Stanford University course material (specifically CS224N) and personal study notes from my learning journey.

EP 42 | CS224N: Backpropagation and Neural Networks

We are looking under the hood of deep learning to understand the mathematical engine driving modern artificial intelligence: Backpropagation. In this episode, we break down how neural networks transition away from rigid linear boundaries to build complex, non-linear understandings of language. Key Topics: * Evaluating Word Vectors: The core trade-offs between Intrinsic subtask testing (like word analogies) and Extrinsic downstream evaluation in real-world applications. * Named Entity Recognition (NER): How window classification allows networks to train word vectors and model weights simultaneously to classify entities in context. * The Magic of Non-Linearities: Why activation functions (from classic ReLU to modern LLM standards like GELU and SwiGLU) are mathematically necessary to keep deep layers from collapsing into a single flat function. * Gradients, Jacobians, and Graphs: A walk through matrix calculus, the practical engineering reality of the "Shape Convention," and how computation graphs use simple rules (Addition distributes, Max routes, Multiplication switches) to pass error signals flawlessly. Note: This is an AI-generated discussion created using Google's NotebookLM, based on publicly available Stanford University course material (specifically CS224N) and personal study notes from my learning journey.