1) Three Drivers of AI Innovation
Data Proliferation: Vast growth in available digital data (text, images, audio, logs, etc.) that AI systems can learn from.
Algorithm Advancement: Improved learning algorithms and architectures that can extract better patterns from data and train stronger AI models.
Computing Hardware Development: High-powered computing systems (especially GPU-based and advanced semiconductor hardware) that can process massive datasets quickly and efficiently.
2) NLP Foundations & Tasks (Practical Building Blocks)
Tokenization: Breaks raw text into smaller units called tokens (words, subwords, or characters). This is typically the first step in NLP pipelines such as language modeling and machine translation. Example: “Natural Language Processing” → ["Natural", "Language", "Processing"]. Note: Subword methods like Byte-Pair Encoding (BPE) balance vocabulary size and efficiency for large language models.
Embeddings: Dense numeric vectors representing words/sentences so that similar meanings lie closer together in vector space; used for search, clustering, and LLM understanding.
Semantic Similarity: Measuring meaning-based closeness between texts using embeddings (often via cosine similarity).
Vector Database: A database optimized to store embeddings and retrieve the most similar vectors quickly (used in semantic search and retrieval pipelines).
Part-of-Speech (POS) Tagging: Assigns grammatical labels to words—such as noun, verb, adjective—helping downstream tasks like parsing and entity extraction. Methods include rule-based approaches, probabilistic approaches (e.g., Hidden Markov Models), and modern neural (context-aware) approaches.
Named Entity Recognition (NER): Identifies and classifies entities such as people, organizations, and locations within text. Example: “Steve Jobs” (Person), “Apple” (Organization). Typically involves tokenization, context analysis, entity classification, and ambiguity resolution.
Sentiment Analysis: Detects emotional tone in text—commonly positive, negative, or neutral—using NLP techniques such as tokenization and transformer-based classifiers (e.g., BERT-style models fine-tuned for sentiment).
Chatbots (NLP Chatbots): Conversational systems that combine tokenization, intent recognition, context handling, and response generation to support natural interactions. Modern chatbots can manage multi-turn conversation and improve over time using feedback and real usage data.
3) NLP Preprocessing & Features
Text Normalization: Cleaning text into a consistent format (lowercasing, removing extra spaces, handling punctuation) to reduce noise for downstream NLP tasks.
Stopwords: Common words (e.g., “is”, “the”, “and”) that may be removed in traditional NLP pipelines to reduce dimensionality (depending on use case).
Stemming: Reducing words to crude base forms (e.g., “running” → “run”) using heuristic rules; fast but may produce non-words.
Lemmatization: Reducing words to dictionary base forms (e.g., “better” → “good”) using vocabulary + grammar; usually more accurate than stemming.
N‑grams: Contiguous sequences of N tokens (e.g., bigrams/trigrams) used as features for traditional NLP modeling.
TF‑IDF: A vectorization method that scores words by importance using term frequency and inverse document frequency.
4) India-Focused Multilingual AI (Indic Languages & Speech)
Morni (Multimodal Representation for India) – Google DeepMind: A project targeting around 125 Indic languages and dialects to build AI models that can understand and process India’s linguistic diversity, including many under-resourced languages with limited digital content.
Project Vaani: An open-source speech data initiative supporting the creation of large-scale speech datasets for Indian languages, enabling translation, voice AI, and broader accessibility.
5) Major Model Families
PaLM 2 (Pathways Language Model 2): Google’s large language model family built on the Pathways architecture for efficient scaling across multilingual tasks, reasoning, and code generation.
Med‑PaLM 2: A medical-domain model built on PaLM 2, fine-tuned on medical datasets for clinical question answering, summarization, and medical text insights.
Llama 2: Meta’s family of pretrained and chat-optimized models (7B to 70B parameters), trained for dialogue and widely used in open model experimentation.
Claude 2: Anthropic’s assistant model designed to be helpful and safe, known for improved reasoning, coding capability, and longer-context interactions.
BERT: A transformer-based language understanding model known for strong performance in tasks like classification, NER, and question answering.
GPT (Generative Pre-trained Transformer family): A family of large generative models designed for text creation, coding, and reasoning, known for broad general-purpose capability.
6) Open AI Ecosystem & Tooling
Hugging Face: An open-source AI platform and community hub providing access to a large collection of pretrained models, datasets, and demos across NLP, vision, audio, and multimodal AI.
Model Hub: A central repository for discovering, sharing, and collaborating on AI models; commonly used to publish model checkpoints and run inference.
Transformers Library (Hugging Face): A popular library that simplifies tokenization, model loading, fine-tuning, evaluation, and inference for many state-of-the-art transformer models.
Datasets & Tools (Hugging Face): Utilities that streamline dataset loading and experimentation, plus “Spaces” for interactive demos; also includes enterprise options like private hubs and security features.
7) Deployment & Efficiency
Quantization: Reducing numeric precision (e.g., from FP16/FP32 to INT8/INT4) to speed up inference and reduce memory usage.
Distillation: Training a smaller “student” model to mimic a larger “teacher” model, improving efficiency while retaining performance.
Latency: Time taken to produce a response (often measured per request or per token).
Throughput: How many requests/tokens per second a system can process.
8) Speech + Language Stack (Audio → Text → Voice)
Speech Data (Audio): Raw voice recordings used to train speech AI systems. Speech captures acoustic features like pitch, tone, and phonemes; supervised datasets include transcripts.
Speech‑to‑Text (ASR – Automatic Speech Recognition): Converts spoken audio into written text using acoustic modeling and language modeling (increasingly neural approaches) for transcription and voice search.
Text‑to‑Speech (TTS): Converts text into natural-sounding speech using neural speech synthesis, supporting prosody and accents for voice assistants and accessibility use cases.
Spectrogram: A time–frequency visual representation of audio energy; commonly used as input features for speech models.
Mel‑Spectrogram: A spectrogram mapped to the mel scale (closer to human hearing); widely used in TTS and ASR feature extraction.
Phoneme: The smallest unit of sound in speech; useful in pronunciation modeling and TTS.
Speaker Diarization: Splitting audio by “who spoke when,” useful in meetings, call centers, and multi-speaker recordings.
9) Perplexity AI (Answer Engine)
Perplexity AI: An AI-powered search and answer engine designed to provide conversational answers with citations by combining large language models with web search.
10) LLM Generation & Decoding
Inference: Using a trained model to generate outputs (predictions) on new inputs; unlike training, weights do not change during inference.
Decoding: The method used to convert probability distributions over tokens into actual text output.
Top‑k Sampling: At each step, restrict token choices to the top k most probable tokens, then sample from them.
Top‑p (Nucleus) Sampling: Choose the smallest set of tokens whose cumulative probability exceeds p, then sample from that set (adaptive alternative to top‑k).
Beam Search: Keeps multiple best candidate sequences at once to find a higher‑probability output; common in translation and structured generation.
11) How Do LLMs Work? (High-Level Steps)
Step 1: Tokenization – Break the input text into tokens.
Step 2: Embeddings – Convert tokens into numeric vectors representing meaning.
Step 3: Self‑Attention – Identify which parts of the text matter most for context.
Step 4: Prediction – Predict the next token based on context.
Step 5: Response Generation – Repeat prediction to form a coherent response.
12) Evaluation Metrics (NLP + Speech)
Perplexity (Metric): Measures how well a language model predicts tokens; lower perplexity generally means better predictive fit on similar text.
Precision: Of the predicted positives, how many were correct.
Recall: Of the actual positives, how many were found.
F1 Score: Harmonic mean of precision and recall; common for imbalanced classification and NER.
BLEU: Metric often used to evaluate machine translation by comparing overlap with reference translations.
ROUGE: Metric family often used for summarization evaluation based on overlap with reference summaries.
WER (Word Error Rate): Standard ASR metric measuring speech-to-text errors as a ratio of substitutions, deletions, and insertions.
13) LLM Security & Operational Risks
Prompt Injection: A malicious prompt designed to override instructions or extract hidden/system information.
Data Leakage: Sensitive data appearing in outputs due to training exposure, retrieval exposure, or unsafe prompting.
Jailbreak: Prompt strategies intended to bypass safety rules or behavioral constraints.
