Unit 1 Case Studies – GenAI for Industry

AI Industry

ChatGPT at Scale: OpenAI's Language Model Revolution

OpenAI launched ChatGPT in November 2022, reaching 100 million users within two months—the fastest-growing consumer application in history. GPT-4, its successor, uses over 1 trillion parameters and powers enterprise adoption across industries worldwide.

Key Facts:

100 million users reached in just 2 months
GPT-4 approximately 1.76 trillion parameters
Adopted by 92% of Fortune 500 companies
Training cost estimated at $100 million+
Market projected to reach $1.3 trillion by 2032

Focus Question

What architectural breakthrough enabled the scaling of language models?

Recurrent Neural Networks (RNNs) Transformer architecture with self-attention mechanism Convolutional Neural Networks (CNNs) Bayesian Networks

Tip: This case illustrates how the Transformer architecture enabled the scaling revolution in language models. The self-attention mechanism allows parallel processing of sequences, making it possible to train on massive datasets efficiently.

Search & Multimodal

Google's Gemini: Multimodal AI in Search

Google deployed Gemini across Search, Workspace, and Android, processing text, images, audio, and video simultaneously. The integration reduced search latency by 40% while improving result quality by 25%.

Key Facts:

True multimodal processing across text, images, audio, video
40% reduction in search latency
25% improvement in result quality
Nano/Pro/Ultra tiers for different deployment needs
1 million token context window capability

Focus Question

How does multimodal processing differ from text-only LLMs?

Uses separate models for each input type Fuses multiple modalities into a unified representation Only supports text and images Requires manual feature engineering for each modality

Tip: This demonstrates the evolution toward multimodal models that can understand and generate across multiple data types. Gemini's architecture processes all modalities in a shared embedding space rather than using separate pipelines.

Open Source AI

Meta's LLaMA: Open-Source LLM Ecosystem

Meta released LLaMA as an open-source large language model, with LLaMA 3.1 405B matching proprietary models in performance. The open release spawned over 100,000 derivative models on Hugging Face, democratizing AI research worldwide.

Key Facts:

Largest open-source LLM at 405 billion parameters
100,000+ derivative models created on Hugging Face
Supports quantization for efficient inference
LoRA fine-tuning for customization
Permissive license enabling commercial use

Focus Question

What techniques make large models practical on limited hardware?

Parallelism only Quantization and LoRA fine-tuning More training data Reduced vocabulary size

Tip: This demonstrates how quantization (reducing precision of model weights) and LoRA (Low-Rank Adaptation) fine-tuning make large models practical for deployment on consumer hardware without significant performance loss.