UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
1. AI Applications – Overview (2025 Landscape)
| Domain | Key Applications (2025) | Leading Examples / Companies |
|---|---|---|
| Healthcare | Diagnosis, Drug Discovery, Personalized Medicine | Google DeepMind (AlphaFold 3), IBM Watson Health, Tempus |
| Finance | Fraud Detection, Algorithmic Trading, Credit Scoring | JPMorgan LOXM, PayPal fraud system, Upstart |
| Transportation | Autonomous Vehicles, Traffic Optimization | Tesla FSD v13, Waymo, Uber ATG |
| Education | Personalized Tutoring, Automated Grading | Duolingo, Khan Academy AI, Gradescope |
| Entertainment | Content Generation, Game AI, Recommendation | Netflix, Midjourney, OpenAI Sora |
| Manufacturing | Predictive Maintenance, Quality Control | Siemens MindSphere, GE Predix |
| Agriculture | Precision Farming, Crop Monitoring | John Deere See & Spray, Blue River Technology |
| Defense & Security | Surveillance, Cyber Defense, Drone Swarms | Palantir, Anduril, Israel’s Lavender system |
2. Language Models (LLMs) – The Core of Modern AI
Evolution of Language Models
| Year | Model Family | Size | Breakthrough |
|------|-----------------------|--------------|-------------------------------------------|
| 2017 | Transformer | — | Attention is All You Need paper |
| 2018 | GPT-1 | 117M | Generative Pre-training |
| 2019 | GPT-2 | 1.5B | Zero-shot capabilities |
| 2020 | GPT-3 | 175B | Few-shot learning |
| 2023 | GPT-4 / Claude 2 | ~1.7T | Multimodal (text + image) |
| 2024 | Llama 3 / Grok-2 | 405B–1T+ | Open-source catching up |
| 2025 | GPT-5 class models | >10T | Reasoning, planning, long context (1M+) |
Key Concepts (2025)
- Pre-training → Instruction Tuning → Alignment (RLHF/RLAIF/DPO)
- Retrieval-Augmented Generation (RAG) – LLMs + external knowledge
- Mixture of Experts (MoE) – Only activate needed parameters (e.g., Mixtral, Grok-1)
- Multimodal Models – Text + Image + Audio + Video (GPT-4o, Gemini 1.5, Claude 3.5)
Real-Life Impact (2025)
- 70%+ of code on GitHub is now AI-generated (GitHub Copilot, Cursor)
- Customer support: 90% of queries handled by AI agents (Ada, Intercom AI)
- Education: Personalized tutors for millions (Khanmigo, Duolingo Max)
3. Information Retrieval (IR)
Finding relevant documents from large collections.
Classic IR → Modern Neural IR (2025)
| Approach | Method | Example Tools (2025) |
|---|---|---|
| Boolean Retrieval | AND, OR, NOT | Old search engines |
| Vector Space Model | TF-IDF + Cosine similarity | Elasticsearch (classic) |
| BM25 | Probabilistic ranking | Still used in many systems |
| Dense Retrieval | Embeddings (BERT, ColBERT) | Cohere, Jina AI, Voyage AI |
| Hybrid Retrieval | BM25 + Dense + Re-ranking | Most production systems |
| Learned Sparse (SPLADE) | Combines best of both | Top performer in BEIR benchmark |
Real-Life: Google Search (2025) = MUM + Dense passages + Re-ranking with Gemini
4. Information Extraction (IE)
Extracting structured data from unstructured text.
Sub-tasks
- Named Entity Recognition (NER) → Person, Org, Location
- Relation Extraction → (Elon Musk, CEO_of, Tesla)
- Event Extraction → (Company X, Acquired, Company Y, $10B, 2025)
- Template Filling
2025 State-of-the-Art
- Fine-tuned LLMs (GPT-4, Llama-3-70B-Instruct) outperform traditional models
- Prompt engineering + JSON output mode = best IE system
Example Prompt for IE (2025 style)
Extract all company acquisitions from the text. Return as JSON:
{
"acquisitions": [
{"buyer": "...", "target": "...", "amount_usd": ..., "date": "..."}
]
}
Text: "Microsoft acquired Activision Blizzard for $69 billion in October 2023..."
5. Natural Language Processing (NLP) Pipeline (2025)
| Task | Traditional Method | 2025 Method |
|---|---|---|
| Tokenization | Rule-based | Byte-Pair Encoding (BPE), Tiktoken |
| POS Tagging | HMM, CRF | Built-in to LLMs |
| Parsing | PCFG | Rarely needed (LLMs understand syntax) |
| Sentiment Analysis | VADER, TextBlob | Prompt GPT-4o or Claude 3.5 |
| Text Classification | BERT fine-tuning | Few-shot with Llama-3 405B |
| Summarization | Extractive (TextRank) | Abstractive with Gemini 1.5 Flash |
| Question Answering | BiDAF | RAG with long-context models |
6. Machine Translation (MT)
Evolution
- Rule-based (1950s–1990s)
- Statistical MT (1990s–2010s) → Google Translate (old)
- Neural MT (2016+) → Transformer-based
- 2025: SeamlessM4T v2, NLLB-200, Google Translate (Universal)
Zero-Shot & Multilingual Models (2025)
- One model translates 200+ languages
- Real-time voice-to-voice (e.g., Google Meet live translation)
7. Speech Processing
Speech Recognition (ASR) – 2025
- Whisper (OpenAI) – Best open model
- Google USM, Deepgram, AssemblyAI – Real-time, high accuracy
- Word Error Rate (WER) < 3% on clean English
Text-to-Speech (TTS)
- ElevenLabs, PlayHT, Respeecher – Voice cloning in seconds
- Emotion & style control
End-to-End Voice AI (2025)
- GPT-4o voice mode: Real-time conversation with emotion detection
8. Robotics – The Physical Embodiment of AI
A. Robot Hardware (2025)
| Component | 2025 Technology | Example Robots |
|---|---|---|
| Actuators | High-torque brushless motors, series elastic | Boston Dynamics Atlas, Tesla Bot |
| Sensors | LiDAR, RGB-D cameras, tactile skins, IMUs | Figure 01, Agility Robotics Digit |
| Compute | NVIDIA Jetson Orin NX (275 TOPS), custom AI chips | All modern humanoid robots |
| Batteries | Solid-state batteries (higher density) | Longer operation time |
B. Perception
- Computer Vision: YOLOv10, Segment Anything Model 2 (SAM-2)
- SLAM (Simultaneous Localization & Mapping): ORB-SLAM3, Kimera
- Tactile Sensing: GelSight, DIGIT sensors
C. Planning & Decision Making
- Task & Motion Planning (TAMP)
- Large Language Models for high-level planning (2025 breakthrough)
- SayCan, Code as Policies, RT-2
Example: LLM + Robotics (2025)
# Pseudo-code: Robot uses LLM for planning
user_command = "Make me a cup of tea"
high_level_plan = llm.generate_plan(user_command)
# Output: 1. Go to kitchen 2. Find kettle 3. Fill with water...
for step in high_level_plan:
low_level_actions = vision_language_model(step + current_camera_image)
execute(low_level_actions)
D. Movement & Control
- Reinforcement Learning (RL) for locomotion
- Model Predictive Control (MPC)
- Whole-body control (Boston Dynamics)
Leading Humanoid Robots (November 2025)
| Robot | Company | Status (2025) |
|----------------|--------------------|--------------------------------|
| Atlas | Boston Dynamics | Electric version, super agile |
| Optimus Gen 2 | Tesla | Walking in factories |
| Figure 01 | Figure AI | Working in BMW plant (pilot) |
| Apollo | Apptronik | Warehouse tasks |
| Ameca | Engineered Arts | Best face/expressions |
Summary Table – Unit V (2025 Perspective)
| Area | Dominant Technology (2025) | Killer Application |
|---|---|---|
| Language Models | Multimodal Transformers (10T+) | AI assistants, code generation |
| Information Retrieval | Dense + Hybrid Retrieval | Semantic search engines |
| NLP | Prompting + Fine-tuning LLMs | Chatbots, content creation |
| Machine Translation | Multilingual seamless models | Real-time global communication |
| Speech | End-to-end neural (Whisper, USM) | Voice AI agents |
| Robotics | LLM-guided + Vision + RL control | Humanoid robots in homes/factories |
Key Takeaway for 2025–2030
We are moving from “AI that talks” → “AI that sees, hears, and acts in the physical world.”
The next revolution = Embodied AI (Robots + LLMs) and AI Agents that can autonomously achieve complex goals.
You now have the complete big picture of AI applications in 2025! 🚀
UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
UNIT V — Applications of AI
Complete Notes with Clear Explanations, Real-Life Examples & Key Concepts (2025 Perspective)
1. AI Applications – Overview (2025 Landscape)
| Domain | Key Applications (2025) | Leading Examples / Companies |
|---|---|---|
| Healthcare | Diagnosis, Drug Discovery, Personalized Medicine | Google DeepMind (AlphaFold 3), IBM Watson Health, Tempus |
| Finance | Fraud Detection, Algorithmic Trading, Credit Scoring | JPMorgan LOXM, PayPal fraud system, Upstart |
| Transportation | Autonomous Vehicles, Traffic Optimization | Tesla FSD v13, Waymo, Uber ATG |
| Education | Personalized Tutoring, Automated Grading | Duolingo, Khan Academy AI, Gradescope |
| Entertainment | Content Generation, Game AI, Recommendation | Netflix, Midjourney, OpenAI Sora |
| Manufacturing | Predictive Maintenance, Quality Control | Siemens MindSphere, GE Predix |
| Agriculture | Precision Farming, Crop Monitoring | John Deere See & Spray, Blue River Technology |
| Defense & Security | Surveillance, Cyber Defense, Drone Swarms | Palantir, Anduril, Israel’s Lavender system |
2. Language Models (LLMs) – The Core of Modern AI
Evolution of Language Models
| Year | Model Family | Size | Breakthrough |
|------|-----------------------|--------------|-------------------------------------------|
| 2017 | Transformer | — | Attention is All You Need paper |
| 2018 | GPT-1 | 117M | Generative Pre-training |
| 2019 | GPT-2 | 1.5B | Zero-shot capabilities |
| 2020 | GPT-3 | 175B | Few-shot learning |
| 2023 | GPT-4 / Claude 2 | ~1.7T | Multimodal (text + image) |
| 2024 | Llama 3 / Grok-2 | 405B–1T+ | Open-source catching up |
| 2025 | GPT-5 class models | >10T | Reasoning, planning, long context (1M+) |
Key Concepts (2025)
- Pre-training → Instruction Tuning → Alignment (RLHF/RLAIF/DPO)
- Retrieval-Augmented Generation (RAG) – LLMs + external knowledge
- Mixture of Experts (MoE) – Only activate needed parameters (e.g., Mixtral, Grok-1)
- Multimodal Models – Text + Image + Audio + Video (GPT-4o, Gemini 1.5, Claude 3.5)
Real-Life Impact (2025)
- 70%+ of code on GitHub is now AI-generated (GitHub Copilot, Cursor)
- Customer support: 90% of queries handled by AI agents (Ada, Intercom AI)
- Education: Personalized tutors for millions (Khanmigo, Duolingo Max)
3. Information Retrieval (IR)
Finding relevant documents from large collections.
Classic IR → Modern Neural IR (2025)
| Approach | Method | Example Tools (2025) |
|---|---|---|
| Boolean Retrieval | AND, OR, NOT | Old search engines |
| Vector Space Model | TF-IDF + Cosine similarity | Elasticsearch (classic) |
| BM25 | Probabilistic ranking | Still used in many systems |
| Dense Retrieval | Embeddings (BERT, ColBERT) | Cohere, Jina AI, Voyage AI |
| Hybrid Retrieval | BM25 + Dense + Re-ranking | Most production systems |
| Learned Sparse (SPLADE) | Combines best of both | Top performer in BEIR benchmark |
Real-Life: Google Search (2025) = MUM + Dense passages + Re-ranking with Gemini
4. Information Extraction (IE)
Extracting structured data from unstructured text.
Sub-tasks
- Named Entity Recognition (NER) → Person, Org, Location
- Relation Extraction → (Elon Musk, CEO_of, Tesla)
- Event Extraction → (Company X, Acquired, Company Y, $10B, 2025)
- Template Filling
2025 State-of-the-Art
- Fine-tuned LLMs (GPT-4, Llama-3-70B-Instruct) outperform traditional models
- Prompt engineering + JSON output mode = best IE system
Example Prompt for IE (2025 style)
Extract all company acquisitions from the text. Return as JSON:
{
"acquisitions": [
{"buyer": "...", "target": "...", "amount_usd": ..., "date": "..."}
]
}
Text: "Microsoft acquired Activision Blizzard for $69 billion in October 2023..."
5. Natural Language Processing (NLP) Pipeline (2025)
| Task | Traditional Method | 2025 Method |
|---|---|---|
| Tokenization | Rule-based | Byte-Pair Encoding (BPE), Tiktoken |
| POS Tagging | HMM, CRF | Built-in to LLMs |
| Parsing | PCFG | Rarely needed (LLMs understand syntax) |
| Sentiment Analysis | VADER, TextBlob | Prompt GPT-4o or Claude 3.5 |
| Text Classification | BERT fine-tuning | Few-shot with Llama-3 405B |
| Summarization | Extractive (TextRank) | Abstractive with Gemini 1.5 Flash |
| Question Answering | BiDAF | RAG with long-context models |
6. Machine Translation (MT)
Evolution
- Rule-based (1950s–1990s)
- Statistical MT (1990s–2010s) → Google Translate (old)
- Neural MT (2016+) → Transformer-based
- 2025: SeamlessM4T v2, NLLB-200, Google Translate (Universal)
Zero-Shot & Multilingual Models (2025)
- One model translates 200+ languages
- Real-time voice-to-voice (e.g., Google Meet live translation)
7. Speech Processing
Speech Recognition (ASR) – 2025
- Whisper (OpenAI) – Best open model
- Google USM, Deepgram, AssemblyAI – Real-time, high accuracy
- Word Error Rate (WER) < 3% on clean English
Text-to-Speech (TTS)
- ElevenLabs, PlayHT, Respeecher – Voice cloning in seconds
- Emotion & style control
End-to-End Voice AI (2025)
- GPT-4o voice mode: Real-time conversation with emotion detection
8. Robotics – The Physical Embodiment of AI
A. Robot Hardware (2025)
| Component | 2025 Technology | Example Robots |
|---|---|---|
| Actuators | High-torque brushless motors, series elastic | Boston Dynamics Atlas, Tesla Bot |
| Sensors | LiDAR, RGB-D cameras, tactile skins, IMUs | Figure 01, Agility Robotics Digit |
| Compute | NVIDIA Jetson Orin NX (275 TOPS), custom AI chips | All modern humanoid robots |
| Batteries | Solid-state batteries (higher density) | Longer operation time |
B. Perception
- Computer Vision: YOLOv10, Segment Anything Model 2 (SAM-2)
- SLAM (Simultaneous Localization & Mapping): ORB-SLAM3, Kimera
- Tactile Sensing: GelSight, DIGIT sensors
C. Planning & Decision Making
- Task & Motion Planning (TAMP)
- Large Language Models for high-level planning (2025 breakthrough)
- SayCan, Code as Policies, RT-2
Example: LLM + Robotics (2025)
# Pseudo-code: Robot uses LLM for planning
user_command = "Make me a cup of tea"
high_level_plan = llm.generate_plan(user_command)
# Output: 1. Go to kitchen 2. Find kettle 3. Fill with water...
for step in high_level_plan:
low_level_actions = vision_language_model(step + current_camera_image)
execute(low_level_actions)
D. Movement & Control
- Reinforcement Learning (RL) for locomotion
- Model Predictive Control (MPC)
- Whole-body control (Boston Dynamics)
Leading Humanoid Robots (November 2025)
| Robot | Company | Status (2025) |
|----------------|--------------------|--------------------------------|
| Atlas | Boston Dynamics | Electric version, super agile |
| Optimus Gen 2 | Tesla | Walking in factories |
| Figure 01 | Figure AI | Working in BMW plant (pilot) |
| Apollo | Apptronik | Warehouse tasks |
| Ameca | Engineered Arts | Best face/expressions |
Summary Table – Unit V (2025 Perspective)
| Area | Dominant Technology (2025) | Killer Application |
|---|---|---|
| Language Models | Multimodal Transformers (10T+) | AI assistants, code generation |
| Information Retrieval | Dense + Hybrid Retrieval | Semantic search engines |
| NLP | Prompting + Fine-tuning LLMs | Chatbots, content creation |
| Machine Translation | Multilingual seamless models | Real-time global communication |
| Speech | End-to-end neural (Whisper, USM) | Voice AI agents |
| Robotics | LLM-guided + Vision + RL control | Humanoid robots in homes/factories |
Key Takeaway for 2025–2030
We are moving from “AI that talks” → “AI that sees, hears, and acts in the physical world.”
The next revolution = Embodied AI (Robots + LLMs) and AI Agents that can autonomously achieve complex goals.
You now have the complete big picture of AI applications in 2025! 🚀