AI Systems Landscape

Generative AI — Interactive Architecture Chart

A comprehensive interactive exploration of Generative AI — the generation pipeline, 8-layer stack, output modalities, foundation models, developer tools, benchmarks, market data, and more.

~55 min read · Interactive Reference

Hameem M Mahdi, B.S.C.S., M.S.E., Ph.D. · 2026

Senior Principal Applied Scientist | Private Equity Leader | AI Innovative Solutions

📄 Forthcoming Paper

The Generation Pipeline

Generative AI follows a five-stage pipeline from raw data to novel content generation. Click any step to learn more.

1
DATA
Collect & clean
2
PRE-TRAINING
Self-supervised on massive corpus
3
FINE-TUNING
Domain specialisation, RLHF
4
INFERENCE
User prompt
5
GENERATION
Novel content

Click a pipeline step

Select any stage above to see details about that phase of the generative AI pipeline.

Did You Know?

1

GPT-4 was trained on an estimated 13 trillion tokens — roughly 10 million books' worth of text.

2

Diffusion models generate images by reversing a noise-addition process over ~1,000 denoising steps.

3

The first transformer paper ('Attention Is All You Need', 2017) has been cited over 130,000 times.

Knowledge Check

Test your understanding — select the best answer for each question.

Q1. Which architecture is the foundation of most large language models (LLMs)?

Q2. What technique aligns LLM outputs with human preferences?

Q3. What do diffusion models learn to reverse during image generation?

The Generative AI Stack — 8 Layers

Click any layer to expand its details. The stack is ordered from foundation (bottom) to application (top).

Output Modalities

Generative AI produces content across an expanding set of modalities — from text to molecules.

Core Architectures

The fundamental model architectures that power generative AI across all modalities.

Foundation Models

The leading frontier foundation models driving the generative AI ecosystem in 2026.

ModelOrganisationKey Highlights
GPT-4o / GPT-5OpenAIMultimodal; leading general reasoning
Claude 4 Opus / SonnetAnthropic#1 SWE-Bench 77.2%; 200K context
Gemini 3 Pro / FlashGoogle DeepMind1M+ token context; strong video
Llama 4 Scout / MaverickMetaOpen-weight; 10M-400B MoE
Mistral LargeMistral AIEuropean frontier; multilingual
DeepSeek R1 / V3DeepSeekMIT licensed; reasoning; trained <$6M

Developer Tools & Ecosystem

The frameworks, databases, and tooling that power generative AI application development.

Orchestration Frameworks

FrameworkFocus
LangChainChain-based LLM orchestration & agents
LlamaIndexData ingestion, indexing & RAG pipelines
HaystackEnd-to-end NLP & retrieval pipelines
DSPyProgrammatic prompt optimisation
Semantic KernelMicrosoft SDK for AI orchestration
InstructorStructured output extraction from LLMs

Vector Databases

DatabaseFocus
PineconeManaged vector search; serverless
WeaviateOpen-source; hybrid search
QdrantRust-powered; high performance
ChromaLightweight; developer-friendly
Milvus / ZillizScalable; GPU-accelerated

Industry Use Cases

Real-world generative AI deployments transforming healthcare, legal, and scientific industries.

Benchmarks

Key performance metrics and arena rankings for leading generative AI models.

Model Benchmarks (% score)

Arena Rankings (Elo)

Market Data

Global generative AI market sizing, enterprise spend, and projected growth trajectory.

Market Snapshot (2024)

Gen AI Market Growth — 2024–2030 (CAGR 46%)

Risks & Limitations

Critical limitations and failure modes to consider when deploying generative AI systems.

Glossary

Key terms and concepts in generative AI — searchable and always accessible.

Visual Infographics

Animation infographics for Generative AI — overview and full technology stack.

Regulation

Detailed reference content for regulation.

Regulation & Governance

EU AI Act

Risk Tier Description Examples
Unacceptable Risk Banned outright Social scoring, subliminal manipulation, real-time biometric surveillance
High Risk Strict requirements: conformity assessments, documentation, human oversight CV screening, credit scoring, medical devices, critical infrastructure
Limited Risk Transparency obligations Chatbots must disclose AI nature; deepfakes must be labelled
Minimal Risk No requirements Spam filters, AI in video games, generative content tools
GPAI Models General-purpose AI model requirements Transparency, copyright compliance, safety testing for frontier models

Key Provisions for Generative AI:


United States

Initiative Status
Executive Order on AI (Oct 2023) Broad safety, security, and trust directives for federal AI
NIST AI RMF Voluntary AI Risk Management Framework; widely adopted
State-Level Laws California (SB 1047 vetoed; AB 2013 passed), Colorado AI Act
FTC Guidance Unfair and deceptive practices rules applied to AI
FDA AI/ML Guidance Regulation of AI-enabled medical devices

Global Regulatory Landscape

Jurisdiction Approach
EU Risk-based regulation; legally binding; AI Act in force
UK Pro-innovation; sector-led; no AI-specific law; AI Safety Institute
USA Sector-specific; voluntary frameworks; state-level laws emerging
China Algorithmic recommendation rules; deep synthesis (deepfake) rules; GPAI regulations
Canada AIDA (Artificial Intelligence and Data Act) — in progress
Brazil AI Bill — in legislative process
India Advisory-based; no binding AI law yet
Japan Principle-based; light-touch; "AI-friendly" positioning

Content Authenticity & Watermarking

Initiative Details
C2PA (Content Credentials) Open standard for content provenance metadata; Adobe, Microsoft, OpenAI, Google
SynthID (Google DeepMind) Imperceptible watermarking for AI-generated images, audio, video, and text
DALL·E Watermarking OpenAI embeds C2PA metadata in all DALL·E outputs
EU AI Act Requirement AI-generated content must be labelled as such
Platform Policies Meta, YouTube, TikTok all require disclosure of AI-generated content

Training

Detailed reference content for training.

Training Techniques

Pre-Training

Technique Description
Next Token Prediction Train model to predict the next token given all prior tokens (GPT-style)
Masked Language Modelling Randomly mask tokens; train to predict them (BERT-style)
Contrastive Learning Train model to align related pairs (image + text in CLIP)
Denoising Train model to reconstruct original data from corrupted versions
Mixture of Experts (MoE) Route tokens to specialised sub-networks; scale efficiently

Fine-Tuning Methods

Method Description When to Use
Full Fine-Tuning Update all model weights on domain data Small models; maximum domain performance
LoRA (Low-Rank Adaptation) Add trainable low-rank matrices; freeze base weights Most fine-tuning scenarios; cost-efficient
QLoRA Quantised LoRA; 4-bit base model + LoRA adapters Consumer GPU fine-tuning (24GB VRAM)
Prefix Tuning Prepend trainable tokens to input; keep model frozen Style and tone adaptation
Adapter Layers Insert small trainable modules between frozen layers Multi-task adaptation
PEFT (Parameter-Efficient FT) Umbrella of LoRA, adapters, prefix; HuggingFace library All efficient fine-tuning
Instruction Fine-Tuning Train on (instruction, response) pairs Making models follow instructions

Alignment Techniques

Technique Description Used In
RLHF Human labellers rank outputs; train reward model; PPO optimise ChatGPT, Claude, Gemini
RLAIF Use AI instead of humans to generate preference labels Constitutional AI (Anthropic)
DPO (Direct Preference Optimisation) Train directly on preference pairs; no reward model needed LLaMA 3, Mistral, most open models
Constitutional AI (CAI) AI critiques and revises its own outputs against a constitution Claude (Anthropic)
ORPO Combines SFT and preference learning in one step Efficient alignment
PPO Proximal Policy Optimisation; core RL algorithm for RLHF Original ChatGPT training

Quantisation & Efficiency

Method Description Benefit
INT8 Quantisation 8-bit integer weights instead of 32-bit float 4x memory reduction
INT4 / GPTQ 4-bit quantisation; minimal quality loss 8x memory reduction
GGUF (llama.cpp) Format for running quantised models locally CPU/GPU inference on consumer hardware
AWQ Activation-aware weight quantisation; better quality Deployment on edge devices
Speculative Decoding Small draft model proposes tokens; large model verifies 2-3x faster inference
Flash Attention 2/3 Memory-efficient attention computation Longer contexts; faster training
KV Cache Cache key-value pairs from previous tokens Faster multi-turn inference
Continuous Batching Process multiple requests simultaneously Higher throughput in serving

Enterprise

Detailed reference content for enterprise.

Enterprise Platforms

Cloud Provider AI Platforms

Platform Provider Key Capabilities
Azure OpenAI Service Microsoft GPT-5, o3, DALL·E, Whisper via Azure; enterprise SLAs
Google Vertex AI Google Gemini, Imagen, PaLM; MLOps; Model Garden
AWS Bedrock Amazon Claude, Llama, Titan, Mistral; multi-model; RAG
AWS SageMaker Amazon Custom model training, fine-tuning, deployment
IBM watsonx.ai IBM Granite models; enterprise governance; OpenScale
Oracle OCI AI Oracle Database-native AI; Cohere integration

Enterprise AI Application Platforms

Platform Provider Highlights
Microsoft Copilot 365 Microsoft AI across Word, Excel, Teams, Outlook, PowerPoint
Salesforce Einstein Salesforce CRM-native AI; Agentforce agents
ServiceNow AI ServiceNow IT, HR, and customer service workflow AI
Workday AI Workday HR, finance, and planning AI
SAP Joule SAP Copilot across SAP ERP ecosystem
Adobe Firefly Enterprise Adobe Brand-safe generative AI for creative workflows
Box AI Box Document intelligence; summarisation; Q&A
Slack AI Salesforce Thread summarisation; search; workflow AI
Zoom AI Companion Zoom Meeting summarisation; smart compose; coaching

AI Gateway & Inference Optimisation

Tool Purpose
LiteLLM Universal LLM proxy; route between 100+ models
Kong AI Gateway Enterprise API gateway for LLM traffic
Portkey AI gateway; fallbacks, retries, cost control
Martian Intelligent LLM routing based on task type and cost
Not Diamond Automatic best-model selection per query

Consumer Tools

Detailed reference content for consumer tools.

Consumer & Prosumer Tools

General-Purpose AI Assistants

Product Provider Highlights
ChatGPT OpenAI 700M+ weekly users; GPT-5, o3; multimodal; tool use
Claude.ai Anthropic Claude 4 Opus/Sonnet; 200K context; best for writing and coding
Gemini Google Gemini 2.0/3; integrates with Google Workspace
Copilot Microsoft GPT-5 powered; integrated across Microsoft 365
Le Chat Mistral European alternative; fast inference; Gmail integration
Grok xAI Real-time X/Twitter data; Grok 3 reasoning
Perplexity Perplexity AI Web-grounded answers; citations; research assistant
You.com You.com Search + AI assistant with app integrations
HuggingChat HuggingFace Open-source models; free; no login required

Image Generation Tools

Product Provider Highlights
Midjourney Midjourney Most aesthetically refined; v7; subscription-based
DALL·E 3 OpenAI Integrated in ChatGPT; prompt adherence; inpainting
Stable Diffusion Stability AI Open-source; fully customisable; runs locally
Adobe Firefly Adobe Commercially safe; integrated in Photoshop/Illustrator
Imagen 3 Google Google's highest-quality text-to-image model
Ideogram Ideogram Excellent text rendering within images
Flux Black Forest Labs Open-weight; state-of-the-art quality; fast
Leonardo.ai Leonardo Game asset and concept art generation
Canva AI Canva Magic Generate; design-integrated image generation

Video Generation Tools

Product Provider Highlights
Sora OpenAI 1080p; up to 60s; cinematic quality
Veo 3 Google Native audio generation; YouTube integration
Runway Gen-3 Alpha Runway Professional VFX-grade; image-to-video
Kling 2.0 Kuaishou High-fidelity motion; strong physics simulation
Pika 2.0 Pika Labs Fast generation; scene modification features
HeyGen HeyGen Avatar video; AI dubbing; lip sync
Synthesia Synthesia Enterprise avatar video for training and comms
Luma Dream Machine Luma AI Fast; smooth motion; 3D-grounded generation

Audio & Voice Tools

Product Provider Highlights
ElevenLabs ElevenLabs Best-in-class TTS, voice cloning, dubbing
OpenAI Voice OpenAI Natural realtime conversational voice in ChatGPT
Suno Suno Full song generation from text; v4 model
Udio Udio Music generation; style control; 3-minute tracks
Descript Descript AI podcast and video editing; voice cloning
Adobe Podcast Adobe AI audio enhancement and transcription
Murf Murf Professional TTS for presentations and e-learning
Play.ht Play.ht TTS API; voice cloning; 900+ voices

Writing & Content Tools

Product Provider Highlights
Notion AI Notion Integrated writing assistant; summarisation; Q&A
Jasper Jasper Marketing copy; brand voice training
Copy.ai Copy.ai Marketing and sales content generation
Grammarly Grammarly AI writing assistant; rewriting; tone adjustment
Writesonic Writesonic Blog posts, ads, product descriptions
Sudowrite Sudowrite AI for fiction and creative writing
Hemingway Editor AI Hemingway Clarity and readability scoring with suggestions

Coding Tools

Product Provider Highlights
GitHub Copilot GitHub / OpenAI #1 coding assistant; GPT-4o + Claude 4
Cursor Cursor AI-native IDE; multi-file editing; composer mode
Windsurf Codeium Agent-native IDE; Cascade multi-file agent
Bolt.new StackBlitz Build full-stack web apps in browser from prompts
Lovable Lovable Generate full React apps from natural language
v0 Vercel Generate and edit UI components with AI
Replit Agent Replit Build and deploy apps in natural language
Claude Code Anthropic CLI coding agent; top SWE-Bench performance
Devin Cognition Autonomous software engineering agent

Overview

Detailed reference content for overview.

Definition & Core Concept

Generative AI is the branch of artificial intelligence focused on systems that can produce new content — text, images, video, audio, music, code, 3D models, molecules, and more — that did not exist before the generation event.

Dimension Detail
Core Capability Creates — does not just classify, predict, retrieve, or respond with pre-written text
How It Learns Learns the statistical patterns and distributions of massive training datasets
What It Produces Novel outputs that are plausible, coherent, and contextually appropriate
Key Differentiator Output is generative, not extractive — the model synthesises, it does not copy

Generative AI vs. Other AI Types

AI Type What It Does Example
Generative AI Creates new original content from learned distributions Write an essay, generate an image, synthesise a video
Agentic AI Pursues goals autonomously using tools, memory, and planning Research agent, coding agent, autonomous workflow
Analytical AI Extracts insights and explanations from existing data Dashboard, root-cause analysis, anomaly detection
Autonomous AI (Non-Agentic) Operates independently within fixed boundaries without human input Autopilot, auto-scaling, algorithmic trading
Bayesian / Probabilistic AI Reasons under uncertainty using probability distributions Clinical trial analysis, A/B testing, risk modelling
Cognitive / Neuro-Symbolic AI Combines neural learning with symbolic reasoning LLM + knowledge graph, physics-informed neural net
Conversational AI Manages multi-turn dialogue between humans and machines Customer service chatbot, voice assistant
Evolutionary / Genetic AI Optimises solutions through population-based search inspired by natural selection Neural architecture search, logistics scheduling
Explainable AI (XAI) Makes AI decisions understandable to humans SHAP explanations, LIME, Grad-CAM
Multimodal Perception AI Fuses vision, language, audio, and other modalities GPT-4o processing image + text, AV sensor fusion
Optimisation / Operations Research AI Finds optimal solutions to constrained mathematical problems Vehicle routing, supply chain planning, scheduling
Physical / Embodied AI Acts in the physical world through sensors and actuators Autonomous vehicle, robot arm, drone
Predictive / Discriminative AI Classifies or forecasts from historical patterns Spam filter, credit score, churn prediction
Privacy-Preserving AI Trains and runs AI without exposing raw data Federated hospital models, differential privacy
Reactive AI Responds to current inputs with no memory or learning Chess engine move evaluation, thermostat
Recommendation / Retrieval AI Surfaces relevant items from large catalogues based on user signals Netflix suggestions, Google Search, Spotify playlists
Reinforcement Learning AI Learns optimal behaviour from reward signals via trial and error AlphaGo, robotic locomotion, RLHF
Scientific / Simulation AI Solves scientific problems and models physical systems AlphaFold, climate simulation, molecular dynamics
Symbolic / Rule-Based AI Reasons over explicit rules and knowledge to derive conclusions Medical expert system, legal reasoning engine