LLM Race Timeline: The Giants' Competition (2017-2025)

• Diagramm 1: 2017-2019

sequenceDiagram
    participant Google
    participant OpenAI


    rect rgb(200, 220, 240)
        Note over Google: Era 1: Foundation (2017-2019)
        Note over Google: Jun 2017: The Birth of Modern AI
        Google->>Google: "Attention is All You Need" - Transformer Architecture Paper
        Note over Google: Vaswani et al. introduce the Transformer - Foundation for all modern LLMs

        Note over OpenAI: Jun 2018: The First GPT
        OpenAI->>OpenAI: **GPT-1 (117M)** - Generative Pre-training breakthrough
        Note over OpenAI: Demonstrates unsupervised pre-training + supervised fine-tuning

        Note over Google: Oct 2018: Bidirectional Understanding
        Google->>Google: **BERT (340M)** - Bidirectional Encoder Representations
        Note over Google: Revolutionizes NLP tasks, SOTA on GLUE benchmark

        Note over OpenAI: Feb 2019: Coherent Text Generation
        OpenAI->>OpenAI: **GPT-2 (1.5B)** - "Too dangerous to release"
        Note over OpenAI: Staged release due to misuse concerns, impressive text generation
    end

• Diagramm 2: 2020-2022

sequenceDiagram
    participant Google
    participant OpenAI
    participant Meta
    rect rgb(220, 240, 220)
        Note over Google, Meta: Era 2: Scaling Era (2020-2022)
        Note over OpenAI: Jun 2020: THE SCALE BREAKTHROUGH
        OpenAI->>OpenAI: **GPT-3 (175B)** - Massive scale enables few-shot learning
        Note over OpenAI: API-first model, demonstrates emergent capabilities

        Note over Google: Oct 2020: Text-to-Text Framework
        Google->>Google: **T5 (11B)** - Unified framework for all NLP tasks
        Note over Google: "Text-to-Text Transfer Transformer"

        Note over Google: Jan 2021: Sparse Scaling
        Google->>Google: **Switch Transformer (1.6T, ~26B active)** - MoE breakthrough
        Note over Google: Demonstrates efficient scaling with Mixture of Experts

        Note over Meta: Dec 2021: Open Research Initiative
        Meta->>Meta: **OPT (175B)** - Open Pre-trained Transformer
        Note over Meta: Democratizing access to large-scale models

        Note over Google: Apr 2022: Pathways Innovation
        Google->>Google: **PaLM (540B)** - Pathways Language Model
        Note over Google: Breakthrough performance on reasoning tasks

        Note over Google: Mar 2022: Optimal Scaling Laws
        Note over Google: DeepMind's **Chinchilla (70B)** - Compute-optimal training
        Note over Google: Demonstrates importance of training data vs model size

        Note right of OpenAI: Nov 30, 2022: THE TURNING POINT
        OpenAI->>OpenAI: **ChatGPT** (GPT-3.5) - Conversational AI for everyone
        Note over Google, Meta: **Triggers Global AI Revolution**
        Note over Google, Meta: 100M users in 2 months - Fastest growing consumer app ever
    end

• Diagramm 3: 2023

sequenceDiagram
    participant Google
    participant OpenAI
    participant Meta
    participant Anthropic
    participant Mistral AI
    participant Stability AI

    rect rgb(240, 220, 240)
    rect rgb(255, 240, 220)
        Note over Google, Stability AI: Era 3: Post-ChatGPT Explosion (2023)
        Note over Google, Stability AI: All major players respond to ChatGPT

        Google-->>OpenAI: Feb 2023: Rushed Response
        Google->>Google: **Bard** (LaMDA-based) - Search-integrated chatbot
        Note over Google: Early version struggles with factual errors in demo

        Meta->>Meta: Feb 2023: **LLaMA 1 (7B/13B/33B/65B)** - Efficient open models
        Note over Meta: Leaked shortly after release, spawns open-source revolution

        Anthropic-->>OpenAI: Mar 2023: Safety-First Alternative
        Anthropic->>Anthropic: **Claude 1 (~52B)** - Constitutional AI principles
        Note over Anthropic: Emphasis on harmlessness and helpful AI alignment

        OpenAI->>Anthropic: Mar 14, 2023: **GPT-4 (1.76T MoE)** - Multimodal reasoning leader
        OpenAI->>Google: GPT-4 sets new standard
        Note over OpenAI: Passes bar exam (90th percentile), advanced reasoning

        Stability AI->>Stability AI: Apr 2023: **StableLM (3B/7B)** - Open-source text models
        Note over Stability AI: Extends Stable Diffusion success to language

        Google-->>OpenAI: May 2023: Powerful Counter
        Google->>Google: **PaLM 2** - Improved efficiency and multilingual
        Note over Google: Powers Bard upgrade, stronger reasoning

        Meta->>Meta: Jul 18, 2023: **LLaMA 2 (7B/13B/70B)** - Commercial license
        Note over Google, Stability AI: Becomes de-facto standard for open-source LLMs
        Note over Meta: Free for commercial use, <700M monthly users

        Mistral AI->>Mistral AI: Sep 2023: **Mistral 7B** - Efficiency champion
        Note over Mistral AI: Outperforms LLaMA 2 13B, grouped-query attention
        Note over Mistral AI: French startup challenges established players

        Google->>Google: Dec 6, 2023: **Gemini 1.0** (Ultra/Pro/Nano) - Multimodal response
        Note over Google: Claims GPT-4 parity, native multimodal training
    end
    end

• Diagramm 4: 2024

sequenceDiagram
    participant Google
    participant OpenAI
    participant Meta
    participant Anthropic
    participant Mistral AI
    participant DeepSeek
    participant Apple
    rect rgb(240, 220, 240)
        Note over Google, Apple: Era 4: Efficiency & Multimodality (2024)

        Mistral AI->>Mistral AI: Jan 2024: **Mixtral 8x7B (47B, 13B active)** - MoE popularized
        Note over Mistral AI: Matches GPT-3.5 performance at fraction of cost
        Note over Mistral AI: Apache 2.0 license, fully open-source

        Google->>Google: Feb 2024: **Gemini 1.5 Pro (1M token context)** - Context breakthrough
        Note over Google: Longest context window in production, multimodal

        Anthropic->>OpenAI: Mar 2024: **Claude 3** Family (Opus/Sonnet/Haiku)
        Note over Anthropic: **Opus (175B)** surpasses GPT-4 on multiple benchmarks
        Note over Anthropic: MMLU: 86.8% vs GPT-4's 86.4%

        Meta->>Meta: Apr 18, 2024: **LLaMA 3 (8B/70B)** - 15T tokens trained
        Note over Meta: SOTA for openly available models, 8K context

        OpenAI->>OpenAI: May 13, 2024: **GPT-4o** ('omni') - Real-time multimodal
        Note over Google, Apple: Sets new standard for speed, cost, and interaction
        Note over OpenAI: Native audio-visual-text, <320ms voice response time
        Note over OpenAI: 2x faster, 50% cheaper than GPT-4 Turbo

        DeepSeek->>DeepSeek: May 2024: **DeepSeek-V2 (236B, 21B active)** - MoE efficiency
        Note over DeepSeek: Chinese competitor with exceptional cost-efficiency
        Note over DeepSeek: Superior code generation, 128K context

        Google->>Google: May 2024: **Gemini 1.5 Flash** - Fast, cost-efficient multimodal
        Note over Google: Designed to compete with GPT-4o and Claude pricing

        Anthropic->>Anthropic: Jun 20, 2024: **Claude 3.5 Sonnet** - Artifacts feature
        Note over Anthropic: Surpasses GPT-4o on many coding benchmarks
        Note over Anthropic: Interactive "Artifacts" UI - Code in live workspace

        Apple->>OpenAI: Jun 10, 2024: **Apple Intelligence** - WWDC announcement
        Note over Apple: Hybrid on-device + cloud AI strategy
        Note over Apple: On-device: Private, fast, offline-capable
        Note over Apple: Private Cloud Compute for complex tasks
        Note over Apple: Partnership with OpenAI for advanced queries

        OpenAI->>OpenAI: Jul 18, 2024: **GPT-4o mini** - Small model revolution
        Note over OpenAI: 60% cheaper than GPT-3.5 Turbo, outperforms it
        Note over OpenAI: 128K context, fast inference for edge deployments

        Meta->>Meta: Jul 23, 2024: **LLaMA 3.1** (8B/70B/405B) - Open-source flagship
        Note over Meta: 405B model rivals GPT-4 and Claude 3.5 Opus
        Note over Meta: 128K context, multilingual, open-weight license

        Mistral AI->>Mistral AI: Jul 2024: **Mistral Large 2 (123B)** - European leader
        Note over Mistral AI: Competitive with GPT-4o, stronger code/math
        Note over Mistral AI: 128K context window

        Meta->>Meta: Sep 2024: **LLaMA 3.2** (1B/3B/11B/90B) - Vision models
        Note over Meta: First LLaMA models with vision capabilities
        Note over Meta: Edge-optimized 1B/3B models for mobile

        OpenAI->>OpenAI: Sep 12, 2024: **o1-preview & o1-mini** - Reasoning models
        Note over OpenAI: Chain-of-thought reasoning, PhD-level problem solving
        Note over OpenAI: o1: 83rd percentile on Codeforces, excels at math/science
        Note over OpenAI: New paradigm: Thinking time vs response speed

        Anthropic->>Anthropic: Oct 22, 2024: **Claude 3.5 Sonnet** (updated) + **Haiku**
        Note over Anthropic: Sonnet improvements in coding/agentic tasks
        Note over Anthropic: Haiku: Fastest model in Claude 3.5 family

        Google->>Google: Dec 2024: **Gemini 2.0 Flash** - Multimodal agent foundation
        Note over Google: Native image/video generation, tool use
        Note over Google: Designed for agentic AI applications
    end

• Diagramm 5: 2025

sequenceDiagram
    participant Google
    participant OpenAI
    participant Meta
    participant Anthropic
    participant Mistral AI
    participant DeepSeek
    participant Apple
    rect rgb(255, 235, 235)
        Note over Google, Apple: Era 5: Agent & Reasoning Era (2025)

        Note over Google, Apple: Focus shifts to reasoning, planning, and autonomous agents

        DeepSeek->>DeepSeek: Jan 20, 2025: **DeepSeek-V3 (671B, 37B active)** - Open MoE champion
        Note over DeepSeek: Matches GPT-4o performance at drastically lower cost
        Note over DeepSeek: Multi-head latent attention (MLA), efficient training

        DeepSeek->>DeepSeek: Jan 27, 2025: **DeepSeek-R1** - Open reasoning model
        Note over DeepSeek: Challenges OpenAI o1 with transparent reasoning
        Note over DeepSeek: AIME 2024: 79.8% vs o1's 79.2%
        Note over DeepSeek: Open-source alternative to proprietary reasoning models

        OpenAI->>OpenAI: Jan 2025: **Operator** - Browser-controlling agent
        Note over OpenAI: First autonomous web agent, performs tasks in browser
        Note over OpenAI: Can navigate websites, fill forms, make purchases

        Google->>Google: Feb 2025: **Gemini 2.0 Pro** - Production-ready multimodal
        Note over Google: 1M token context, native code execution
        Note over Google: Enhanced agentic capabilities

        Anthropic->>Anthropic: Feb 2025: **Claude 3.5 Opus** - Flagship returns
        Note over Anthropic: Most capable Claude model, enhanced reasoning
        Note over Anthropic: Focused on complex multi-step tasks

        OpenAI->>OpenAI: Apr 2025: **GPT-4.5** - Incremental improvement
        Note over OpenAI: Enhanced reasoning, better tool use
        Note over OpenAI: Bridging GPT-4 and GPT-5

        Meta->>Meta: Apr 2025: **LLaMA 4 (100B/300B/600B)** - Next-gen open models
        Note over Meta: Significantly improved reasoning and multimodal
        Note over Meta: Continues open-weight philosophy

        Google->>Google: May 2025: **Gemini 2.5 Pro** - Long-context reasoning
        Note over Google: 2M token context window, extended reasoning
        Note over Google: Deep Think mode for complex problems

        Mistral AI->>Mistral AI: May 2025: **Mistral Large 3** - European flagship
        Note over Mistral AI: Competitive reasoning capabilities
        Note over Mistral AI: Maintains EU data sovereignty focus

        Apple->>Apple: Jun 2025: **Apple Intelligence 2.0** - iOS 19 / macOS 16
        Note over Apple: Enhanced on-device capabilities with larger models
        Note over Apple: Deeper system integration, proactive assistance

        DeepSeek->>DeepSeek: Jul 2025: **DeepSeek-V4** - Fourth generation MoE
        Note over DeepSeek: Further efficiency improvements
        Note over DeepSeek: Stronger math and code reasoning

        Anthropic->>Anthropic: Aug 2025: **Claude 4** - New architecture
        Note over Anthropic: Major update with enhanced safety features
        Note over Anthropic: Constitutional AI 2.0, improved alignment

        OpenAI->>OpenAI: Sep 2025: **o1** (full release) - Reasoning model production
        Note over OpenAI: Full capabilities unlocked, extended thinking time
        Note over OpenAI: Accessible to all users, not just preview

        Google->>Google: Oct 2025: **Gemini 3.0** (anticipated) - Next generation
        Note over Google: Expected major architectural improvements
        Note over Google: Enhanced multimodal reasoning and agent capabilities
    end