Where did each AI come from, who built it, and how does it relate to everything else? The complete map of the AI landscape — from the foundational 2017 research to the products people use today.
Perplexity uses third-party models (GPT-4, Claude, Sonar) and adds real-time web search with source attribution. The product is the interface and retrieval layer.
Own model work
Sonar models
Perplexity's own models fine-tuned for search and citation tasks. Available via the Perplexity API.
The neural network design introduced by Google in 2017 that all major LLMs are built on. The "attention mechanism" allows the model to weigh the relevance of every word to every other word in the input.
RLHF — Reinforcement Learning from Human Feedback
Training method where human raters rank model outputs, teaching the model to produce responses humans prefer. Used by OpenAI (ChatGPT), Anthropic (Claude), and most consumer AI products.
Open weights vs open source
Open weights = model parameters are publicly downloadable (Llama, Flux Dev). Open source = code and weights released under permissive licence (Flux Schnell, Mistral). Closed = no access to model internals (GPT-4, Claude).
Mixture of Experts (MoE)
Architecture where multiple "expert" sub-networks exist, but only a subset activates for any given input. Allows much larger effective model capacity at lower inference cost. Used by Mixtral (Mistral) and GPT-4 (reportedly).
Context window
The maximum amount of text a model can process in one conversation — measured in tokens (roughly 3/4 of a word). GPT-4: 128k. Claude: 200k. Gemini Advanced: 1M. Larger = can handle longer documents.
Multimodal AI
Models that understand multiple types of input — text, images, audio, video. GPT-4o, Gemini, and Claude 3+ are multimodal. Earlier LLMs were text-only.
The timeline — key moments
2017
Transformer architecture published
Vaswani et al., Google Brain. "Attention Is All You Need". The foundation of every major AI model that follows.
2018
BERT released by Google
Bidirectional transformer for language understanding. Transformed NLP benchmarks. The precursor to modern search and language AI.
2020
GPT-3 released by OpenAI
175 billion parameters. Demonstrated emergent capabilities — the model could do tasks it wasn't explicitly trained on. Sent shockwaves through AI research.
2021
Anthropic founded
Dario Amodei, Daniela Amodei, and colleagues leave OpenAI to found Anthropic, focused on AI safety research.
2022
Stable Diffusion released
Rombach et al. release open-source image generation. The first high-quality image model freely available to run on consumer hardware.
Nov 2022
ChatGPT launches
GPT-3.5 with RLHF. Reaches 1 million users in 5 days, 100 million in 2 months. AI goes mainstream.
2023
Llama released by Meta as open weights
Meta releases model weights publicly — anyone can download and run a capable LLM. Triggers the open-source AI ecosystem.
Mar 2023
GPT-4 released
Multimodal, significantly more capable than GPT-3.5. Passes bar exam, medical licensing exam. Sets new benchmark standard.
Jun 2023
Mistral AI founded
Ex-DeepMind and Meta researchers raise €105M before shipping a model. European AI research powerhouse.
Sept 2023
Mistral 7B released
Outperforms Llama 2 13B at 7B parameters. Open source. Signals that parameter count isn't everything.
Dec 2023
Gemini 1.0 and Grok-1
Google releases first natively multimodal model. Elon Musk's xAI releases Grok inside X (Twitter).
2024
Claude 3, Llama 3, Gemini 1.5
Major capability improvements across all labs. Gemini 1.5 introduces 1M token context. Arms race accelerates.
Aug 2024
Flux.1 released
Black Forest Labs (ex-Stability AI team) releases Flux.1 — new standard for image generation quality and prompt adherence.
2025
Reasoning models, agentic AI
OpenAI o1/o3, Claude 4. Extended thinking becomes standard. Agentic AI frameworks (LangChain, CrewAI) reach mainstream adoption.
April 2026
AI embedded everywhere
Meta AI in 3B+ WhatsApp/Instagram users. Copilot in all Windows/Office. Gemini in all Google Workspace. AI is infrastructure.