ElemedexA pocket encyclopedia of AI architectures
ENES

Types

Each species has one or two types, as in the games. A type is not a specific architecture but a family of computational ideas: a way of representing, processing or learning. The colors are borrowed from the Pokémon types, and the "analog" says which one.

Symbolic

Analog: Steel

Manipulates explicit symbols through written or learned rules: logic, search, decision trees. GOFAI (symbolic AI)

  • McCulloch–Pitts neuron
  • Logic Theorist
  • ELIZA
  • Expert system (MYCIN)
  • Decision tree (ID3/C4.5)
  • Deep Blue

Probabilistic

Analog: Electric

Explicitly models probability distributions and reasons with them (counts, Bayes, Markov chains). Probability distribution

  • n-gram language model
  • Naive Bayes
  • Expert system (MYCIN)

Geometric

Analog: Ice

Treats data as points in a space and decides by distances, boundaries, or regions. Vector

  • k-nearest neighbors
  • k-means
  • Support vector machine
  • AlphaFold 2

Neural

Analog: Normal

Networks of simple units with adjustable weights. The "normal" type of contemporary AI: almost everything else presupposes it. Neural network

  • McCulloch–Pitts neuron
  • Perceptron
  • Multilayer perceptron (backprop)
  • Simple recurrent network (Elman)
  • LSTM
  • Neural probabilistic language model
  • Hopfield network

Recurrent

Analog: Water

Processes sequences step by step, carrying an internal state that flows from one step to the next. Recurrence

  • Simple recurrent network (Elman)
  • LSTM
  • Seq2seq
  • ELMo
  • Mamba
  • Hopfield network
  • SSM hybrid (Jamba, Qwen3.5)

Convolutional

Analog: Bug

Looks through small filters slid across the input, like a compound eye: detects local patterns and composes them. Convolution

  • Convolutional network (LeNet)
  • AlexNet
  • AlphaGo

Attention

Analog: Psychic

Every element queries every other and weighs how much each matters. The heart of the transformer. Attention

  • Bahdanau attention
  • Transformer
  • Vision Transformer (ViT)
  • AlphaFold 2
  • SSM hybrid (Jamba, Qwen3.5)

Encoder

Analog: Rock

Turns a whole input into representations (vectors) that capture its content. It reads; it does not write. Encoder

  • BERT
  • Sentence-BERT
  • Diffusion language model (LLaDA)

Decoder

Analog: Fire

Produces an output element by element, each step conditioned on what it has already produced. Decoder

  • GPT-1
  • GPT-2
  • GPT-3
  • Chinchilla
  • ChatGPT (InstructGPT)
  • LLaMA
  • GPT-4
  • LLM agent
  • Mamba
  • Mixture of experts (Mixtral)
  • Reasoning model (o1, R1)
  • Claude
  • Gemini
  • DeepSeek-V3

Transducer

Analog: Poison

Encoder and decoder together: reads a whole sequence and writes another (translate, summarize, rewrite). Encoder–decoder

  • Seq2seq
  • Bahdanau attention
  • Transformer
  • T5
  • Whisper

Generative

Analog: Fairy

Learns how data are distributed and can sample new examples: text, images, sound. Generative vs discriminative models

  • n-gram language model
  • VAE
  • GAN
  • GPT-1
  • GPT-2
  • T5
  • GPT-3
  • Diffusion model
  • Chinchilla
  • LLaMA
  • Mixture of experts (Mixtral)
  • DeepSeek-V3
  • Diffusion language model (LLaDA)

Discriminative

Analog: Fighting

Learns to separate: given an input, decides which class it belongs to or which value it gets. Classification

  • Perceptron
  • Naive Bayes
  • k-nearest neighbors
  • Decision tree (ID3/C4.5)
  • Multilayer perceptron (backprop)
  • Support vector machine
  • Convolutional network (LeNet)
  • AlexNet
  • GAN
  • BERT
  • Vision Transformer (ViT)

Embedding

Analog: Flying

Its main product is a space: it places words, sentences, or images as points whose closeness means something. Embedding

  • Neural probabilistic language model
  • word2vec
  • ELMo
  • Sentence-BERT
  • Retrieval-augmented generation (RAG)
  • CLIP

Reinforcement

Analog: Dragon

Learns by acting: tries actions, receives rewards, and adjusts its policy to get more. Reinforcement learning

  • AlphaGo
  • ChatGPT (InstructGPT)
  • Reasoning model (o1, R1)
  • Claude
  • AlphaStar

Latent

Analog: Ghost

Works with hidden variables: compresses data into a latent space and rebuilds it (or makes it emerge from noise). Latent space

  • VAE
  • Diffusion model

Multimodal

Analog: Grass

Relates more than one modality (text, image, audio) within a single representational system. Multimodality

  • CLIP
  • Gato
  • GPT-4
  • Gemini
  • Whisper

Agent

Analog: Ground

A system, not just a model: it plans, uses tools, consults sources, and acts on an environment. Agent

  • Retrieval-augmented generation (RAG)
  • Gato
  • LLM agent
  • AlphaStar

Effectiveness chart

How well each type fares at each family of tasks. The chart is indicative and debatable (that is the point): ×2 means the type is built for it; ×½, that it can but it isn’t its thing; ×0, that it simply doesn’t apply.

ClassifyGenerate textTranslateSemantic searchPerceive imagesCreate imagesPlan / actExplain itself
Symbolic×½×½×½×0×0×2×2
Probabilistic×2×½×½×2
Geometric×2×0×0×2×½×0×½
Neural×2×½
Recurrent×2×½×½×½×½×½
Convolutional×2×½×½×2×½
Attention×2×2×2×2×2×½
Encoder×2×½×2×½×½×½
Decoder×2×½×½×½
Transducer×2×½×½×½×½×½
Generative×2×½×½×2×½
Discriminative×2×0×½×2×0×½
Embedding×0×½×2×½×0
Reinforcement×½×½×½×½×½×2×½
Latent×½×2×½×½
Multimodal×2×2×2×2×½
Agent×2×2

×2 super effective · ×1 normal · ×½ not very effective · ×0 no effect