As with the script cheat sheet, that’s not a post, at least in the classical sense 😊 Rather it’s a collection of knowledge that I will keep updating, for me to find easily.

Main taxonomies of LLMs/AIs

(products listed as of June 2025)

1. By Architecture
   • Transformer-based: GPT, BERT, T5, PaLM, Claude
   • Recurrent: LSTM, GRU-based models (more info)
   • Convolutional: CNN-based language models (more info)
   • Hybrid: Models combining multiple architectures
2. By Training Approach
   • Autoregressive: GPT series, LLaMA, PaLM
   • Masked Language Modeling: BERT, RoBERTa, DeBERTa
   • Encoder-Decoder: T5, BART, UL2
   • Reinforcement Learning from Human Feedback (RLHF): ChatGPT, Claude, Bard
3. By Scale/Size
   • Small: <1B parameters (DistilBERT, MobileBERT)
   • Medium: 1-10B parameters (GPT-2, T5-Base)
   • Large: 10-100B parameters (GPT-3, PaLM-62B)
   • Very Large: 100B+ parameters (GPT-4, PaLM-540B, Claude)
4. By Modality
   • Text-only: GPT-3, BERT, T5
   • Multimodal: GPT-4V, DALL-E, Flamingo, Claude 3
   • Vision: CLIP, ALIGN
   • Audio: Whisper, MusicLM
   • Code: Codex, CodeT5, GitHub Copilot
5. By Capability/Purpose
   • Foundation Models: GPT, BERT, T5 (general-purpose)
   • Specialized: BioBERT (biomedical), FinBERT (finance)
   • Conversational: ChatGPT, Claude, Bard
   • Code Generation: Codex, CodeT5, StarCoder
   • Reasoning: PaLM-2, GPT-4, Claude
6. By Training Data
   • Web-trained: Most large models (Common Crawl, web scrapes)
   • Curated: Models trained on filtered, high-quality datasets
   • Domain-specific: Models trained on specialized corpora
   • Synthetic: Models incorporating AI-generated training data

Architecture Categories

1. Transformer-based: Uses attention mechanisms to process sequences in parallel. Self-attention allows the model to weigh relationships between all tokens simultaneously.

Self-Attention: Attention(Q,K,V) = softmax(QK^T/√d_k)V

Multi-Head: MultiHead(Q,K,V)=Concat(head₁​,…,head_h​)W^O

Processes all positions simultaneously with O(n²) complexity.

2 . Recurrent (LSTM/GRU): Processes sequences step-by-step using memory states. Information flows through hidden states that capture context from previous tokens.

Hidden State Update: h_t​=f(W_h​h_t−1​+W_x​x_t​+b)

LSTM Gates:

Forget: f_t=σ(W_f⋅[h_t−1,x_t]+b_f)

Input: i_t=σ(W_i⋅[h_t−1,x_t]+b_i)

Output: o_t=σ(W_o⋅[h_t−1,x_t]+b_o)

Sequential processing with O(n) complexity

3. Convolutional: Applies sliding filters across text sequences to detect local patterns and features, similar to image processing.

1D Convolution: (f∗g)[n]=∑_mf[m]⋅g[n−m]

Feature Maps: y_i,j=ReLU(∑_kw_k⋅x_i+k,j+b)

Local pattern detection with sliding windows

4. Hybrid: Combines multiple architectures (e.g., transformer + CNN) to leverage different strengths.

Combines architectures: Output=Transformer(CNN(input))

Or parallel processing: Output=α⋅Trans(x)+β⋅RNN(x)

Key difference: Transformers use parallel attention (O(n²) memory), RNNs use sequential states (O(n) memory), CNNs use local convolutions.

Main pros & cons:

• Transformer-based

Pros: Fast parallel processing, excellent at understanding context

Cons: Memory-hungry, struggles with very long texts

• Recurrent (LSTM/GRU)

Pros: Memory-efficient, good at sequential patterns

Cons: Slow training, forgets distant information

• Convolutional

Pros: Fast, good at detecting local patterns

Cons: Limited long-range understanding, less flexible

• Hybrid

Pros: Combines strengths of multiple approaches

Cons: More complex, harder to optimize

In layman’s terms:

• Transformers is reading the whole page at once
• Recurrent is reading word-by-word while taking notes
• Convolutional is scanning for specific phrases, and
• Hybrid is using multiple reading strategies together.

Training Approach Categories

– Autoregressive: Predicts the next token given previous tokens. Trained left-to-right on text sequences.

– Masked Language Modeling: Randomly masks tokens in text and learns to predict them using bidirectional context from both sides.

– Encoder-Decoder: Encodes input into representations, then decodes to generate output. Useful for translation and summarization tasks.

– Reinforcement Learning from Human Feedback (RLHF): Fine-tunes models using human feedback as rewards. Trains the model to generate responses humans prefer through reinforcement learning.

Main pros & cons:

• Autoregressive (GPT-style)

Pros: Great at creative text generation, coherent long-form writing

Cons: Can’t “look ahead” in text, slower for some tasks

• Masked Language Modeling (BERT-style)

Pros: Understands context from both directions, excellent for comprehension

Cons: Poor at generating new text from scratch

• Encoder-Decoder (T5-style)

Pros: Flexible for many tasks, good at text-to-text transformations

Cons: More complex architecture, requires more computational resources

• Reinforcement Learning from Human Feedback (RLHF, ChatGPT/Claude-style)

Pros: Produces helpful, harmless responses aligned with human preferences

Cons: Expensive to train, can be overly cautious or verbose

In layman’s terms:

• Autoregressive is like writing a story word-by-word
• Masked is like fill-in-the-blank exercises
• Encoder-Decoder is like translation between languages, and
• RLHF is like having a human tutor guide the AI’s responses.