GPT-5.3

Ingredient list

1. #1Grouped-Query Attention (GQA)

   Invented by Google · 2023-05 · Velocity 3y

   “GQA is widely adopted in state-of-the-art LLMs for inference efficiency; GPT-5.3 likely incorporates similar optimizations.”

   architecturemedium
2. #2ReAct (Reason + Act)

   Invented by Princeton / Google · 2022-10 · Velocity 3y

   “GPT-5.3-Codex is described as an 'agentic model for autonomous software development,' aligning with ReAct's reasoning+action paradigm.”

   agentsmedium
3. #3FlashAttention

   Invented by Stanford · 2022-05 · Velocity 4y

   “OpenAI's models since GPT-3 have utilized attention optimizations; FlashAttention is a standard for efficient large-scale attention.”

   inferencemedium
4. #4Zero-Shot Chain-of-Thought

   Invented by University of Tokyo · 2022-05 · Velocity 4y

   “GPT-5.3 exhibits zero-shot reasoning, likely prompted with step-by-step directives.”

   reasoningmedium
5. #5Self-Consistency

   Invented by Google · 2022-03 · Velocity 4y

   “GPT-5.3's advanced reasoning capabilities suggest use of sampling-based verification like self-consistency.”

   reasoningmedium
6. #6Reinforcement Learning from Human Feedback (RLHF)

   Invented by OpenAI · 2022-03 · Velocity 4y

   “OpenAI pioneered RLHF with InstructGPT; GPT-5.3 continues this alignment approach.”

   alignmentmedium
7. #7Chain-of-Thought Prompting

   Invented by Google · 2022-01 · Velocity 4y

   “GPT-5.3-Codex demonstrates step-by-step reasoning in autonomous software development tasks.”

   reasoningmedium
8. #8Instruction Tuning (FLAN)

   Invented by Google · 2021-09 · Velocity 4y

   “GPT-5.3 follows OpenAI's instruction-tuned lineage (InstructGPT) for strong zero-shot generalization.”

   trainingmedium
9. #9Rotary Position Embedding (RoPE)

   Invented by Zhuiyi Technology · 2021-04 · Velocity 5y

   “RoPE is a standard position encoding in modern LLMs; GPT-5.3 likely uses it for better length extrapolation.”

   architecturemedium
10. #10Transformer Self-Attention

    Invented by Google · 2017-06 · Velocity 9y

    “GPT-5.3 is a Transformer-based model, using self-attention as its core architecture.”

    architecturehigh
11. #11Mixture of Experts (Sparse MoE for LLMs)

    Invented by Google · 2017-01 · Velocity 9y

    “OpenAI has explored MoE architectures (e.g., GPT-4); GPT-5.3 likely uses sparse MoE for efficient scaling.”

    architecturemedium