Continuous FunctionStart learning

Foundation Lab

Infinite Context Architectures

Turns entire repos/books into "single prompt" territory

Concept 100 of 100EfficiencyPhase 13
Inspect equationAll Foundations
#100InfCtxEfficiency
key equation
Mt+1=Update(Mt,Kt,Vt)M_{t+1} = \text{Update}(M_t, K_t, V_t)
Reading map and next steps

Selected Foundation Object

Keep the equation fixed; move through the evidence.

Concept 100 of 100InfCtxEfficiency / Phase 13: Cutting-edge 2024-2025 research
Current question

Compressive: old context summarized into memory state

Mt+1=Update(Mt,Kt,Vt)M_{t+1} = \text{Update}(M_t, K_t, V_t)
PredictionCommit before tracing the equation.

Ask what should change under a concrete input, then trace that expectation through the equation.

EvidenceCompare the equation and source.

Use the key equation and canonical papers as the available witnesses, without implying that a runnable panel exists.

InvariantName what survives notation changes.

The useful learning product is the reusable mechanism you can carry into another model, paper, or engineering tradeoff.

Next moveContinue through the atlas.

Use prerequisites, dependents, and semantic links to repair the next gap without leaving the object behind.

Inspect equationBrowse atlasAll foundations

Why It Matters for Modern Models

  • Turns entire repos/books into "single prompt" territory
  • Streaming: process unbounded sequences with fixed memory
  • 1M+ tokens: Gemini 1.5, LongRoPE, Ring Attention

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Compressive: old context summarized into memory state
  • Ring: sequence chunks processed in ring topology across GPUs
  • Hybrid: combine attention with SSM-style recurrence

Visualization Status

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and cited sources. Come back here for the full walkthrough.

Key Equation
Mt+1=Update(Mt,Kt,Vt)M_{t+1} = \text{Update}(M_t, K_t, V_t)

Compressive memory for bounded cost:

Attn(Q,K,V)=softmax(QK⊤d)VO(n2)\mathrm{Attn}(Q,K,V) = \mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d}}\right)V \quad O(n^2)

Infini-attention: Maintain memory MM updated online, cost bounded w.r.t. nn.

Ring Attention: Distribute long sequences across devices via blockwise ring communication.

Canonical Papers

Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention

Munkhdalai et al.2024Google
Read paper →

Connections

Prerequisites

📏Long Context🔀SSMs & Hybrids

Next Moves

Choose the next question to carry this object forward.