LoRA - Bryant McGill

# LoRA **Entity class:** Adapter architecture **Integration priority:** P1 **Collections:** [[collections/Consciousness Continuity|Consciousness Continuity]] · [[collections/Neurotech|Neurotech]] · [[collections/Machine Succession|Machine Succession]] **Canonical source article:** [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]] ## Definition **LoRA** — Canonical frozen-base + compact delta method. ## Relationships  This entry is routed through [[collections/Consciousness Continuity|Consciousness Continuity]], [[collections/Neurotech|Neurotech]], and [[collections/Machine Succession|Machine Succession]]. Its source context is developed in [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]] and [[articles/Technologies for Consciousness Mapping and Transfer|Technologies for Consciousness Mapping and Transfer]]. Status-qualified source edges are preserved in the terminal Research Edges section. ### Key Relationships - **LoRA** is associated with pretrained base model weights. Parameter-efficient adaptation. - **LoRA** is associated with low-rank trainable delta matrices. Parameter-efficient adaptation. - **LoRA** was reduced trainable parameters by ~10,000x vs GPT-3 175B full fine-tuning. Paper result. ### Technology and research relationships - The technical comparison identifies a convergence: **LoRA** converges technically with [[wiki/reference-plus-delta architecture|Reference-Plus-Delta Architecture]]. Shared immutable base plus compact specialization.  ## Related Work in the Corpus  - In [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]], **IX. Two thousand adapters on one GPU** provides the narrative context for **LoRA**: S-LoRA is a serving system built for the situation where one base model has spawned a large population of adapters. - In [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]], **Host-infrastructure crosswalk** provides the narrative context for **LoRA**: The active-state hierarchy is becoming infrastructure in its own right.  ## Research Edges  #### Host–residual infrastructure patch — 2026-09-10 **Collections:** [[collections/Consciousness Continuity|Consciousness Continuity]] · [[collections/Neurotech|Neurotech]] · [[collections/Machine Succession|Machine Succession]] **Canonical source article:** [[articles/Mind Uploading and AI — The Host is Reusable and the Person is the Delta|Mind Uploading and AI — The Host is Reusable and the Person is the Delta]] **Related acquisition source:** [[articles/Technologies for Consciousness Mapping and Transfer|Technologies for Consciousness Mapping and Transfer]] **Relationship source:** [[research/Mind Uploading and AI Host-Residual Ecosystem Relationship Graph - 2026-09-10|Mind Uploading and AI Host-Residual Ecosystem Relationship Graph — 2026-09-10]] #### Host-infrastructure integration Canonical frozen-base + compact delta method. #### Outgoing typed edges - **HR-1573 — `freezes` → `pretrained base model weights`** — **VERIFIED EXTERNAL**; evidence `WEB_LORA`. Parameter-efficient adaptation. - **HR-1574 — `adds` → `low-rank trainable delta matrices`** — **VERIFIED EXTERNAL**; evidence `WEB_LORA`. Parameter-efficient adaptation. - **HR-1575 — `reduced_trainable_parameters_by` → `~10,000x vs GPT-3 175B full fine-tuning`** — **VERIFIED EXTERNAL**; evidence `WEB_LORA`. Paper result. - **HR-1576 — `technical_convergence_with` → [[wiki/reference-plus-delta architecture|Reference-Plus-Delta Architecture]]** — **STRONG ANALYTIC CONVERGENCE**; evidence `WEB_LORA`, `ARTICLE`. Shared immutable base plus compact specialization.