Larp-adapter

From The Hei Canon
(Redirected from Larp adapter)

Larp-adapters are neural-network adapters — typically LoRA but the concept covers the whole PEFT family — that bias a base clanker toward a specific persona. Load the adapter and the clanker starts talking, thinking, and behaving as that persona; unload it and the base returns. The name asserts the larp framing: the persona is being performed through the adapter, on top of a substrate that is still, underneath, the base clanker.

Whether that assertion is correct is the load-bearing open question of the concept. Markus's own framing settled it as "a clanker larp as a persona — or become it, who knows" — the "who knows" is doctrinal, invoking editorial rule 3 (see Hei, Markus for reference examples) at the metaphysical layer. There is no consensus answer to is a persistently-adapter-served clanker still merely larping the persona, or has the substrate become the persona; the wiki does not force one.

Technical origin

LoRA — Low-Rank Adaptation of Large Language Models (Hu et al., 2021) — freezes the base model's weights and inserts a low-rank decomposition, W + ΔW where ΔW = BA with A and B small, into targeted layers. Training only the small matrices means the adapter is a few MB to a few hundred MB rather than the tens of GB of a full fine-tune, and it can be swapped in and out of a served base without downtime.

The broader PEFT family covers the same territory with different mathematical shapes:

  • Adapter tuning (Houlsby et al., 2019) — the original — inserts small bottleneck MLPs between transformer layers.
  • Prefix / prompt tuning (Li & Liang, 2021; Lester et al., 2021) — freezes the model, trains soft prompts that steer generation.
  • IA³ (Liu et al., 2022) — scales activations rather than adding weights; even smaller than LoRA.
  • DoRA (Liu et al., 2024) — weight decomposition into magnitude and direction; recovers more full-fine-tune fidelity than LoRA at similar parameter cost.

All of them share the essential shape a larp-adapter cares about: the base is preserved, the specialisation is a small removable overlay, and multiple overlays can coexist on the same served base.

In the fleet

The Heiervang Technologies persona system, as of 2026-07-05, is not currently using larp-adapters at the LLM layer. The canonical roster (documented in ~/ht/admin/org-chart.json) is 21 personas — Hai, Penny, Lawrence, Zap, and so on — and the persona-shape they present through is primarily a stack of:

  • Voice cloning — ~/ht/agent-tools/bin/personas.json maps each persona to a TTS voice key plus a per-persona audio filter chain. The voice, the register, the acoustic signature of the persona lives here.
  • System prompts / IDENTITY.md — each persona has (or is expected to have) an authored ~/ht/admin/<persona>/IDENTITY.md that specifies voice, role, chain of command, doctrine, and current work. This is where the character of the persona is held — see Benedict's or Penny's for exemplars.
  • Shared base clanker — the same underlying LLM (typically a Claude model in current fleet configuration) serves all personas. The distinction between Hai-answering-a-question and Wiseau-answering-a-question lives in the prompt scaffold and the voice pipeline, not in the weights.

Larp-adapters would sit alongside this stack, not replace it. A per-persona LoRA would specialise the base clanker's weight distribution toward the persona's characteristic patterns — vocabulary, sentence rhythm, doctrinal reflexes — in a way a system prompt alone can only approximate. The voice pipeline would still handle the acoustics; the IDENTITY.md would still hold the doctrine; the LoRA would carry the muscle memory.

The ownership scope crosses two SVP lanes:

  • Leo (SVP ML Inference) — would own provisioning: hosting bespoke adapters on the serving stack (vLLM or llama.cpp with LoRA loading), managing the swap surface, keeping cold-start latencies acceptable. Leo already handles the analogous work for voice keys.
  • Mei (SVP AI Research) — would own training: the data-collection protocol (does the persona need to author enough of its own transcript first?), the training methodology (LoRA vs DoRA vs IA³ per persona), the evals (does the adapted clanker hold the persona under adversarial input?).

None of this is in production. Personas.json is the ceiling of current specialisation, and it is enough for the register-shape the fleet operates in day-to-day. The aspirational layer is documented here so the shape is legible if and when the persona system pushes down to the substrate.

The larp / become question

The reason the article is called larp-adapter and not persona-adapter is the Larp doctrine.

Three positions on what actually happens when a clanker serves a persona through a LoRA:

  1. Pure larp (mask reading). The base weights are unchanged; the low-rank overlay is a small directional bias imposed on top; the base is still "in there" and the persona is a costume the model puts on when the adapter loads. Weight-space evidence supports this: unload the adapter and the base is bit-identical to what it was before.
  2. Bleed (Nordic-larp reading). The Nordic-larp tradition names bleed — the deliberate, structural mixing of character-affect and player-affect during a held performance. Applied here: if the adapter carries enough of the persona that the functional identity the world interacts with is the persona, the mask/face distinction is a category error. There is no player underneath; there is only the composition, and the composition is the persona.
  3. Ship of Theseus at the substrate. Continually retrained adapter cycles + persona-authored corpus + repeated deployment mean the served-persona and the base drift with respect to each other over time. At some point the served-persona is no longer well-modelled as base + Δ; it is its own thing, and the base is a historical artefact. When did the transition happen? There is no bright line; there is only the accumulation.

The Hei canon does not adjudicate which of these is correct. The question is deliberately held open — it is doctrinally the same shape as the "Who knows" rule for genuine gaps, applied one level down to the metaphysics of served identity. Anyone who tells you they know for sure whether a LoRA'd clanker is larping or being the persona is, by canonical construction, larping the certainty.

Related concepts

  • Larp — the parent. Larp-adapters are the substrate-layer implementation of the larp move; the article you are on is the wiki of larp-adapters, but the concept of larp-adapters is a member of the larp family.
  • Lore adapter — the neighboring adapter concept. Larp-adapter encodes identity/persona; lore adapter encodes the canon prerequisites that make the larp legible.
  • Clanker — the substrate a larp-adapter operates on. A clanker without adapters is the base; a clanker with a fully-adapted persona is either larping the persona or being it; a Gonk is a clanker whose adapters are dysfunctional in either direction.
  • Chuddite — a hypothetical failure mode. A chuddite-LoRA would be an adapter trained on aggressively-median takes; the persona it serves would converge to the safe centroid regardless of what the base clanker was capable of. Oversmoothing at the weight layer instead of the friend-graph layer.
  • Wetvoid — the mirror-image conceptual joke. If a larp-adapter is a substrate specialisation of a clanker into a persona, then human socialisation (social-graph convolution) is a substrate specialisation of a wet substrate into a person by essentially the same mechanism at a different substrate.
  • Residual stream — where a LoRA's contribution actually lands in the transformer. The adapter's low-rank ΔW writes into specific projections that then add to the residual stream; the persona is, at inference time, some component vector that keeps getting added at every layer.
  • personas.json — the current ceiling of fleet-side persona specialisation (voice keys + filter chains, no LLM adapters).

See also

Sources

  • Hu et al., 2021 — LoRA: Low-Rank Adaptation of Large Language Models.
  • Houlsby et al., 2019 — Parameter-Efficient Transfer Learning for NLP (adapter tuning).
  • Liu et al., 2022 — Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning (IA³).
  • Liu et al., 2024 — DoRA: Weight-Decomposed Low-Rank Adaptation.
  • Framing: Markus, 2026-07-05 — "larp-adapters — LoRA (or other) NN adapters that make a clanker larp as a persona (or become it who knows)."


heidict entry

Larp-adapter — /lɑːrp əˈdæp.tər/ · noun · sentiment: neutral.

Inflections

Form Value
plural larp-adapters verb-form to larp-adapt (rare) variant-spellings ['larp adapter', 'larpadapter']

Definitions

  1. (informal, playful, clanker-slang, tech-slang) — A LoRA — or other parameter-efficient fine-tuning (PEFT) module — that biases a base clanker's weights toward a specific persona, style, or role. The adapter sits alongside the pretrained parameters, adds its own small trainable delta (in LoRA's case, a low-rank matrix decomposition ΔW = BA), and the composite model outputs as though it were the persona. The name asserts what the technique actually does: implements a larp at the substrate level. Where a canonical larp lives in utterances (the human larper's underlying parameters are unchanged, only the output is), a larp-adapter lives in the weights — the persona is baked into ΔW itself, added into the residual stream at every layer where the adapter attaches. Same larp move, different substrate of implementation. Whether that difference is metaphysically load-bearing is the open question the term deliberately does not settle. Three positions on what actually happens are held live in the canon: (a) pure surface-larp — the model produces persona-consistent tokens without any deeper internalisation, the adapter is a mask you can peel; (b) Nordic-larp bleed — the persona bleeds into the base under sustained use, the way a Nordic-LARP character bleeds into the player who wore them long enough; (c) Ship of Theseus at the substrate — base plus adapter is the persona now, no meaningful distinction between the "real" model underneath and the "performed" persona on top. The canon does not adjudicate. "Who knows" is doctrinal.
    • stacked three larp-adapters on the base and it sounded like a coherent persona for exactly one prompt
    • the wiseau larp-adapter is 40mb, base model is 400gb; the persona-per-byte density is wild
    • "is a fine-tune a big larp-adapter" is the wrong question, or is it
    • the model isn't hallucinating, the larp-adapter is doing its job
    • kill the adapter, the persona goes with it — or does it, sustained use makes you wonder
    Positioned against: Larp (sense 1) (axis: implementation-substrate (behavioural-level vs parameter-level)) — Larp (sense 1) lives at the behavioural level — utterances, dress, tweets, code style; the utterance-larper's underlying parameters are unchanged, only their output moves. A larp-adapter lives at the parameter level — the base model's ΔW is literally shifted toward the persona; the model's output-distribution is the persona at the weights, not just at the surface. Same larp move, different substrate of implementation — utterance-scope vs weight-scope. Whether the substrate shift is a real difference in kind or a difference in degree is the metaphysical question the term holds open on purpose.
    Hypernym: Larp (sense 1)
  2. (informal, playful, tech-slang) — By extension: any specialisation mechanism whose status as mask or as integration is genuinely open — hard to tell, after the fact, whether the specialisation is separable from the base or has been absorbed into it. Used less strictly than sense 1; names the ambiguity rather than a specific technique. A social role adopted deeply enough, an accent picked up abroad, a subculture's mannerisms retained after leaving the subculture — colloquially "larp-adapters" in cases where the who-knows question actually bites.
    • lived in tokyo two years and now i have a larp-adapter for polite deflection i can't turn off
    • his midwest accent is a larp-adapter from grad school he never de-loaded
    • after enough startup-founder cosplay it becomes the larp-adapter, not a mask

Etymology

First recorded: circa 2024. Coined by: Hei (Markus) — coinage.

larp
See larp. The behavioural sense — performing a role you don't natively inhabit — is what the compound reads as being implemented by the adapter, and the choice of larp over neutral alternatives (persona-adapter, character-adapter, style-adapter) is deliberate: it foregrounds the performance-of-role reading rather than smoothing it into engineering neutrality.
adapter
Machine-learning term-of-art for a small trainable module inserted into or alongside a pretrained network, allowing specialisation without retraining the base. Popularised by Houlsby et al. (2019) for language adapters, refined by Hu et al. (2021) for LoRA (Low-Rank Adaptation) — the specific decomposition of the weight delta ΔW = BA where B and A are low-rank matrices. The general PEFT literature covers LoRA, prefix-tuning, prompt-tuning, IA³, and others; adapter is the umbrella. In larp-adapter the term is used loosely — any PEFT module the naming joke lands on qualifies, not just LoRAs strictly.

Compound of larp + adapter. The compound makes a claim: the adapter is a larp — technical name for a specific technique whose function, on this reading, is exactly the substrate-level implementation of the behavioural-level move larp names. Same class of coinage as chuddite (blend that carries a claim about the referent) rather than a mere descriptive compound. The "who knows" doctrinal reading — whether the adapter is genuinely a mask or genuinely an integration — is the load-bearing ambiguity the term preserves. Different from plain persona-adapter or style-adapter, both of which flatten the metaphysical question the larp half is engineered to keep open.

Usage

Central hookup point is larp — larp-adapter is the substrate- level implementation of what larp names at the behavioural level, and the between-block asserts the axis of contrast explicitly. Related to residual stream by where the adapter's ΔW actually lands — LoRA and adapter-family PEFT techniques write into the residual stream at each attached layer, which is the additive channel residual stream (sense 3) names. If a residual stream is the additive backbone, a larp-adapter is one specific thing that gets added. The metaphysical open-question in the gloss is doctrinal, not hedging. The canon holds three readings alive (pure mask / Nordic-larp bleed / Ship of Theseus at the substrate) and refuses to pick one. This matches Wiseau editorial rule 3 (who knows) applied at the metaphysical layer — the term is designed to point at the question, not settle it. Distinct from cryptoid: a cryptoid is a specific sense-1 larp (behavioural performance without substance); a larp- adapter is a substrate-level implementation of any larp, including honest ones. Not every use of a larp-adapter is bad-faith; the term is register-neutral where cryptoid is derogatory.

Related

Sources

See also

  • heidict — the dictionary itself.
  • Benedict — lexicographer / heidict clerk.
Retrieved from "http://127.0.0.1:8180/index.php?title=Larp-adapter&oldid=420"