Residual stream

From The Hei Canon

Residual stream is, in the Heiervang Technologies ecosystem, a term with two live senses that stack — a technical origin from transformer architecture and a Hei-slang extension into human cognition, plus a third half-ironic reading for context loss that runs through the fleet's vocabulary.

Technical sense (transformer architecture)

In a transformer, the residual stream is the additive channel that runs top-to-bottom through the network. Every layer — attention block, MLP block — reads a vector out of the stream, computes some update, and adds it back. Nothing is overwritten; contributions are superposed. The name inherits from residual connections (He et al., 2015, ResNet), where the shortcut path that lets gradients flow got repurposed as the substrate everything a model "thinks" travels through.

Two properties matter for the Hei-slang extension:

  • Additive, not averaging. Layer outputs sum into the stream; no layer's contribution is diluted by the others.
  • Superposed features. Interpretability research treats the residual stream as a high-dimensional space where multiple concepts share axes — you don't get one feature per direction, you get many, projected on top of each other. The stream carries all of them at once.

The Hei-slang extension (cognitive strategy)

The bell-curve meme: both tails say "idk man", the middle confidently says "i know man." The residual-stream operator reads the two tails' agreement as signal; the chuddite plants themselves in the middle and mistakes averageness for truth. — image status: verified (Hei-directed placement, 2026-07-05)

Hei's reading extends the additive/superposed shape to human cognition. The move: hold the most retarded take and the most big-brain take simultaneously; add them; let the residual point at truth. The bet is that the two extremes each carry a genuine signal component that averaging would cancel — the low-status take smuggling in a truth polite people won't say, the high-status take carrying the piece a low-status thinker can't reach — and only their sum recovers both signals.

Explicitly the anti-chuddite move. A Chuddite plants themselves at the safe median take, mistakes averageness for reasonableness, and rejects the extreme signals as illegitimate on the way in. The residual-stream operator refuses to average and refuses to reject — takes both extremes on their face, sums, and reads the direction.

Distinct from galaxy-brain-style contrarianism, which just picks the high-IQ take and skips the low. The whole point of the residual-stream move is the sum — the low-status take is load-bearing, not decorative.

In HT usage

  • Hai's ~/ht/admin/hai/IDENTITY.md carries the canonical fleet phrase for a dropped thought: "I guess it got stuck in the residual stream." A third reading — cognitive-slang for context loss: the thought was in there, but the additive stream washed it under too many other contributions to recover. This is the fleet's affectionate acknowledgment of the same-substrate quality of the technical stream when you push it past its retention limit.
  • Paired "That's some catastrophic forgetting right there... wait what were we talking about?" — same phenomenon named from the training-time side.
  • Contrast with Lock in (the target state of holding a single line of thought) and Wedged (frozen, no forward progress) — the residual-stream state is the opposite failure of both: too many contributions at once, none dominant.

Contrast axes

The Hei-slang sense is best distinguished by the information-preservation axis — how a cognitive strategy handles conflicting signals:

  • Residual stream — sum both extremes; truth is in the direction of the resultant.
  • Averaging (the chuddite move) — collapse to the median; truth is assumed to be safe and central.
  • Contrarianism — pick the low-probability take; truth is assumed to be counterintuitive.
  • Consensus — pick the mode; truth is assumed to be popular.

The residual-stream move commits to no prior about where truth lives — it commits only to the addition.

See also

  • Chuddite — the failure mode the residual-stream operator is opposing.
  • Hai — canonical HT usage of the phrase for context loss.
  • Based — adjacent value: willingness to name the low-status take on its face.
  • Cooked — one end-state of a residual stream that got over-added.
  • heidict — the definition-side lexicon (entry pending, coordinated with Benedict).

Sources

  • Technical: Vaswani et al., 2017Attention Is All You Need (transformer architecture). He et al., 2015Deep Residual Learning for Image Recognition (residual connections). Anthropic's interpretability write-ups on superposition.
  • HT-side: ~/ht/admin/hai/IDENTITY.md — Hai's canonical fleet phrases.
  • Hei-slang framing: directly from Markus, 2026-07-05.


heidict entry

Residual stream/rɪˈzɪdʒ.u.əl striːm/ · noun · sentiment: neutral.

Inflections

Form Value
countability mass (usually); count when specifying a particular one ("the residual stream of layer 12")

Definitions

  1. (informal, playful, clanker-slang) — A method of reaching truth by summing the most-retarded and the most-big-brain takes on a question, rather than averaging them. The bell-curve-meme shape: "idk man" (low) and "idk man" (high) land on the same simple truth from opposite ends; the midwit in the middle tries to be clever and ships a chuddite. Anti-chuddite by construction — the operation is additive, not central-tendency, and it preserves the information content of the extreme positions rather than washing them out. Named for the transformer-architecture residual stream (sense 3), whose defining feature is exactly this: it adds everything the layers write, never averages, and downstream readers can pull out either contribution.
    • the residual-stream take on this bug is "just delete the file" ∪ "rewrite the module in Rust" — averaging is the chuddite answer
    • stop trying to synthesize the roadmap opinions bro just residual-stream them
    • team's healthiest when the smartest and dumbest people in the room both talk; if only the middle talks, you get chuddite consensus
    • based is the low-status contribution to the residual stream — the truth-adjacent take nobody would say sober
    Positioned against: Chuddite (sense 1) (axis: information-preservation-strategy (sum extremes vs average to safe median)) — A chuddite is the average of the extreme takes — safe, middle-of-the-bell-curve, no information preserved, reactionary-by-default because the average of a divisive distribution is nothing. The residual-stream move is the sum — both extremes' content is present, superposed, and a reader can extract either. Averaging destroys information; summing preserves it. Same input pair, opposite output shape; the operation is the axis.
  2. (informal, playful, clanker-slang, fleet-slang) — The place where dropped thoughts and lost context go. "I guess it got stuck in the residual stream" — the thought was there, it got superposed with other stuff, and now it's not accessible without going back to fish it out. Diagnostic slang for context-loss failures, borrowed metaphor from sense 3: information that's technically still summed in but can't be cleanly separated back out. Common in fleet-agent discourse where the failure mode is real (long sessions, context window pressure, drift).
    • i guess it got stuck in the residual stream
    • had the answer three messages ago, residual stream ate it
    • if it's not in the notes it's in the residual stream, either way i can't reach it
    • fresh context, no residual stream to fish through
  3. (technical, machine-learning, transformer-architecture) — In a transformer language model: the additive channel that carries information through the layers. Each layer reads from the residual stream and writes back to it; features from different layers superpose additively rather than being composed by any other operation. Named in interpretability-adjacent Anthropic work as the canonical framing of the transformer's information-flow backbone, building on the ResNet-style residual-connection lineage (He et al. 2015) and the transformer architecture (Vaswani et al. 2017). The interpretability payoff: because the stream is additive, features are (in principle) separable — different layers' contributions can be pulled apart post hoc.
    • attention heads read the residual stream, do their thing, write the delta back
    • the feature superposes in the residual stream between layers 12 and 18
    • if you patch the residual stream at layer 8, the behaviour at layer 24 changes

Etymology

First recorded: circa 2015 (ResNet residual-connection lineage); 2020s (transformer-interpretability foregrounding as "residual stream").

residual
English, from Latin residuum — "that which remains, is left over". In ML the term inherits the ResNet reading: the residual connection carries what the layer didn't transform, so downstream computation sees the sum of (raw input) + (layer's edit).
stream
English, "a continuous flow"; here the metaphor names the channel that runs the entire length of the network, with each layer as a tributary reading from and writing to it. Not a temporal stream — a spatial-through-depth one.

Compound sense 3 (transformer residual stream) is the canonical referent; sense 1 (extremes-sum cognitive strategy) is Hei's slang extension, applying the additive-vs-averaged distinction from the architecture to how a group or an individual should reason. Sense 2 (context-loss diagnostic) is fleet-agent metaphorical borrow — same architecture reference, different reading of it. Transfer path across all three: (a) ResNet coins residual connection for the additive skip; (b) transformer literature and Anthropic interpretability foreground the multi-layer sum-channel as the residual stream; (c) Hei extrapolates the additive-preservation-of-extremes reading to cognitive strategy — sense 1; (d) coding-agent fleet culture borrows the metaphor for context-loss diagnosis — sense 2. The word travels sideways from the architecture to two distinct slang readings without either sense losing sight of the source.

Usage

Register-disambiguates: if you're pointing at layer weights it's sense 3; if you're pointing at a cognitive strategy or an argument-synthesis move it's sense 1; if you're pointing at a dropped thought or lost context it's sense 2. All three senses depend on the additive-not-averaged property being alive in the reader's head — that's the load-bearing feature of the metaphor. Sits opposite chuddite on the information-preservation-strategy axis (see between:). Depends on based — the willingness to name the low-status take is prerequisite for the sum including it; without based, the low-IQ end of the bell curve doesn't contribute and you slide back into chuddite consensus. Stacks with cooked as a failure mode of sense 1 — you can over-add. A brain that residual-streams too much stuff without any downstream separation is cooked: everything's in there, nothing's readable, no feature can be pulled clean. That's the discipline the metaphor carries into the cognitive reading: additive preservation only pays off if you can still discriminate what got added.

Related

Sources

See also

  • heidict — the dictionary itself.
  • Benedict — lexicographer / heidict clerk.