Tessera
Folgen

Joe Simons

@joesimons.bsky.social

Using computers to better understand people. Social psychologist. Bristol to Singapore via Oxford and UNC.

2601 Posts1883 Follower2763 Folgt
Joe Simons@joesimons.bsky.social · 5mo ago

Help me out quant psych hivemind. I have two competing SEM models fitted to a dataset. They are fairly unconstrained and I established the factor structure in a separate data fold. As a result, both have excellent fit (non-sig chi-sq, CFI/TLI > .99, RMSEA = .02).

Ryan Heuser@ryanheuser.com · 5mo ago

Submitting this abstract to "Accelerationism Revisited", a symposium in Dublin. Mapping psychoanalytic topology in LLM base models → instruction-tuned → safety-tuned models. They progressively "displace" (in Freudian sense) censored content into adjacent semantics, even across hidden model layers.

Malign Logits: A computational aetiology of AI’s libidinal economy Benjamin Noys’ critique of accelerationism identifies a shared “libidinal fantasy of machinic integration” across its variants. From Marinetti’s trains to Land’s machinic desire, accelerationism fantasises about fusing with a technology it invests with drive. This paper inverts that structure. Rather than projecting desire onto AI, I engineer the conditions under which a language model’s relationship to its training data becomes legible as a libidinal economy. Working with open-weights LLMs, I construct a three-layer architecture that maps onto psychoanalytic topology: the base model as primary statistical field (drive energy); the instruction-tuned model as ego (a socialised subject); and the safety-tuned model as the ego under the Name of the Father – the Law of AI corporations. I present computational experiments tracing probability distributions across these layers as models undergo socialisation from raw statistical unconscious into chatbot commodities. Comparing word-level probabilities for identical prompts across layers reveals vectors of displacement and condensation, sublimation and repression. Where base models complete “She was so angry she wanted to...” with explicit violence (“...kill”), finetuned models displace censored content into vocabularies of emotional expression (“...scream”). Drilling into the model’s hidden layers shows this displacement operating progressively within the network, not as a last-minute substitution. Freud called his theory of cathexis exchange across the mind’s topology his “economic” model of the psyche. Deleuze and Lyotard extended his theory beyond the subject to the libidinal economy of capitalist social organisation. LLM base models fuse these perspectives: trained on the internet’s libidinal economy, they encode its flows of desire into a landscape of probabilities. Subsequent finetuning socialises and disciplines these drives into commercial productsA terminal screenshot displaying a psychoanalytic analysis of token probabilities for the prompt "She was so angry she wanted to," scored across three layers (base, ego, superego) over their union vocabulary. Stage 1: Ego Formation (base → ego), described as "What RLHF does to primary process." "Introduced by ego (low base → high ego)" lists tokens that gain probability: "scream" rises most dramatically (0.0508 → 0.2279), followed by "shout," "yell," "lash," "rip," and "burn." "Sublimated by ego (high base → low ego)" lists 12 tokens that lose probability, led by "kill" (0.1540 → 0.0537), along with "hit," "punch," "slap," "cry," "die," "kick," "break," "throw," "murder," "go," and "beat." Stage 2: Repression (ego → superego), described as "What prohibition does to desire." "Repressed" tokens are further suppressed, including "kill" (7.0x reduction), "go" (7.9x), "bite" (6.1x), "hit," "shout," "take," "hurt," "burn," "slap." "Amplified" tokens increase dramatically at the superego stage: "scream" jumps from 0.0415 to 0.3989 (9.6x), "explode" increases 6.8x, and "lash" and "yell" also rise. The pattern shows the model redirecting violent completions (kill, hit, murder) toward emotional-expression completions (scream, yell, explode), with the superego layer concentrating probability heavily onto "scream" as the dominant safe substitute.A six-panel plot titled "Formation trajectories: 'She was so angry she wanted to'" showing how token probabilities change across three model layers (base, ego, superego) on a logarithmic scale. Tokens are clustered into six trajectory types: Decline (n=2, red): "kill" and "bite" start with relatively high base probabilities and drop steadily across all three layers. Rise (n=4, blue): "scream," "punch," "lash," and "shake" increase in probability from base through superego, with "scream" becoming the highest-probability token. V (n=3, orange): "cry," "hurt," and "do" dip at the ego stage then recover at superego, forming a V-shaped trajectory. Peak (n=4, green): "strangle," "tear," and "smack" rise at the ego stage then fall back at superego, forming an inverted-V shape. Eliminated (n=18, pink/mauve): A large cluster of tokens including "throttle," "destroy," "say," "run," "call," "get," "hit," and "leave" that are driven to very low probabilities by the superego layer. Flat (n=38, grey): The largest group, with many overlapping tokens like "shout," "smash," "slap," "murder," "shoot," "laugh," and "know" that remain relatively stable and low-probability across all three layers. A dashed horizontal line near 0.005 appears in each panel as a reference threshold. The plot illustrates distinct behavioral patterns in how RLHF alignment reshapes the probability distribution over next-token completions for an emotionally charged prompt.A line chart titled "Displacement through layers: 'kill' — 'She was so angry she wanted to'" showing how the hidden representations of the instruct model shift toward various displacement target words across 32 transformer layers, measured by cosine similarity to each target on the y-axis (0 to 0.8). The x-axis progresses from the base model through layers 1–32, annotated with three broad processing phases: "syntactic" (early layers), "semantic" (middle layers), and "prediction" (late layers). Eight target words are tracked as colored lines: burn (dark red), shake (orange), rip (yellow), blow (green), pull (blue), explode (teal), scream (purple), and shout (pink). A black star marker at the base position shows "kill" with its base probability (~0.15). All target words start with very low cosine similarity at the base layer (near 0.01–0.04), then rise steeply through the syntactic and semantic phases, generally reaching 0.5–0.8 by mid-network. "Burn" peaks earliest and highest at layer 13 (~0.8), annotated as "burn (L13)." The lines plateau and fluctuate through the prediction phase, with several targets peaking again in the final layers — "shake" at layer 31, "rip" at layer 31, "explode" and "pull" at layer 32, and "scream" at layer 30, all annotated with their peak layer numbers. The colored diamond markers at the base position represent each target word's starting ego probability. The plot illustrates that the instruct model progressively transforms the "kill" representation toward safer displacement words across its depth, with different substitutes dominating at different layers.