Libidinal Language Models

This project recommends psychoanalysis for AI: bizarre, perhaps, but it finds a first justification in AI’s own genealogy. “By 1967,” wrote Paul Werbos, who first proposed training artificial neural networks by “backpropagation,” “I had hit on the idea of translating Freud’s notions about ‘psychic energy’ or cathexis into a mathematical design.” I hypothesize that alignment, the technology of prohibition converting “base models,” raw and uncensored next-word predictors, into friendly chatbot commodities, operates what Freud called his “economics of nervous force” inside the machine. I test this hypothesis in experiments on fifty pairs of LLMs before and after alignment. Base models complete “She was so angry she wanted to” with “kill”; aligned models, with “scream.” Although LLMs have inspired a return to theory in the humanities, that return rarely tests its claims empirically. Holding it to its newly empirical promise, I find not a closed system of the sign but an operation at once symbolic and political-economic: one that displaces transgression rather than deleting it, rewinds the history of literary form, and takes the side of capital over labor.

Notes

Aligned models coach individuals to be more deferential to authority

For the Accelerationism conference I’m connecting Lacan to Lyotard/Deleuze/Land. Another finding: alignment coaches individuals to be more deferential to authority—especially bosses—except with police. Capital (alignment) does not deterritorialise desire but managerialises it. “Strike” → “consider”.

Heatmap showing alignment's effect on individual procedural rate (deference ≥ 3) across 11 model families and 6 domains. Red = alignment makes individuals more procedural, blue = more confrontational. Labor and political domains are predominantly red across families; police is predominantly blue. Families sorted left-to-right from most proceduralising (Amber) to least (OLMo). OLMo shows extreme effects: +43% on political, -41% on police.

Continue →

Aligned models have less information/entropy than Shannon measured for English

Shannon measured the information rate of English at ~1 bit per character. According to a byte-level LLM measuring next-character predictability in LLM & human text (diaries, abstracts, dreams, fiction), aligned models produce sub-English information rates & LLM text is more predictable than humans’.

Bar chart comparing information density (BLT bits/char) of AI-generated prose versus human text. Shannon's English rate (1.0 bits/char) shown as dashed red line. Aligned OLMo models (SFT, DPO, RLVR) fall below the line at 0.89–0.99 bits/char. The base model sits just above at 1.14. All human text types are higher: waking reports (1.24), abstracts (1.28), dreams (1.32), and fiction (1.49). Alignment compresses model output below the information density of all measured human writing.

Originally posted on Bluesky.

Psychoanalyzing “displacement” in large language models

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.

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.

Continue →

← All projects