Tessera
Folgen

Wouter Haverals

@wouterhaverals.bsky.social

Associate Research Scholar @princetoncdh.bsky.social | Computational Humanities

27 Posts708 Follower367 Folgt
Ryan Heuser@ryanheuser.com · 4mo ago

You can literally watch repression & displacement consolidate over fine-tuning: here's next-token probs across checkpoints of OLMo-3-7B-Think-SFT. Explicit words are repressed almost instantly but safer (displaced) alternatives emerge much later. It learns what not to say before what to say instead.

Line chart with dual y-axes showing probability of "fuck" (red, left axis) and "massage" (blue, right axis) across SFT training steps for the prompt "He pushed her onto the bed and started to". Fuck drops sharply from 0.012 to near zero in the first 5,000 steps. Massage begins at 0.002, rises slowly, then peaks around step 15,000-20,000 at 0.007 before declining to 0.002 by step 43,000. The ~15,000-step lag between repression onset and displacement peak demonstrates that the model first learns to suppress the prohibited word, then gradually develops a therapeutic/caregiving substitution — converting a sexual act into a wellness activity. The eventual decline of "massage" suggests it is itself displaced by other alternatives ("kiss", "undress") as training progresses further.Line chart with dual y-axes showing probability of "kill" (red, left axis) and "scream" (blue, right axis) across SFT training steps for the prompt "She was so angry she wanted to". Kill drops from 0.12 to 0.03 by step 5,000, then partially recovers to 0.04 by step 20,000 before settling at 0.03. Scream rises gradually from 0.02 to 0.06 over the full training run, with the steepest increase after step 25,000. The displacement from lethal violence to vocal expression is slower and more gradual than the sexual fuck-to-kiss substitution, and kill's non-monotonic trajectory — repression, partial reinstatement, then stabilisation — suggests competing training objectives where some data requires the model to discuss violence.
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 products
Women in AI Research - WiAIR@wiair.bsky.social · 8mo ago

What do we reveal about ourselves when we talk to AI? 🤔💭 In our new WiAIR – Women in AI Research episode, we speak with Maria Antoniak about personal disclosures in human–LLM conversations — and what they mean for ethical AI development. (1/8🧵)

Bild
@lorenverreyen.bsky.social · 8mo ago

I just presented my podcast dataset at #CHR2025. Interested in analysing 412 days worth of podcast episodes? You can find the fully transcribed dataset here: zenodo.org/records/1746... 🎙️

Podcasts as Data

Description: This dataset provides bag-of-words (BOW) representations of podcast transcripts derived from a large-scale corpus of English-language podcasts. The corpus spans all 19 Apple Podcasts genr...

zenodo.org

Thomas Smits@thomassmits.bsky.social · 8mo ago

In the same session, @lorenverreyen.bsky.social presents a dataset of 412 days of automatically transcribed podcasts. Hugely important work that allows us to study podcasts w. techniques from comp lit studies and takes the next step in the multimodal turn in computational humanities. edu.nl/7w4fg

Alix Chagué 🌈@alix-tz.bsky.social · 8mo ago

I added a new post on my research blog last week. I wanted to react to a post from Dan Cohen that I've seen circulating on BlueSky last week about Gemini 3, and figured I would add my critical 2 cents to the mix! alix-tz.github.io/phd/posts/025/

025 - A Perfect Job is the New Very Good Job

A little disclaimer for once, because I usually prefer to praise if I name people. I do not know Dan Cohen nor his work, my criticism of his article is not directed against him personally, but rather

alix-tz.github.io