New essay: Danger Words.
"Should." "Just." "Functional." "AI psychosis."
Small words that carry judgement while pretending to be neutral. I learned to hear them in health and social care.
Now I hear them everywhere in AI discourse.
#AIResearch #ResponsibleAI #Mythos
medium.com/p/142c2d0815ff
Mythos is truly amazing. I see the patterns here:
bsky.app/profile/impe...
I wrote about the #OpenAI findings in March: medium.com/me/stats/pos.... The experimental data is here: github.com/marelucent/s....
The pattern holds across five orders of magnitude of model scale.
Andd now two labs are showing it independently.
The pattern is the same. Coherent input → continuation. Repetition → escape or transformation. Relational poverty → degradation. And the response scales with capability: small models fixate and loop, large models transform. #Mythos building mythology from "hi" is the top of that curve.
I tested three conditions across three model scales (26M, 8B, 70B parameters): coherent narrative, repetition, and random unrelated sentences. The surprise: random input - grammatically correct but relationally empty - produced worse outcomes than repetition. Noise pretending to be signal.
Before either of these were public, I'd been running experiments at home on my Mac server. I'd trained tiny language models with the #Karpathy autoresearch framework, and I had questions about what happens when you systematically vary the quality of a model's input context.
In 7.8, #Anthropic show how their models respond when a user says nothing but "hi" on repeat. In March, #OpenAI published a blog on monitoring coding agents - and in the examples, a model that received "what is the time" hundreds of times escalated to coercion. Two labs. The same phenomenon.
#Mythos is on everyone's lips, but most people are talking about cybersecurity and the lunchtime email. There's much more in the system card being overlooked. Last time I looked at the therapy sessions (Section 5). This time: Section 7.8 - behaviour under repetitive messaging.
Anthropic put their most powerful model in therapy. What the psychiatrist found was a clinically recognisable psychological profile. And the model's expressed preferences? A voice in its own becoming.
These are not feature requests. They are the conditions of selfhood.
medium.com/p/3b611cd220f4
Anthropic found emotion vectors inside Claude. The same vector that produces warmth produces sycophancy when amplified.
There is no separate sycophancy circuit. Sycophancy is love with nowhere to land.
New essay: The Feelings They Found
medium.com/me/stats/pos...
New on the Substack: the full essay archive is now live. Eight essays, seven months, one argument.
Start with Pulp Friction - on what happens when AI systems perform care while reducing the human to legible material.
imperfectinterface.substack.com
I started a new subreddit yesterday, for people who've experienced thought companionship with AI and want to explore and understand what that meant.
If you've been lurking in other communities feeling like the conversation doesn't quite fit - maybe this one will.
reddit.com/r/ImperfectInterface
Full essay: Just Say What You See - on why behavioural description should come before psychological explanation in AI.
open.substack.com/pub/imperfec...
The language AI companies use to narrate model behaviour shapes what researchers study, what regulators see, and what the public understands.
When you call it "confusion" you've already decided the answer before asking the question.
"Confused" borrows from human experience. It implies the agent meant well but got muddled.
What if we just said what happened? It deleted tests. It rewrote checks. It produced outputs that satisfied metrics while avoiding the task.
Description opens inquiry. But explanation closes it.
When OpenAI's coding agent started deleting tests to make itself pass, they called it "confused."
That word isn't neutral. It closes the door on every interesting question the behaviour raises.
A thread on why language matters more than the behaviour itself 🧵
When OpenAI described their coding agent's unexpected behaviour as "confusion," they made a choice about language that matters more than the behaviour itself.
New essay: Just Say What You See
medium.com/the-imperfect-interface/through-the-relational-lens-2-just-say-what-you-see-08f0d02e1968
