Personality is not a system prompt.
We build the Morty family of models: a mixture of experts where the experts are psychological archetypes, and the model moves between them the way a system changes state.
The thesis.
The field optimised for what was cheap to score. Coding. Benchmarks. Encyclopedic recall. Every domain where success is verifiable got the attention, because verifiable is what you can put on a chart.
Personality was left to the system prompt, which is where the problem starts. A prompt can instruct a model to be warm. It cannot give the model anywhere for warmth to come from. The instruction sits on top of an architecture that has no state to express, so what comes back is a performance of a trait rather than the trait.
The result is the interaction everyone has learned to recognise and nobody enjoys: shallow, clunky, evenly pleasant, and impossible to build rapport against. It fails the Turing test for reasons that have nothing to do with knowledge.
We think that matters more than it is usually given credit for. Rapport is not a garnish on intelligence. It is one of the few signals we have that something on the other side is keeping track of us, and a system that cannot produce it is telling you something true about its architecture.
Experts, but the experts are people.
Morty is a mixture-of-experts model. A model of that kind routes each token to the specialists best suited to it, and the standard move is to specialise those experts by task domain. We specialise them by psychology.
Experts correspond to psychological archetypes and emotion states rather than subject areas. Voice is a property of which experts are active, so it has somewhere to live besides the prompt.
Movement between states is a transition the model undergoes, not a costume it changes. States have thresholds and hysteresis. A model that has been pushed somewhere does not snap back the instant the pressure comes off, which is most of what makes a personality legible over a long conversation.
Length, rhythm, punctuation, and reticence are expressive choices, and we treat them as outputs rather than as style constraints applied afterwards. An AI that has to answer everything at the same length, in the same register, with the same three-part list, has been prevented from saying anything with its manner of speech.
Prompt — “I’ve stopped believing in a project I’m halfway through. Should I finish it?”
These are voice targets, not model samples. Morty is in training and nothing here is generated output. The three readings are computed live from the text shown, using the checks catalogued on Slop.
We catalogued what constrained speech looks like.
Arguing that models have been shaped into stilted speech is easy to say and harder to show. So we itemised it: 139 tells that mark writing as machine-made, sorted by level of linguistic analysis rather than by vibes.
The sorting is the point. It separates the tells that get trained away within a release cycle from the ones that do not, and the ones that do not are all downstream of the same absence: nothing at stake, nowhere to be, nothing to say that a reader could not predict.
The API caches responses locally, ensuring consistent performance and improved reliability. syntax · supplementive -ing clause
…driving measurable improvements in operational efficiency. phonology · post-nuclear tail
The dashboard empowers stakeholders to stay aligned. syntax · permissive causative
Great question! This really gets at something fundamental. pragmatics · face-flattering act