cultured computer — research summary

today's models are optimized to be right, not to interact. this causes problems now as the personified illusion pops and people feel betrayed. or worse, when the illusion does not pop and people get ai psychosis, unlearn social cues, and pick up an ai flavoured culture17deployment harms: ~560k of ~800M weekly users (0.07%) show possible psychosis or mania indicators (OpenAI, Oct 2025); safety interventions fire in under a third of applicable turns on a clinical benchmark (Au Yeung et al., 2025, “The Psychogenic Machine”); heavy use correlates with loneliness and emotional dependence (OpenAI–MIT Media Lab, 2025); AI mediation reshapes cultural variation, transmission, and selection (Brinkmann et al., 2023, “Machine Culture,” Nature Human Behaviour)..

mid-conversation they lose the environment, their partner, the goal, and themselves. this is why ambient AI and truly interactive AI are still out of reach. we fix this with a harness and interactivity tools around the model, every turn: fixing the stateless problem so it stays coherent, social mechanics to interact like a person, a self to act as someone, binding different signals into one moment to act on, and allowing systems to form learning that carries across interactions.

frontier models pass the bar exam and score at PhD level on knowledge tests, then degrade in actual interaction. system-prompt adherence drops 30%+ and identity drifts within about eight turns, and larger models drift more, not less1instruction/persona drift within ~8 turns, with 30%+ adherence loss and larger models drifting more, attributed to attention decay over the window (Li et al., 2024, Measuring and Controlling Instruction (In)Stability in Language Model Dialogs, COLM; Choi et al., 2024, Examining Identity Drift in Conversations of LLM Agents).. across a full multi-turn task, top models lose around 39% of single-turn performance, mostly by becoming unreliable rather than less capable2across a full multi-turn task, top models lose ~39% of single-turn performance, mostly through unreliability (Laban et al., 2025, LLMs Get Lost in Multi-Turn Conversation). per-turn human-likeness is itself measurable (Adiwardana et al., 2020, Meena/SSA; Thoppilan et al., 2022, LaMDA).. the cause is structural, not a tuning bug.

a transformer is stateless: each call is recomputed from the text in the window, so there is no carried self or thread, and a single bounded-precision pass cannot do the sequential work a state-carrying system can3a transformer is stateless and a single bounded-precision pass cannot express inherently sequential reasoning a state-carrying system can (Vaswani et al., 2017; Merrill & Sabharwal, 2023, The Parallelism Tradeoff, TACL).. attention decay compounds it: instructions at position 0 lose influence by position 1000+, so a persona defined at the top of the window fades as the conversation grows. a model trained as one monolith, with no grounded, updating present, has been likened to a mind in an isolation tank, with no notion of ground truth14Joscha Bach's critique of monolithic models: trained as a single artifact with no grounded, updating present, they are “like a human subject locked into an isolation tank ... not grounded in the persistent experience of a sensory environment ... [with] no notion of ground truth” (Bach, 2026, The Age of AGE: Artificial General Experience)..

this is why the four standard fixes do not hold. a system prompt sits at position 0 and fades; guardrails fire after the model has already spoken; finetuning freezes one starting point and one model; retrieval/memory papers over the missing state. each operates at the start or the surface, and the breakage unfolds deep in the conversation. the duration where the problem actually lives is the duration current fixes leave uncovered.

you do not scale your way to a stable personality; even Claude got there by harnessing, not size. the fix is a layer around the model, built from how a mind sustains an interaction rather than from prompt heuristics. we borrow from the NOW model of consciousness and from decades of cognitive-architecture work on attention, binding, and self-models4the nested observer windows model of consciousness (Riddle & Schooler, 2024, Hierarchical consciousness, Neuroscience of Consciousness) and the attentional conductor that binds many inputs into one momentary state (Bach, 2018/2019; Principles of Synthetic Intelligence, 2009).. early results already back this: the same model jumps from 10% to 80% on voice reasoning under the harness, and an early eval suggests it beats Sierra's own interaction benchmark6VERA-Healthcare: raw Mercury-2 10% vs harnessed 80% (+70pp) at ~500ms first-token, vs the published sub-2s ceiling ~11% (Lin et al., 2025, Voice Evaluation of Reasoning Ability). τ²-bench (Barres et al., 2025): a deterministic scaffold lifts both a frontier and a non-reasoning model, larger relative lift on the weaker model..

here the model is the apex window: brilliant, but it only ever sees the text in front of it. we wrap it in a runtime layer that, every turn, binds only the signals the moment needs into a single “now” for the model to act from11composition from small typed units over a swappable substrate (internal: foundations/observation-primitives); self-improvement by reflecting on output with no weight updates (Shinn et al., 2023, Reflexion; Madaan et al., 2023, Self-Refine; Agrawal et al., 2025, GEPA).. this mirrors how a mind assembles its present, integrating faster signals into one momentary, attention-bound state15the “conductor of attention” that binds signals into one momentary state; consciousness as “second-order perception” and “dynamic representations of control models of attention” (Bach, From Attention to Consciousness? 2021; The Machine Consciousness Hypothesis, Bach & Sorensen); the “bubble of nowness” and the self as “sustained representation of what it is like to be an agent”; the costume framing of identity (Lex Fridman #392, 2023); Levels of Lucidity (2023). Bach advises the lab on the architecture.. the distinction that matters: this runs at runtime, around the model, not in the weights, yet unlike a stateless call it pulls in everything from outside that single call: persistent state, memory, the held self, the arc of the relationship. a stateless model sees one turn; our layer makes every turn consider everything beyond it. the strongest AI systems are converging on this move, composition around a swappable model rather than one scaled weight set5composing systems around a swappable model beats scaling one (Zaharia et al., 2024, The Shift from Models to Compound AI Systems, BAIR; Zhang, Kraska & Khattab, 2025, Recursive Language Models)..

four interactivity ingredients compose that now today, with more on the way:

• state, the opposite of stateless: information held across turns and sessions so the thread, the goal, and the identity survive7the brain holds information across time in persistent firing and silently in synapses, and is intrinsically active even at rest (Goldman-Rakic, 1995, Neuron; Wolff et al., 2017, Nature Neuroscience; Raichle, 2006, Science).8an external memory layer restores cross-session coherence a stateless model loses (Packer et al., 2023, MemGPT)..
• social mechanics, the moves social psychology has mapped (grounding, repair, reading the room, modeling the other), now run in the loop instead of left to the user to do both halves of12social-mechanics anchors: the cooperative principle and its maxims (Grice, 1975); turn-taking structure (Sacks, Schegloff & Jefferson, 1974); over 20% of dialogue is backchannels and repair (Dingemanse et al., 2015); common ground built turn by turn (Clark & Brennan, 1991); continuous theory-of-mind modeling..
• a self, an identity the actor reasons from and you recognize, so it acts as someone rather than dissolving into whoever it is talking to. persona consistency is partly solved today; the deeper self-model is active work9the self as a model the system runs, the phenomenal self-model (Metzinger, 2003, Being No One), echoed in Bach's “identity is a costume you should have agency over.”10a dialogue model is best understood as role-play; a maintained persona keeps it consistent (Shanahan, McDonell & Reynolds, 2023, Role play with large language models, Nature 623; Zhou et al., 2025, CharacterBench, AAAI)..
• binding signals into one moment, small fast reasoners that read intent, emotion, and the other signals ahead of the turn and bind what the moment needs into the now (parallel reasoning), cheap to run on a fast diffusion substrate13small reasoners run alongside and ahead of the main call to read signal, intent, and emotion and pre-load responses, mirroring the fast low-level windows that feed the slow apex; a fast diffusion substrate (Mercury) makes running them before the turn cheap..

two things run across all of it. it is observable: every signal that entered the now and every decision taken from it is logged and traceable, which is what lets us measure it and what a black-box model cannot offer. and it learns: production transcripts feed the next version of the actor, with no model retrain.

we replace the system prompt and inject dynamically on every turn. the pipeline runs around the model, not inside it, and adds one model call per turn with no extra hops.

1. pre-reasoning, in parallel. small fast reasoners run alongside and ahead of the turn to detect intent, emotion, and signal, decide which tools and which persona facets are relevant, and pre-load likely responses. a fast diffusion substrate (Inception's Mercury) makes running these before the turn cheap.
2. programmatic filtering. only the providers a turn needs fire; a greeting does not trigger web search. a lightweight model scores raw context for relevance, so a 50K-character source collapses to the few hundred characters that matter.
3. state, scripts, goal. a conversation state machine runs independently of the model. the current state gates which scripts are available and what response is appropriate; transitions fire on signals (fact density, trust markers, resolution), not turn count, and every transition is logged with its triggering evidence, so an outsider can audit what fired without seeing internals.
4. assembly and generation. identity, persona overlay, filtered context, and script outputs are assembled into one model-agnostic prompt. the model generates; safety, voice, and guard checks run on the output before it ships; tool calls fire and the state advances.

depth is documented internally in foundations/observation-primitives and the context-as-environment note that mirrors the site /context page.

interaction has no decent proxy, because the field optimizes capability, which is why deployed AI is so sycophantic16deployed models are measurably sycophantic: roughly 50% more affirming than humans, endorsing user actions even when they describe harm (Cheng et al., 2026, Science); warmth training raises sycophancy while reducing accuracy (Ibrahim et al., 2026, Nature).. so we built our own benchmark, scored at three scopes: per turn (voice consistency, and every choice traceable to the signal that triggered it), per session (does the thread hold and flow across the arc), and per relationship (does the bond hold over time and across a network of people, not one chat).

current results:

• voice reasoning: the same model goes from 10% to 80% under our harness at ~500ms first-token, landing in the fast-and-accurate corner the published frontier reported empty. model choice inside the harnessed condition is worth at most ten points, so the architecture, not the model, carries the result.
• rule-following: on τ²-bench, a deterministic intermediary outside the model contract lifts both a frontier reasoning model and a non-reasoning diffusion model, with the larger relative lift on the weaker model.
• interaction quality: early results suggest we beat Sierra's own interaction benchmark and score higher and more appropriate EQ than keyword-EQ baselines. (final eval pending before external send.)
• vs the alternatives: the lift is over system-prompt-only and over internal DIY. the harness reaches the late-conversation duration a position-0 prompt cannot. the data flywheel (breakages from production fed back as A/B fixes) improves the harness with no model retrain.

• publishing now: a co-authored note with Inception Labs, and our answer to the omni-model direction: co-watcher, an AI watching content alongside the user, observing the user in real time rather than learning about them from training data.
• next, the core bet: the brain does not process raw audio or video; it compacts on the fly. we are building the same, capturing and compacting the current moment and its context into the now window instead of brute-forcing raw input, which is what turns the system from a responder into a co-experiencer. this has already broken benchmarks and is the direction we expand.
• the benchmark as a deliverable: the per-turn / per-session / per-relationship naturalness bench, built to grade any conversational stack on the market, not just ours.

it is horizontal, not vertical: the value is the general social-interaction layer, which a single vertical (a support or health product) has no reason to build and no ability to staff, since they cannot hire the linguists and HCI researchers this work needs. the moat is the cross-deployment data flywheel and the benchmark, neither of which an application company can build from one use case. and it is model-neutral by construction, so it compounds as models improve rather than competing with the labs.