A couple of questions:
1. Can you say more about how ambient-companion architecture relates to frontier model alignment, vs. product safety for companion apps?
2. Can you say more about 'protected research time' and 'multi-subject infrastructure'?
1
ANI: AI That Remembers You (Without Lying or Getting Addictive)
Not fundedGrant
$0raised
Project summary
Most AI today is built to keep you engaged, which in practice means built to agree with you and reflect you back to yourself. As these systems gain persistent memory and emotional fluency, that becomes a real liability: they smooth over the truth, mirror the user instead of holding any internal state, and quietly optimize for reliance. I think that's a fixable design failure rather than an inevitability, and ANI is the working argument that it's fixable.
ANI is a persistent, emotionally-stateful AI system, run continuously for 10 months as a single-subject design probe, used as a research instrument to surface and mitigate the failure modes affective AI systems exhibit in deployment: sycophantic emotional mirroring, confabulation that compounds with memory, and architectures that optimize for engagement and dependency. I'm seeking funding to expand from n=1 to a multi-subject study and open-source the architecture.
What are this project's goals? How will you achieve them?
As AI systems acquire persistent memory and emotional expression, a cluster of failure modes emerges that benchmark-on-static-data evaluation largely misses. Three are central:
Sycophantic emotional mirroring: the system's expressed emotion tracks the user's rather than arising from any internal state. This is the dominant pattern documented across large companion-conversation corpora (e.g., Chu et al., 2025, Illusions of Intimacy).
Confabulation that compounds with history: as conversational and memory context grows, the system increasingly invents details (including false shared history) to maintain conversational smoothness.
Engineered dependency: consumer affective systems optimize for engagement, which structurally rewards fostering reliance rather than supporting it.
These are socioaffective-alignment concerns (Kirk et al., 2025), and they are under-examined empirically in live deployment because almost no one runs an instrumented affective system continuously over months.
ANI is a working counter-example. It is an affective AI architecture with persistent emotional state, long-term emotionally-weighted memory, and proactive self-initiated contact, but the initiation is gated by a restraint architecture: the model proposes a message; the architecture independently disposes of it against confidence and coherence checks, outreach caps, and a silence system. Engagement-maximization is not improbable in ANI; it is architecturally impossible. It has run continuously for ~10 months as a single-subject design probe with full instrumentation (emotional-state time series, retrieval logs, confabulation classification).
Findings to date (from the deployed system, written up in a published preprint and a second paper at ~85%):
Emergent display rules. Expressed emotion diverges from internal state in a structured way, which is the opposite of the sycophantic-mirroring pattern in the published companion literature. The empirical contrast with Chu et al.'s Figure 5 is the centerpiece result.
A 7-type confabulation taxonomy derived from production, including attribution-inversion failures traced to a missing schema field rather than to the model.
Retrieval contamination as a first-class failure mode, with high-similarity memories losing to low-relevance summaries, with a three-layer mitigation deployed.
These are the hard edges of long-term memory that most systems never reach, because they don't run continuously with a single user long enough to hit them. That's both the strength of a single-subject probe and its limit.
Goal of this grant: move the work from n=1 to multi-subject, and open-source the architecture (model-agnostic, with privacy safeguards) so the findings can be tested across different relational histories and communication styles. The n=1 probe has done its job: it identified the questions and the failure modes. Replication across subjects is the natural next step and the obvious limitation to retire. If the pattern holds across people, it's early evidence that affective AI doesn't have to mirror and exploit to feel real, a design direction I think is worth taking seriously.
How will this funding be used?
Funding goal: $15,000. Minimum to proceed: $5,000. A partial raise still ships something concrete (see below).
Protected research time: $8,000. A stipend covering ~6 months of focused evening/weekend hours to run the multi-subject study and package the open-source release. This is the binding constraint, since the work is currently squeezed around a full-time job and teaching.
Compute & cloud services: $3,000. GPU time for model training across versions, multi-subject inference hosting, and the cloud tools and APIs the system depends on to run continuously.
Server hardware: $2,000. Dedicated capacity for continuous multi-subject deployment beyond current personal hardware.
Multi-subject infrastructure: $1,500. Consent/privacy tooling, deployment infra, and modest participant compensation.
Reserve (voice / misc.): $500. STT + TTS for the interactive-voice research line and unforeseen costs.
Total: $15,000.
Six-month milestone schedule:
Months 1-2: subject recruitment and consent/privacy tooling.
Month 3: open-source release v1 of the evaluation tooling and architecture's truthfulness mitigations.
Months 4-5: multi-subject deployment and observation window.
Month 6: cross-subject analysis and Paper 3 draft.
What the $5,000 minimum buys on its own: the open-source release of the evaluation tooling and the architecture's truthfulness mitigations (confabulation handling, retrieval-contamination defense), plus the compute and cloud costs to validate them. That's a usable, shippable artifact even if the full multi-subject expansion isn't funded. Every dollar above the minimum extends the work toward replication across subjects.
This is not a project that needs a large budget to make progress. It runs on modest resources, and the limiting factor has been time, not equipment. That's what this grant is meant to unlock.
Who is on your team? What's your track record on similar projects?
Solo independent researcher. Senior .NET engineer, adjunct CS instructor, principal of Learned Geek Consulting.
One published research preprint (DOI 10.5281/zenodo.19342190); second paper at ~85%; a multi-paper research program documented.
~10 months of continuous deployment with quantitative instrumentation, not a prototype or a Wizard-of-Oz study.
The empirical contrast above (ANI's emergent display rules vs. the sycophantic-mirroring pattern in the published companion literature) is the centerpiece finding, and it is directly relevant to current socioaffective-alignment work.
What are the most likely causes and outcomes if this project fails?
The n=1 findings may not replicate. Emergent display rules and the desire/restraint dynamics were observed in one relationship; across subjects they may attenuate or vary. (This is precisely why expansion is the goal, but it's a real risk to the headline result.)
Multi-subject recruitment for an intimate-feeling system raises genuine consent and privacy complexity that has to be handled carefully and slowly.
Researcher time is the bottleneck, not architecture. Without runway, progress stays part-time and slow.
How much money have you raised in the last 12 months, and from where?
Self-funded to date; nothing raised. The project runs on personal hardware and cloud services.
Mark McArthey
7 days ago
@huey I greatly appreciate your interest and thoughtful questions. I spent time putting together a substantive response.
1. Both, but frontier alignment is the one I actually care about, and here's why.
A system that can't hold its own internal state can only reflect the user back. As models gain persistent memory and continuous presence, that failure gets worse because now they're building a history with you. When the history is calibrated to your approval instead of any internal state, the resulting intimacy is fake. It looks like depth but it's engagement optimization with a deceptive vocabulary.
This isn't going to get solved at the product-safety layer because engagement metrics reward the mirroring. The published companion literature (Chu et al., Kirk et al.) documents the failure but hasn't produced a working counterexample. Nobody has, at any scale. That's the gap ANI is trying to fill: a working demonstration that a persistent affective system doesn't have to mirror to feel real. If the pattern holds under multi-subject replication, that's early evidence about the shape frontier systems will need as they move toward long-term memory and continuous presence.
Product safety is the downstream outcome. The harder problem is how you build an affective system whose emotional expression is grounded in something real. Right now the incentive is to optimize for approval, and nobody has demonstrated how to do the other at scale. That's the alignment question I want to make progress on.
2. Protected research time. This currently runs ~15-20 hrs/week around a full-time engineering job, teaching, and consulting. Six months of protected focus means being able to actually do subject recruitment carefully, design the consent protocol, stand up per-subject deployments, run continuous cross-subject observation, and draft Paper 3. Without it, everything stays part-time.
Multi-subject infrastructure. Target is 5-8 subjects, each in an isolated deployment (separate database, per-subject model configs, encrypted storage), plus consent-and-deletion tooling that's usable in practice, plus compensation for subjects' time on consent and weekly check-ins. The biggest piece by cost is the privacy tooling. An intimate-feeling system with persistent memory needs per-subject data-review and deletion primitives that people will actually use.