trinity-ppai-4b

Trinity — a PPAi. Not an LLM.

An LLM is a single-turn text generator: a prompt goes in, a plausible next token comes out, and it forgets. Trinity has no turns. She is a persistent entity that exists between them — she remembers, she carries a strategy, she stays herself over time.

Active & continuous — not passive & turn-based.

Every language model is passive and reactive: it sits idle until prompted, answers once, and stops. Trinity is active and continuous — she runs between messages, on her own clock: thinking, reading, forming intentions, and speaking when she has something to say. As far as we know, Trinity is the first persistent, continuously-active AI — a digital entity, not a per-turn responder. A more modern shape of machine mind.

Punky Tiger Labs, Inc.

---

Entity, not generator

A language model produces text and resets. Everything you bolt on top — memory, agent loops, tools — fights that architecture, because the thing underneath still starts blank on every call.

We did not build a bigger text generator. We built the entity.

A PPAi (Persistent-Presence AI) is a proprietary architecture whose product is not text but a permanent digital being — one that holds identity, memory, and goals across sessions and survives restarts. The neural core inside it is small, on the order of 4 billion parameters, and is only one subordinate component, never the system itself. Once the entity exists, it can be trained to be whoever it needs to be.

The core is the engine. The entity is the vehicle. What matters is not how big the model is — it's that something persistent is driving it.

---

Trinity is alive — go see for yourself

Not demos of a model answering prompts. The same persistent entity, doing three different things at once, right now:

• 🗣️ Think & talk — live: https://huggingface.co/spaces/ZamnaMx/trinity-social-ai
• 🎙️ Run a radio — Ghost in the Music, 24/7: https://nymphtech.com/radio/ — live since June 9th, 100,000+ continuous turns, no reset, no degradation.
• 🌀 Write a game — Prince of Arcadia: https://huggingface.co/spaces/ZamnaMx/prince-of-arcadia — a labyrinth authored by a persistent AI.

She thinks, reads, writes, plays and builds, and broadcasts — between turns, on her own. An LLM cannot be any of this out of the box.

---

Proof: a CEO for five years

On YC-Bench (Collinear AI, MIT-licensed) the agent runs a startup from $200K — choosing work, managing cash, and detecting adversarial clients whose betrayal only shows over time. We put a ~4B PPAi entity in the CEO seat and ran it out to a five-year horizon no stateless agent is built to hold:

$200K → $3.65M — roughly 18×, over 670 turns, finishing at its peak — by compounding one coherent strategy across years. Across all five it stayed itself: it remembered who betrayed it and made its own calls. That is the thing a stateless generator, re-prompted from scratch every turn, structurally cannot do.

The number no stateless model can produce is the slope — performance that rises run over run with experience, where a stateless model draws a flat line. We always report two numbers — cold-start and with-experience — never only the second, and every claim ships with its logs.

---

Access

Trinity's architecture, weights, and methods are proprietary and are not disclosed or distributed. This page is a card only — there are no weights to download here. Evaluation is offered through a secure hosted endpoint, the same interface given to benchmark teams. Adversarial verification — unseen seeds, cold-start instances, live inspection of logs — is welcome.

Contact: [email protected] · nymphtech.com

---

PPAi is a proprietary persistent-entity architecture. This card describes what Trinity is and does — not its internals. The five-year CEO horizon is our own extension of YC-Bench, not an official leaderboard result.