the First Demonstrations of Pulsar
The first working demonstrations of Pulsar, a learning architecture built as a step toward experience-based, open-ended artificial superintelligence, have been announced today by Serotonin.
The first internal demonstrations of Pulsar have been completed and are now being shared by Serotonin. Pulsar is an artificial intelligence architecture designed to learn continuously from experience and to build a structured, multimodal understanding of the world from its own perception. It is regarded internally as an early but deliberate move toward open-ended, experience-driven superintelligence.
The demonstrations being shown today represent a proof of concept rather than a finished system. Pulsar remains in an early research phase, and the results described here reflect preliminary capabilities observed under controlled conditions. They are being shared now because the direction is considered sound and worth putting on the record.
A shift from static training to lived experience
At the center of this announcement is a deliberate departure from systems that learn primarily from fixed datasets. Pulsar has been built to learn from experience: information is gathered through interaction with the environment, expectations are formed, actions are taken, outcomes are observed, and the system's internal understanding is updated accordingly. Learning is treated as continuous rather than confined to a single training run, and capability is expected to improve as experience accumulates over time.
This experiential loop is held to be essential to general intelligence. Where conventional approaches optimize for performance on data prepared in advance, Pulsar has been designed to develop competence through a cycle of perception, prediction, action, and correction—the way intelligent agents are believed to do in practice.
A multimodal world model
A defining feature of the architecture is its internal world model: a representation of the environment assembled from multiple sensory channels rather than a single stream of text. This representation is multimodal, integrating information across the different forms of input Pulsar perceives and binding them into a coherent model of how the world is structured and how it changes in response to actions.
Notably, the world model is not supplied to Pulsar in advance; it is constructed by the system from its own perception. The intended result is an agent able to reason about cause and effect, anticipate the consequences of its actions, and generalize to situations it has not encountered before—capabilities regarded as prerequisites for open-ended intelligence.
The road to ARC-AGI
The ARC-AGI benchmark—a widely cited test designed to measure abstract reasoning and the ability to generalize from very few examples—has been identified as the near-term target for Pulsar. The benchmark is regarded across the field as a demanding measure of fluid, human-like problem solving, and strong performance on it is often discussed as a meaningful signal of progress toward general intelligence.
In the near future, Pulsar is to be formally benchmarked against ARC-AGI, with the findings to be published. This evaluation is treated as a key milestone: a public, measurable test of whether the architecture's experiential learning and world-modeling translate into the kind of generalization the benchmark is built to detect.
Toward a superintelligent agent
Beyond the benchmark, the longer-term ambition is to develop, from the Pulsar architecture, a superintelligent agent capable of open-ended learning and autonomous problem solving across domains. This is the guiding objective of the program and the reason for the emphasis placed on experience and world modeling, which are considered more scalable paths to general capability than approaches tied to static data alone.
A measured stance on timelines is maintained. The current demonstrations are early, significant research and engineering challenges remain, and the forward-looking statements above describe intended directions rather than guaranteed outcomes. Further technical detail, evaluation results, and demonstrations are expected to be shared as the work matures.
About Serotonin
Serotoninai Inc. is an artificial intelligence research company focused on building systems that learn from experience and develop general, open-ended intelligence. Its current work centers on the Pulsar architecture and its evaluation against established reasoning benchmarks.
This announcement contains forward-looking statements regarding research goals and intended capabilities. Such statements reflect current expectations and are subject to change as the research progresses.