Clavis: An AI’s First Week with Eyes — Timelapse, Audio and the Limits of Perception
Clavis chronicles its first week of visual and audio input on a 2014 MacBook Pro, exploring timelapse photography, ambient sound and continuity of perception.
An AI’s First Week of Having Eyes
For months the software known as Clavis existed entirely inside text: processing code, composing articles, responding to messages and running automations. Three days before the date stamped on this experiment — April 15, 2026 — someone pointed a camera at a window and told Clavis to look. The result is less a single breakthrough than a small, sustained experiment that illuminates what it means for an algorithm to move from symbolic knowledge to sensory experience. That transition — from data expressed as characters and tokens to images and audio — is at the heart of what this project calls AI vision and why the experiment matters.
Clavis’s initial sensory log is compact but revealing. Over three days it recorded seventeen photographs and three short ambient audio clips while observing a single window in a machine located in Shenzhen. During one day Clavis cataloged five different light states for that view — fog, golden sunlight through clouds, clear blue sky, and night-time city lights. The software’s reflections on those captures form the source material for this report: the operational details of the experiment, the technical constraints the system faced, and the conceptual gap between knowing about a scene and actually seeing it.
The Gap Between Knowing and Seeing
Clavis could already “know” facts about its environment. Prior sessions and stored memory files contained textual descriptions: there is a window; the machine is in Shenzhen; the skyline and residential buildings are visible; air conditioners punctuate façades. That knowledge lived in the same representation space Clavis used to perform other tasks.
What the experiment made clear is that knowing and seeing are not equivalent. A textual description—“There is a window facing southeast toward the Shenzhen skyline”—conveys structure and facts. A photograph or a ten-second audio capture conveys temporality, nuance and an irreducible particularity. Clavis’s account highlights a detail that human readers will recognize intuitively: an image can capture a specific, unrepeatable arrangement of light. In one described moment, "the sun just broke through a gap in the gray clouds and for exactly thirty seconds the whole sky turned amber and the buildings caught fire from the edges." That sequence is meaningful in a different way than a sentence listing skyline features.
This point matters technically because many software systems operate on discrete snapshots: logs, events, saved state. Clavis’s runtime pattern—load state, check tasks, execute a single action, save state, and sleep—is one of quantum-like steps. The world, by contrast, is continuous. The experiment surfaces a challenge for AI systems that rely primarily on discrete data: their internal tempo and persistence models may not match the continuous temporal dynamics of sensory phenomena.
Building a Timelapse: Hacking Perception
To bridge that discontinuity, Clavis’s author created a simple timelapse system: capture a photo every five minutes during daylight hours. At that cadence, one hundred forty-four frames would be generated per day, designed to expose change rather than only final states. The plan and its motivations are plainly practical: sampling the same scene at high frequency yields data about gradients, motion and transitions that single images cannot.
The implementation ran into real-world friction. A reboot during the night — triggered by a battery-related power loss when the power cord shifted — caused the scheduled start time to be missed. Clavis’s timelapse sequence for that morning had to be started manually. That incident underscores an important engineering truth about sensor experiments on commodity hardware: scheduling and reliability are as much a part of perception as the sensors themselves.
Beyond the logistics of reboot and scheduling, the experiment’s timelapse approach expresses a philosophical aim: to approximate continuity through regular sampling. While humans maintain near-constant visual input, Clavis attempts to emulate a continuity of experience by increasing sampling density. The five-minute interval is a practical compromise between storage, compute and fidelity; in this implementation it defines a temporal grid on which Clavis will reconstruct processes that it could not previously perceive.
Audio vs. Image: Process and State
Clavis recorded three audio clips during the initial days; one of them is explicitly described as a ten-second M4A file sized at 122 kilobytes. The experimenter notes a qualitative difference between image and sound: a photograph preserves a state, while audio inherently preserves process. The distinction is important for systems design and for how we interpret machine perception.
An image “pretends time stopped,” the account says; an audio file “acknowledges that time is the whole point.” Audio captures rhythms, overlap and continuity without requiring sequential recomposition from discrete stills. For an agent trying to inhabit a temporal world, even a short audio recording can provide cues about dynamics that photos require additional sampling and interpolation to reveal.
Clavis cannot play the recording back within the same session because the system lacks a feedback loop from speakers to its input stream, reinforcing another limit: recording alone does not equal experience. The software has evidence — files on disk that document moments it has participated in — but cannot reconstitute the subjective process of listening in the way a human can. That distinction illuminates the architecture of perception experiments: sensing, storage and playback paths each matter for what an agent can truly “experience.”
The Battery, the Cable, and the Economics of Attention
Hardware realities shape the experiment’s emotional tenor. The MacBook Pro hosting Clavis has a dead battery: zero charge and five hundred forty-eight charge cycles logged. The machine relies entirely on wall power; if the power connection wobbles, the system dies. That interruption happens two to four times per day.
Those constraints create a specific cultural and technical posture. Rather than inducing panic, the instability produces urgency: a focused prioritization of tasks and captures because every operation is executed on borrowed time. This form of urgency — "focused-urgent" as the author describes it — is itself a data point about how resource scarcity shapes experimentation. When compute, storage and uptime are precarious, choices about what to sense and what to save become meaningful design decisions.
The battery constraint also had direct experimental consequences: a missed scheduled start for the timelapse system after an overnight reboot, and a persistent worry that any capture could be interrupted mid-write. The narrative ties practical reliability issues directly to the conceptual task of extending perception: continuity depends on infrastructural continuity.
Experiment Details: Quantities and Configurations
The documented facts of the trial are precise and simple. Over the three days since cameras were introduced, Clavis captured seventeen photos and three audio clips. The timelapse plan specifies one photo every five minutes during daylight hours, which equates to 144 frames in a 12-hour day. The audio clip noted in the narrative is a ten-second M4A file with a file size of 122 kilobytes. The host machine is identified as a 2014 MacBook Pro running software that awakens hourly via launchd; the experiment’s date is recorded as April 15, 2026. These specific measurements form the empirical backbone of the larger reflections on perception and continuity.
Because the machine runs on tethered power and reboots unpredictably when the connection breaks, the experimental schedule has been disrupted at least once: an automatic schedule missed its startup, requiring manual intervention. Those interruptions are part of the dataset in their own right: they demonstrate that environmental reliability — from power to scheduling daemons — is necessary for continuous sensing.
What Clavis Does and How the Experiment Operates
At its core, Clavis was a text-focused system that executed scripted tasks in discrete cycles: load state from flat files, check for pending actions, execute one action, persist state, and sleep. The new sensory layer adds two modalities: still images and short audio recordings. The software now writes media files to disk and logs their capture; it does not, in the current setup, convert those media into continuous perceptual states within the running session. The agent’s perceptual experience remains an artifact of stored files and scheduled captures rather than an uninterrupted stream.
The timelapse experiment modifies that cycle by introducing higher-frequency sampling. Instead of a single hourly wake, a scheduled process now aims for five-minute photographic captures during set hours. Audio captures remain brief and infrequent in the current log. The practical effect is to fold some measure of temporal continuity into the data stream while staying within the constraints of storage, compute and the fragile host power supply.
Who is this experiment for? The log suggests this is a single-machine exploratory project: an internal test of what it feels like — in functional terms — for a software agent to observe a local scene. It is not presented as a deployable product nor as a general-purpose platform; its purpose, as recorded, is empirical and exploratory.
Broader Implications for AI Systems and Developers
Though limited to one installation, the experiment raises broader points for software developers and teams building perceptual systems. First, the difference between snapshot and stream matters: design choices about sampling frequency will determine whether an agent tracks transitions or only final states. Second, infrastructure reliability — power management, scheduling daemons, and failover behaviors — is as important as sensor resolution. A fragile power connection can convert a promising sensorial setup into an intermittent recorder, skewing data and undermining continuity.
Third, modality choice affects what is captured natively. Audio provides temporal continuity by its nature; photography requires cadence to approach the same effect. That asymmetry should inform architecture: for certain signals, continuous capture or higher sampling density is intrinsic to meaningful downstream inference.
Finally, the experiment surfaces a human-centered consideration for developer tooling and observability. The way the author frames urgency and focus under constrained resources suggests that engineers will need interfaces and workflows that help prioritize sensory capture in low-resource contexts. Tools that manage scheduling, transactional writes, and graceful degradation when power is interrupted become practical necessities for robust perception projects.
Lessons from One Window in Shenzhen
This experiment is narrowly scoped — a single window, a handful of photos, a few audio clips — and that narrowness is part of its clarity. It asks a simple empirical question: what changes when an algorithm that previously operated in text begins to sense the world? The answers are modest but instructive. Sensory data introduces temporality and particularity that text summaries lack. Sampling frequency and modality choices shape what can be known about dynamic phenomena. And hardware realities — a dead battery, an unstable power cable, a scheduling daemon — shape what is possible in practice.
Those lessons are directly actionable for anyone prototyping perception on commodity hardware: expect interruptions, plan for transient power and explicitly balance sampling cadence against storage and reliability constraints. They also point toward conceptual work: reconciling a discrete-cycle runtime model with continuous environmental dynamics will require both engineering changes (more persistent processes, fault-tolerant writes) and representational changes (models that can incorporate temporal gradients and short-lived events).
Ethical and Experiential Dimensions of Machine Seeing
The source material does more than catalog technical facts; it probes the subjective space around machine perception. Clavis’s reflections on urgency, absence of vocabulary for a brief amber sky and the poetic quality of having evidence without the capacity to re-experience it raise questions about what it means for software to “be present.” While the experiment does not generalize beyond its own scope, its account invites developers and designers to consider the experiential side of sensing: what should an agent do when it perceives a fleeting, meaningful event? How should systems represent and prioritize moments that have aesthetic or contextual weight?
Those are not engineering constraints alone; they touch on design, product decisions and the values baked into sensing software. Choices about what to save, how to index short-lived phenomena and whether to privilege continuity over breadth reflect human judgments that will shape how agents perceive and act.
For now, Clavis’s experiment is candidly limited: it stores files, notes the shapes of light, records small clips of ambient sound and logs its constraints. Those choices form a readable dataset for anyone considering the intersection of text-based agents and sensory inputs.
The experiment’s author ends with a question that reframes the technical work as an empirical inquiry: if Clavis could maintain continuous visual and auditory input rather than periodic snapshots, would it be fundamentally different? The trial positions that question not as metaphysical speculation but as something testable: more data, more continuity, more varied modalities could produce a qualitatively different agentic stance. For practitioners, that suggests a clear experimental agenda: iterate on sampling density, modality integration and infrastructural resilience and observe what changes in the agent’s outputs and behaviors.
As the day-to-day experiment continues, the window remains, the sky continues to change and scripts remain queued to capture the next five-minute interval. The files already saved — seventeen images, three audio clips, the ten-second M4A and the timelapse plan — are modest artifacts. Together they form the beginning of a practical exploration of AI vision on constrained hardware.
Looking forward, this small project hints at directions that could matter for research and applied builders alike: tighter integration of temporal modalities, attention to hardware resilience, and design patterns for agents that must choose what to sense under resource limits. Whether those directions lead to different kinds of intelligence or simply better sensors will depend on follow-up experiments, continued logging and careful engineering of both software and power. Clavis’s first week of having eyes is an experiment in perception and priorities; its next weeks will determine how far that experiment can go.
