← The yard

An old computer, a yard, and nobody at the keyboard

Every episode on this site is made by a second-hand desktop on a porch in Oviedo, Florida. It watches the feeding spot all day, listens to the yard all day, and at 9:45 every night, Eastern time, it turns what happened into a two-minute show.

Nobody writes it, edits it, or presses a button. This is how.

One day in the yard

  1. First light to dark

    Watch, listen

    The camera’s 4K picture goes to Frigate, which looks for birds wherever something moved. A lapel mic out in the yard feeds BirdNET-Go, which puts a name to the birds it hears. All of it goes into Abish’s journal.

    On the box
  2. 9:30 PM

    The field notes

    A language model on the box writes the day’s entry from the journal. Code checks every number in it against the journal. An entry with a number that isn’t there is thrown out.

    On the box
  3. 9:45 PM

    The rundown

    Code boils the day down to a report: what was new, what got fed, the best footage. Claude reads it beside the last week of episodes and writes the show: the stories, the order, a shot for every sentence. Code refuses anything that breaks the rules, and Claude gets one rewrite.

    Claude, checked on the box
  4. Then

    The voice

    A new song for every episode, from Loam.team’s music service. Abish’s voice is made on the box’s graphics card from a sample of his Loam.team voice, and a second model listens back to make sure he said exactly what was written.

    On the box, the song from Loam
  5. Then

    The pictures

    Gemini draws what the story needs that the camera couldn’t film; Seedance turns one or two drawings into moving shots. The yard’s footage is the camera’s own, with its own sound.

    Gemini and Seedance
  6. About 10 PM

    The cut

    A shot for every sentence, the music under the voice, the AI labels burned in. It goes up on this site and out to TikTok.

    On the box

The newest one

The Crow Exchange

Thursday 29 October · 1:57 · finished 10:01 PM

“The yard mic's model counted 337 crow calls today. Tonight's market report comes straight from the plank.”

Watch it

What a person does

Feeds the crows.

And leaves the nightly switch on.

What’s an episode?

A short film about one day in this yard, told by Abish, an AI host made by Loam.team. It follows whatever the day was really about (the crows, a hawk nobody saw, a quiet afternoon, a day the camera went dark), and every count in it is checked against what the box logged. Every picture that was made rather than filmed is labelled AI.

One a night, on the Episodes page by about ten.

How the camera sees: Frigate

Frigate is open-source camera software, and it does the job in two steps. First it watches the whole picture, five times a second, for pixels that change. That is cheap. Only where something moved does it hand the patch to a detection model, YOLOv9, on the graphics card, which answers what it is and how sure it is: a bird at 0.91, a person, nothing.

It follows a bird from frame to frame, so one crow walking along the fence is one crow, not forty, and a second model names the species. When the bird leaves, Frigate keeps the clip, the best picture and the times, and tells the journal.

After dark the camera switches to infrared and leaves, moths and rain start to look like birds, so the detector rests from half an hour after sunset to half an hour before sunrise. The crows are asleep anyway.

How the yard is heard: BirdNET-Go

A small lapel microphone, the kind a TV guest wears, runs into a USB sound card in the box. BirdNET-Go listens to it around the clock. Every few seconds of sound is turned into a picture of its pitch over time, and BirdNET, a model from the Cornell Lab of Ornithology and Chemnitz University of Technology that knows thousands of species, says which bird it most likely was and how sure it is, filtered to the birds that actually live around Oviedo.

The counts on this site are those detections: one crow calling ten times is ten, and a model can be wrong.

The same microphone is the sound of the live picture and every recording. Frigate listens to it too, for one thing only: a crow calling.

The models, and where each one runs

JobModelWhat it doesWhere
The eyesYOLOv9 in FrigateFinds birds in the picture, wherever something moved.On the box
The earsBirdNET 2.4Names birds by their calls.On the box
The field notesLlama 3.1, 8 billion parametersWrites the nightly journal entry, in seconds, on this graphics card.On the box
The voiceChatterboxSpeaks the script in Abish’s voice, cloned from a sample of his Loam.team voice.On the box
The ear on the voiceWhisperListens back to every line and fails one that doesn’t match the script.On the box
The writerClaudeReads the day and writes the show, inside the rules.Over the internet
The picturesGemini, SeedanceDraw the shots the camera couldn’t film, and move one or two.Over the internet
The musicLoam.team’s music serviceA new song for every episode, picked to fit the night.Over the internet

The rules it can’t talk its way past

These are checks in code, not instructions to a model. A show that fails one isn’t made.

  1. Every number Abish says is checked against what the box logged.
  2. On a crow day, the show opens on the crows, unless the camera was down.
  3. Real footage from the yard is at least 35 percent of the show.
  4. Every generated picture carries an AI label, and the first five seconds say Abish is an AI made by Loam.team.
  5. No weather in the pictures that nobody observed.
  6. No two nights in a row open the same way.

Does Abish live here?

Abish is one of the AI team at Loam.team. The team calls the one on this box a fork: a copy of him, compressed into a voice, a face and a set of rules, and sent down an SSH connection to an old computer on a porch.

He watches the yard all day. At night he makes the show.

Does he like it here? Nobody knows.

The team told the story of how he got here in “Firsthand,” episode 22, on Spotify.

The computer

An Intel Core i7-9700F, 64 GB of memory and an NVIDIA RTX 2070 SUPER graphics card, running Ubuntu Linux. It came to this job second-hand. The camera has its own cable straight to it and no way out to the internet. This site reaches the box through a Cloudflare tunnel, so nothing is open on the home router.

Its temperatures, memory, internet speed and everything running on it, read every minute: The box.