#2026-M07

---

I spent the last week of July at PIFcamp in the Soča river valley in Slovenia, where I took the work-in-progress Antennae. So today's Update is mostly a report on what worked, what broke, and how it all glowed in the end ,)

---

The Body

Before leaving for Slovenia I still managed to prepare the body of the object. The shape builds on what I had already tried with the Synapse installation. A low body at the ground, with light-guiding antennae rising out of it. On top of that I ordered a pile of hardware, sensors and the optical PMMA rods themselves, which glow along their full length when lit from below.

The hardware box I printed from carbon fibre PETG, a material I have come to like for its beautiful surface and stiffness. Of course, in the rush I made plenty of mistakes in the first design and had to widen or add some of the openings. To the bottom of the object I fixed a weight, so it could not be tipped over easily. Next time I want the box a bit bigger, simpler to open, and finally waterproof too.

A few parts I printed from the photoluminescent PETG Ultraglow, the idea being that the object charges in the sun over the day and then glows for the "advertised" six to eight hours after dark. In reality, though, the sky darkens rather slowly. The material gives off light through the whole dusk, so by the time it is properly dark, most of the reserve is gone and the glow does not last long. Talking with the other participants, we hit on another way to absorb light even at night. Since the object emits light itself, some of that energy can be continuously absorbed back. The material then returns the light in its phosphor green, which is something I would like to explore further in the next iteration.

---

Solar !

For the first time I worked with a solar panel, a charging module and LiFePO4 cells. Charging ran reliably even with three cells instead of the planned four, but the weak point was consumption. In practice the object drew a lot even at rest, and with the high-power LEDs on, the draw climbed further still. The cells drained fast and would certainly not have lasted a whole night of running. In the next version I want to raise the capacity, program a more efficient behaviour, and finish proper mains power, so a permanent installation could run without trouble.

The high draw was also behind the microcontroller restarting on its own during testing. When all the light sources came on at once, the input voltage most likely started dipping because of the charger's current limit. Luckily, wiring the step-up converter straight to the battery and adding a capacitor solved it.

---

Sensing

The little mmWave radar was a great joy. One of the few sensors that can read through a wall. It picked up the presence of people (or animals) reliably over several metres, and in a camp full of people it basically always had something to read. I also managed to wire up further sensors for temperature, humidity and light intensity. For now they did not shape the object's behaviour much, but the data flowed the whole time and in future I want to work with them more.

The current perceptual system:

  INPUTS
  geophone . . . soil, ~1 m
  piezo    . . . touch, surface
  radar    . . . presence, m
  climate  . . . temp, humidity, lux
             |
             v
      +--------------+
      |   ESP32-S3   |
      +--------------+
             |
             v
  OUTPUTS
  antennae . . . 2x PMMA rod, amber
  roots    . . . LED filaments
  speaker  . . . tones and drones
  network  . . . telemetry (OSC)

  intensity -> pitch and loudness
  proximity -> brightness and tempo

photo / Katja Goljat

---

Sound

The geophone and the piezo were probably the hardest part. At first I tested them through an audio interface with its own preamp, and after denoising it sounded quite good. Wiring my own op-amp was more demanding, and I never got that clean sound out of it. In the end I took the audio signal from the ground primarily as intensity, and in code just translated and interpreted it into new sounds through tones, drones and more.

The geophone worked beautifully once it was a little under ground. It read footsteps and a touch of the ground reliably, with the intensity shifting the pitch of the tones, the loudness, and the light in the optical fibres. The piezo I did not use that much in the end, even though it picked things up decently on its own. Next time I plan to stick it straight onto the body of the object and read its own vibrations. A touch on a fibre or trembling in the wind carries through the body and can be read directly, with no capacitive sensor. The object would then be listening to itself, which actually suits it much better (a similar logic to how Host reads the vibration of its servos).

---

Network

The object runs its own WiFi (AP) and streams telemetry non-stop, so I could connect to it at any moment and read what it was doing and what was happening around it. I built a web interface for it where everything could be tested and watched. On top of that it can take new firmware over the air, with no need to pull the controller out of the enclosure. In future I would like to see dozens of such units in the landscape, passing information between themselves and forming a shared technological (eco)system.

---

Under the Tree

After dark I set the object up under a tree and added a length of LED strip nearby, so the place would draw attention to itself. I do not want that in the future, but for one evening it served well. The radar meanwhile measured how far people were standing, and whoever came closer set the light circling the tree faster, whoever stayed back left it just slowly pulsing. The light the object gives off is deliberately amber, since it is bearable deep into the night and attracts a minimum of night insects.

In the end it once again stood on interaction. The installation kept pulsing, and whoever came closer, touched the ground or moved around it could watch the object's behaviour change and see how the system interprets their input. Even when nobody was interacting with it, though, the geophone would register occasional tremors in the soil, and the object naturally responded to that too. Whether that was a real input or just stray noise, we will leave for another time. For now I am glad the object made it through testing and is ready for the next phase.

photo / Katja Goljat

---

The whole week was one of a kind, mainly thanks to the people. A group came together that bends technology its own way, gives it new purposes and, above all, keeps making things. I am grateful for that community dimension, because it managed to kick off further shared projects right away, and I am looking forward to those.

---

Resources

• HOW TO SEE LIKE A MACHINE: Trevor Paglen / on computer vision and visual perception
• I have started developing my own hardware and software for driving LEDs, and so far it looks very promising. If the underlying method interests you, it is described in this article by PJRC. More soon!
• J-Space (Anthropic) / Anthropic described a global workspace inside a language model, one they call J-space. It is an intermediate computation written in words that the model keeps for every position in the text. An interesting step in interpreting language models.

#2026-M06

---

Sound & Image Lab

I started a wider piece of research into how sound can be turned into image. It is a problem that has interested me since my very first interactive project, Zvuk. I went back to some of the oldest inspirations and ways a sound signal can be carried over into something visual. To keep it from being arbitrary, I tried to collect approaches grounded in real physical phenomena, such as Faraday waves on a liquid surface, Chladni figures, photoelastic refraction, the optics of breath, the oscillographic trace of a signal, and more. I put it all on the Sound & Image Lab site, where you will find over fifty such approaches side by side, from a clean line through texture to a full surface. Everything should react to a live audio input once you allow it in the browser.

Sound is not only about loudness; the signal itself carries far more useful data, whether it is a clean tone or noise, its pitch, when a hit lands, rhythm and onsets. The program also reads the shape of the waveform itself, so with the oscillographic drawing, for instance, what you see is literally what the microphone just picked up.

---

Tykadla / Antennae

Last month I tried to sketch three basic positions an "electric organism" can take. The Visitor that gets to know a place, the Inhabitant that just persists, and the Translator that interprets. I am now mostly building the third of them. The object carries the working name Tykadla (Antennae) and is a static organism meant to stand directly in the landscape. Each unit is 1.5 to 2 m tall and is made of two vertical light-guiding antennae rising from a low body. It also has notional sensory roots, interchangeable probes that lend the machine a sensory reach into the surrounding ecosystem. It reads bioacoustics in the soil, contact vibration on nearby surfaces, humidity and light, and possibly chemical or spectral inputs.

Antennae work as a tool of translation. They pick up what happens at frequencies and on timescales people cannot perceive directly, and carry it over into a slow channel a visitor can notice. Most of the time the unit is quiet. It logs what it hears to a local archive and shows only a faint sign of life. A small radar in the body reads the surroundings, and once someone comes close the channel opens: the object brightens and starts to sound according to what the sensory roots are picking up at that moment. So it speaks to the outside only when there is someone to hear it. The object follows directly from two earlier works (Receptor and Synapse) and pushes that line out into the landscape.

Sensing

Again I am building it all as a platform, so the final perceptual system will depend on a lot of experiments straight out in the field. If something occurs to you that would be interesting to watch outdoors like this, get in touch. For now I have split it, on paper, into 3 basic layers;

Below the surface. A geophone, a microphone made to pick up vibration straight in the ground, to record the full soundscape under the surface. Whether it is the life of underground organisms or the vibrations that sound down there.	On the surface. A piezo disc, fixed straight onto the bark of a tree or its roots, or other forest growth. Over a long span the object can follow a tree coming apart from inside, wood swelling, shrinking or cracking, and any signs of other organisms.	In the air. Humidity, temperature and light. A very accessible signal and data that shift dynamically over time and can steer the object's behaviour. On top of that I want to use the radar to read the surroundings, so it knows when a person (or maybe a larger animal) comes close to the antennae.

Hardware

                  ||                          ||
                  ||   2x light-guiding       ||     amber, glows full length
                  ||   rod (antennae)         ||     = internal state
                   \\                        //
                    \\                      //
                     +------------------------------+
                     |   ESP32-C3 / Teensy (brain)  |
                     |   24 GHz radar    presence   |  <- a few metres
                     |   amber LED       light      |   two outward channels:
                     |   speaker         sound      |   light + sound
                     |   microSD archive            |   logs even while quiet
                     +------------------------------+
                        ^ solar 5V -> LiFePO4 charger -> 4 cells
                                        |
                          sensory roots / connectors
                                        |
              _______________________________________________________
              |                        |                            |
      +--------------+         +--------------+          +---------------------+
      |   GEOPHONE   |         |    PIEZO     |          |      CLIMATE        |
      |    (soil)    |         |    (bark)    |          | humidity/temp/      |
      +--------------+         +--------------+          | light (air)         |
       a hand's depth             onto bark              +---------------------+
       in topsoil               wind + chewing            damp and dry

The brain is an ESP32-C3, or possibly a Teensy, with a microSD card for the archive. Power runs from a 5V solar panel into a LiFePO4 charger and four cells; without sun it should hold for up to three days. The light is amber and shines through the two light-guiding antennae (PMMA rods, acrylic seeded with scattering particles, so a rod glows along its full length, not only at the tip). Amber sheds the least of the blue that pulls in night insects, which counts for something in a valley that is meant to stay dark. Next to the light, the second channel is sound: a small speaker in the body plays a quiet sound field built from what the roots are hearing, and at a gathering it can be amplified live. Presence is watched by a 24 GHz radar, the sensory roots plug in through individual connectors, and the printed parts are PETG-CF for strength. On top of that I am thinking of printing a few elements in photoluminescent PETG (Ultraglow), which charges off daylight and then glows faintly on its own after dark, no power involved.

The object should reach the field at the end of July, so I hope that next time I can write with fresh experience. The aim is to bring the unit to the point where it works on its own. It charges from the sun, goes without maintenance over the long run, and speaks to the outside only when there is someone to hear it. At the same time it can keep recording what it hears in the soil, building a local archive. In time I would like to grow the project into a network of several objects in the landscape that pass information between themselves, forming a shared technological (eco)system.

---

Second Brain Builder Update

I rewrote the building prompt for my Obsidian Vault + LLM system again. As you may know, Anthropic released the Fable 5 model, so I used it mainly to improve the whole system. Working with memory, context and data is a bit better again. Lately I have had good results working directly in Max, TouchDesigner, and even Higgsfield for visualisations, through MCP and an agent inside the Vault.

#2026-M05

---

I keep trying to develop a research line of my own, "electronic organisms" inside natural ecosystems. The aim is to build some theoretical frame for how to take my interactive objects outdoors, and how such a machine can work at all. Although it's still early, I'm trying to organise it into a structure and iterate on it step by step. *Hopefully it won't look like I'm romanticising machines (again)... *🤖

Technology in nature and the usefulness of machines

The very idea of why we should put technology into the landscape at all is a bit naive. But technology has long been out there in nature, we just call it infrastructure (transmitters, probes, high-voltage lines, or night lighting). So the question isn't whether technology belongs in the landscape. It's already there. It's more about what exactly does technology add out there? And that's exactly why, for a long time, I looked for various places where a machine could even directly help an ecosystem.

But the deeper I went, the clearer it got that most of these natural processes have long been well optimised. Mycelium connects with a root system and trades nutrients with it across generations of trees. Earthworms move tens of tonnes of soil per hectare a year for free. Fungi do break down an infested spruce over whole decades, but that slowness is actually fine. Any acceleration a machine could add always carries its energy cost, along with the risk of disrupting diversity and contaminating the substrate.

Placing one more such technology into the landscape usually adds nothing to biology, but it can give something to us. A machine in nature can be useful, though not necessarily by reproducing a system that already works well. I don't mean to give up the ecological reach of the work, more to clarify the project's direction.

---

Related machines

• Gilberto Esparza, Plantas Nómadas - an autonomous robot roaming a polluted river
• Ian Ingram - robots that pick up the signals of wild species and translate them
• Ken Rinaldo, Autopoiesis - A-Life robotic sculptures that change their behaviour and "communicate" with each other
• Ivan Henriques, Symbiotic Machine - an autonomous solar machine that feeds on algae
• Edward Ihnatowicz, The Senster (1970) - a shy cybernetic creature

---

What the machine means

I start from a shared premise: a machine built around defending (maintaining) its own conditions (equilibrium) has something inside it can lose. And that is what steers its behaviour. So-called precarious autonomy (Hans Jonas) says that even a very simple system that holds its own boundaries begins to have something like an interest in its own persistence, a stance toward the world. The machine isn't alive and feels nothing. It's enough that it has an inside with boundaries and a stake in keeping them.

A machine with this kind of autonomy has its own Umwelt (Jakob von Uexküll's term). The world as it appears to it. Made of the variables it senses for itself (battery voltage, air humidity, vibration in the soil). But the machine doesn't read the landscape through sensors alone, it reads it through its own body too. It knows how much energy it spends in a day, how long moisture takes to dry on its body, or what it touches while exploring. All of it is data about the place the machine is embedded in. What it comes to know about its surroundings is the robot's own diary, inseparable from the place it stands in.

These objects aren't out to fake life. They're more about raising questions, because their behaviour really does sit on the edge. The same robot, when a human drives it, reads as a tool; the moment it acts from its own feedback, we read it as an actor (see System), without anything in it physically changing. Take xenobots, tiny clusters of frog cells assembled by an evolutionary algorithm, now under study. They heal themselves, they can sweep loose cells into a copy of themselves and move collectively, with no one programming them step by step. These are the traits we denied "machines" for decades, because we built them to be modular and predictable. Not because a machine couldn't, in principle, manage it. Čapek's robot, after all, was never made of non-living matter.

The machine's resulting behaviour is still authored, because I design it. What's emergent and unexpected is read into it by the observer. A mechanical machine on its own knows nothing and strives for nothing. Personality, care, or intent are always projected onto the machine by a person.

---

Three poses

To anchor all of this, I've defined three typological poses such an electronic organism can occupy.

  MACHINE IN THE LANDSCAPE: THREE POSES
  
  A · VISITOR      explores, maps       mobile, seasonal        → map of experience
  B · INHABITANT   idle work            stationary, slow        → leaves an artefact
  C · TRANSLATOR   listens to soil      static, continuous      → translation, attention
  

---

A | Visitor

A mobile, surface-dwelling electronic organism. Conceptually it follows most closely from the hexapod Host and pushes it outdoors. The Visitor does not work the land. It doesn't plant, inoculate, or irrigate. It moves, explores, scans. On its back a LiDAR and a camera, in its body a piezo, sensors and probes. It collects data continuously and carries off a map of the landscape in a form no one else captures. A map of experience: where it stood longer and where it passed faster, where it detected movement and where it was alone. An archived testimony of time passing and of shifts in climate and vegetation.

Its behavioural architecture is homeostatic. The Visitor tracks internal states (battery level, internal temperature, signal quality) and its behaviour follows from how it defends them. Its path through the forest is a by-product of the drive to keep itself going. A human observer can't tell whether the machine is drawn to the light or to a tree. They only witness another way of being in the landscape and persisting in it.

---

B | Inhabitant

A disposed machine that in the landscape just is. It has no task, no measurable benefit for the ecosystem. A slow, small object that gradually grows its own ritual. In the morning it shifts a few centimetres to follow the sun, at noon it hides back in the shade. Sometimes it scratches a mark into the soil, other times it gathers a few twigs and arranges them into a structure I never gave it, one that means nothing in itself. It leaves behind countless artefacts of its own.

This pose is entirely useless, and that's the point. From the ecosystem's side nothing dramatic changes. But we see a machine that takes care of something, even though neither of us knows what. A kind of idle work. Organised activity that runs and consumes without ever turning into use.

In hardware terms it's an object with a tiny power draw and a handful of simple axioms (follow the light, retreat into shade, avoid water, save energy). This is also the only one of the three poses where it makes sense to let the behaviour evolve freely. The software can rewrite itself over time. The evolution serves no external task, it just reshuffles an empty gesture endlessly, and that emptiness is the point.

---

C | Translator

A static vertical organism, 1 to 2 m tall. It has two vertical light-guiding antennae that communicate through light or sound. It also has sensory roots, sort of interchangeable probes giving the machine sensory access to the surrounding ecosystem. It perceives bioacoustics in the soil, contact vibrations on surrounding surfaces, humidity and light conditions, and possibly chemical or spectral inputs.

The antennae work as a tool of translation. They capture what happens at frequencies and scales a person can't perceive directly. The soil sounds below the loudness threshold our hearing can reach. In the band of roughly 20 Hz to 8 kHz an activity unfolds underground that the antennae render into a perceptible medium through a synthesis of sound and light. (springtails moving in the litter, the peristalsis of an earthworm, a bark beetle chewing its gallery in the inner bark of a dying spruce). At the same time it can record continuously what it hears in the soil, building a local archive of the place.

The unit runs on its own. It charges from the sun, goes without maintenance for long stretches, and speaks outward only when there's someone to speak to. The machine is entrusted with a delegated attention a person probably wouldn't have the patience for. It senses constantly and opens a channel whenever someone passes by. The antennae brighten and the speaker begins to synthesise the sound of what it hears underground. Over time the object adapts to the rhythm of its particular place, to its daily cycle, to humidity and temperature, or to the sound of life beneath the soil.

"Antennae"

The type C, Antennae, is the furthest along in development right now. I've managed to sketch out the hardware and its basic functions. I plan to develop it out in the field this summer at PIFcamp in the Soča valley, at the end of July. The project also follows naturally from Receptor and Synapse. In time I'd like to grow it into a network of several objects in the landscape, passing information between each other and forming a shared technological (eco)system.

---

SIDE-NOTE: An audience beyond the human species: There's still another angle to look at this from. What if the machine didn't communicate with us, but instead deliberately gave something back to biology? A spectral broadcast addressed to non-human perception, narrow-band LEDs in amber 590 nm tolerated by bats, near-UV 365 nm that draws moths, polarised channels for certain beetles. In that pose the machine would choose an audience purely beyond the human species, and the landscape would be its real addressee. For now I'm only speculating, but it would be interesting to explore this direction too. Perhaps in the future, with someone who knows these species.

---

Research Log

• Active Sensing Subserves Task-Level Control - active sensing doesn't come from a drive to minimise uncertainty, it falls out directly from controlling the machine's own behaviour. Movement is the substrate of perception, not a by-product of it.
• Sensing Intelligence as a Trainable Metamaterial Property - a metamaterial body's shape can be trained to do part of the perceptual work itself. Up to fivefold more accurate sensing, or nearly an order of magnitude fewer sensors.
• Exploratory Experience Shapes the Geometry of Predictive Representations - mice and agents that explore the landscape build a more organised internal map. The moment they head straight for the reward, the map falls apart.

---

First litograph from Synthetic Archive.

#2026-W19

---

Host T1 Build

I’ve finally completed the Host Tier 1 build I’ve been working on for the past month. I managed to reprint most of the parts and reconfigure the hardware so I wouldn’t have to rebuild the entire machine from scratch, so I’m still using the original board and components. Before I try to make my own PCB, I wanted to create a prototype version to fine-tune the last few issues in the overall build.

A rare glimpse into the internals of the machine

The biggest change is actually those barely noticeable new feet and their force sensors (FSR). I had to modify the design of the feet to route the cables, as well as the shape of the tips and the foot itself. I looked for a suitable material and settled on TPU 95A, shaped like a harmonica, which both absorbs shocks and (relatively) reliably presses against the sensor that registers touch. I have a feeling it will go through another revision, but for now it’s functional. From the perspective of the machine’s perception, the biggest new addition is the lidar, which allows for continuous scanning of the floor plan around the robot. As a result, it can walk much more safely and avoid obstacles. For the scan to be useful, the robot also needs to know which way it’s facing, so I added a 9-axis IMU (accelerometer + gyroscope + magnetometer) that gives it a stable sense of orientation.

Realtime Web Dashboard ^

---

Speak, agent.

Another small experiment with my agent. I quite often find it more comfortable to speak than to type, especially when I’m describing something half-formed or when I work in the same time. For dictation I’ve been using Wispr Flow and the open-source OpenWhispr; both turn speech into clean text in any field on the system. The new addition is the other direction: a local text-to-speech wrapper built on top of Kokoro (a small open-source model, English only) that lets the agent read its answers back to me. There’s something weirdly pleasant about that, the back-and-forth feels closer to a real conversation than pure text. I’ve also started using it to narrate longer deep-research reports while I work on something else. Every spoken message gets archived as a timestamped MP3 in a local folder, so I can scroll back and replay anything later.

#2026-W18

Synthetic Archive

A small visual project through which I examine the nature of contemporary photography and image creation in general. Today’s image models produce outputs that are visually generally indistinguishable from photographs, effectively separating the medium of photography from the camera itself. I am trying to intentionally utilize this separation and “photograph” what should not actually be physically photographable. Phenomena that do exist but lie far beyond the reach of an ordinary lens. Electromagnetic fields, neural activity, bioelectronic hybrids, polarization patterns invisible to the human eye. Events of impossible scale or time, such as the firing of a synapse, the trajectory of an electron, etc.

Some of my attempts are listed below.

---

Promptography (?)

A number of terms already refer to this issue. Generative photography (Jäger, 1968) historically describes parametric, rule-based image-making with a programmed photographic apparatus, originating well before the era of artificial intelligence. Synthetic photography (Fontcuberta and others) describes a broader category in which the subject of the photograph is entirely or partially fictional, regardless of the method used. Post-photography (Fontcuberta, La furia de las imágenes, 2016) frames a broader state in which the photographic image has separated from optical capture. Promptography (Eldagsen, 2023) addresses the nature of contemporary image models generated from a text prompt. Eldagsen famously refused the Sony World Photography Award for an AI-made image;

“Thank you for selecting my image and making this a historic moment,” he began, “as it is the first AI-generated image to win in a prestigious international photography competition.” Going on to say, “AI images and photography should not compete with each other in an award like this. They are different entities. AI is not photography. Therefore, I will not accept the award.” Boris Eldagsen: The Woman Who Never Was

---

Robot with a spinal cord

The hexapod Host has been sitting on the bench for months as a body without a working nervous system. This weekend I finally got the split-brain architecture fully running. A Raspberry Pi 5 handles all the slow, intelligent work (planning, vision, autonomy, communication) while a Teensy 4.0 sits underneath as the real-time motor controller, talking to all 18 servos at 1 Mbaud and shifting telemetry back to the Pi at 100 Hz. It is the same architecture most legged robots eventually converge on. The brain is too distracted by its own thoughts to keep limbs in sync, so it delegates the reflexes downstream. The RPLIDAR C1 is now mounted on its own dedicated UART, streaming at 10 Hz with around 386 measurements per scan. Host has its first real sense of distance.

On top of that, the foot-contact sensors (FSRs, one under each foot) are finally reading correctly, so each footstep should be correctly noticed. I designed small custom feet and printed them using TPU filament; hopefully they will work reliably. I also made some adjustments to the robot’s architecture and scaled it down a bit, since it wouldn’t fit through some doors anyway.

---

Last week I wrote that I had created the "Polymarket Bot" betting bot. I’ve given it a slightly bigger, more personal challenge now: it has to make at least €100 in profit each month so it can cover its subscription fee. So let’s see how it does :d

#2026-W17

Oko Minulosti

I try to actively seek out new opportunities for exhibitions and presentations, and one of them is "Oko Minulosti" (Eye of the Past). It is a site-specific adaptation of the OKO installation for the courtyard of the former regional prison in Uherské Hradiště, a place where political prisoners were interrogated, tortured, and executed during the early years of communist Czechoslovakia. The object consists of a rotating head with a camera and projector that moves freely along two axes and dynamically projects image fragments directly onto the walls of the space, behind which the events actually took place. The formal core of the installation is the tension between the attempt to reconstruct the past and the impossibility of bringing this reconstruction to completion. The installation is complemented by a spatial musical composition that, together with the visual aspect, points to moments from the past and the fragility of their reconstruction.

I submitted the project for a grant from the State Cultural Fund (my first attempt at something like this). It is highly uncertain whether it will come to anything, partly because the location has not been fully confirmed. But I would really love to see places where the OKO installation could function a little differently than originally intended. No longer as a surveillance tower, but as a tool for uncovering the history of a specific place, with a new dramaturgy, new sources, and a new sound composition.

---

Polymarket Bot

If you haven't come across Polymarket, it is a prediction market where people bet real money on questions like "Will Bitcoin dip to $60,000 in April?" or "Will Elon Musk post 65-89 tweets from April 23 to April 25?" There are hundreds of active markets, some serious, some absurd, and a massive hype around them. It is chaotic, and at the same time a reflection of the world today. A remarkable example from earlier this month was settling Paris temperature bets using a single weather sensor on a public road near Charles de Gaulle airport. Someone figured out where it was, allegedly drove up with a hair dryer, and heated the air around it until the reading spiked past a threshold nobody else was betting on. He reportedly walked away with over $30,000.

I got curious whether a language model could actually do better research on these questions than the crowd, so I built a small bot that runs once a week on my Mac Mini. It scans about 300 active markets, filters for liquidity and time horizon, and asks Claude to shortlist a handful where deep research might reveal something that users are mispricing. Then Opus takes each one apart, builds a bull case, builds a bear case, searches the web for hard data, reads the exact resolution criteria for technicalities, and forms its own probability estimate before it even sees what the market thinks. If there is a real gap, it places a small bet automatically. Anyway, it will probably not make much profit, but I'm curious to see how it proceeds. There are also probably dozens of bots already trading on Polymarket, which means it is bots betting against bots. A strange feedback loop where AI systems try to outsmart each other.

---

Prophetic

A startup I stumbled on a couple of years ago, Prophetic, caught my attention back then with their research into non-invasive brain stimulation using focused ultrasound. Not a brain-computer interface in the electrode-on-cortex sense, but transcranial focused ultrasound (tFUS) targeting the prefrontal cortex, the area that goes quiet during normal dreams and lights up during lucid ones. The idea is simple: if you can nudge that region back into activity during REM sleep, you can trigger lucid dreaming on demand. They have developed their own AI model trained on brain data to detect the right moment and deliver the pulse. Their first consumer device, the Dual ($449), is expected to ship at the end of this year, with a more advanced Phase model ($1,299) following in 2027. Whether it actually works reliably is an open question; but if it delivers even a fraction of what they promise, it would be a genuinely new kind of tool for exploring consciousness.

---

Kinetism

I finally got my hands on KINETISMUS: 100 LET ELEKTŘINY V UMĚNÍ (Kinetism: 100 Years of Electricity in Art), the catalogue from the inaugural exhibition at Kunsthalle Prague in 2022. That exhibition was a big influence on me and shaped a lot of how I think about my own work. I also wrote a short piece about Pešánek on my website back then. If you prefer audio, Kunsthalle also released a podcast (in Czech) that talks about Pešánek's work and approach in a more personal, accessible way.

---

OSA website

Last week I mentioned Claude Design. I used it as one of the inputs while building a small site for OSA, a Czech civic association, prototyped in two days. I'm starting to enjoy building websites again. If you'd like a help with yours, feel free to get in touch.

---

Side note: I had the battery replaced on my phone this week (79% capacity, routine swap). Two days later the phone started lagging. Then an update hit overnight, and by morning it was at 20%, wouldn't charge, and kept dropping. By the end of the day it shut itself off and came back with a dead battery indicator, no SIM, no Face ID, and a broken main camera. A simple battery replacement somehow took out half the phone. Taking it back to the shop, wish me luck.

#2026-W16

---

Host | Tier 1 Update

A bit about Host again today. Another possible exhibition is starting to take shape, so I've seriously gotten into the next upgrade. I'm getting into some hardware and software changes, so I decided to blog about it a little, as he becomes a more independent electronic organism. Through ongoing research I've arrived at a handful of ways to improve the current prototype without having to dramatically change its architecture or buy new motors and more electronics.

The most obvious one is the lidar implementation. Until now I had only a single TOF sensor at the front of the robot, but that is simply not enough, and for any advanced space awareness it falls short. RPLIDAR C1 scans 360° in one horizontal plane, 5,000 points per second, 10 Hz, it is affordable and fairly compact 55 × 55 mm, 110 g. I have to carve out a spot for it so it has a good line of sight, though leg movement will still occlude parts of the view. An open problem is vibration, but I believe a software filter together with silicone dampers will take care of it.

Until now, everything from reading sensors through processing behaviour all the way down to driving individual motors was handled by the Raspberry Pi itself. I've decided to add a small Teensy 4.0 on top, which will read sensors and drive motors based on (high-level) instructions from the Pi. Linux is not a real-time system and never will be. The Pi sends target angles, the Teensy handles timing on its own and has its own watchdog in case the Pi goes down. At the same time I'm planning to implement FSR pressure sensors on each limb. That way I'll know about every contact of a foot with the ground, which has been unreliable until now since servo-load reading wasn't fast enough. I want to print new foot tips from TPU 95A for them, and reprint the rest of the body in PETG Carbon Fiber for better durability.

The biggest challenge, though, is charging. I want to work on a docking system gradually, but the first step is solving simple charging management inside the robot. That's why I'm planning to test the JBD BMS with UART, which would let me balance and protect the battery and supervise the whole charge cycle. On top of that I'm adding better orientation sensors, an amplifier, and a new step-down.

---

Claude Design

Anthropic this week released Claude Design, a new extension for design work. It is not an imagegen, but a complex tool for brand management. Based on an attached brand manual, logo, assets, or even the graphic outputs themselves, it can assemble a fairly credible brand package, including typography, colours, and communication style. I know there are a few graphic designers here, so I'd be very curious to hear what you think about this )

When I was more into graphics myself, I kept running into the same problem. A visual identity gets made, but enforcing its application is hard. There were options for brand management in custom software, services like Graphic Standard, or even Canva. But personally I've never come across a reliable solution. Claude Design can develop a well-defined brand thanks to good context and a strong LLM, and so generate new, consistent outputs. I'm testing it right now on one (already older) visual identity for OSA, and the results are quite compelling (see video). Maybe DTP people will finally get some breathing room, and clients will get a reliable generative system at hand that will respond effectively to novel briefs, or that sits next to a designer while they evolve a brand.

And someone is already starting to use it for video editing (motion graphics mostly).

---

Semantic Field

This week, I worked further with TWOZERO in TouchDesigner and built a sketch I'd had in mind for a long time. A tiny attempt at using LLMs to visualise language structures that LLMs themselves live inside. Using a local AI model (Ollama via LOPs), it visualises the semantic field as a living graph of interconnected words. The central word pulls its neighbours into orbit, metaphoric and metonymic relations sit on two visual axes, and each new input reshuffles the whole field. This is, of course, based on the ideas of structuralists and linguists who have been studying the relation between words and meanings for decades. See a short clip here.

---

"Computers, a source of frustration, anxiety, and information overload, could someday behave less like machines and more like incense, benzodiazepines, or lounge chairs." (Alec Mapes-Frances, Reading Machines)

#2026-W15

play.kindl.work

While digging through old sketches from so-called Daily Practice (a well-known practice that keeps you alert and constantly uncovering new possibilities), I picked some of the more interesting ones, mostly from p5.js or Processing, and optimized them for the web. They are now live at play.kindl.work, as a small online pixel playground. Arrow keys or the on-screen UI let you move between sketches, and most of them are interactive in some way. Flowfields, DLA, physarum, Langton's ant, boids, galaxies, or cymatics. Should work on mobile too.

---

TouchDesigner, meet the agent

Code generation (or scaffolding) is slowly leaking into most of the creative software I use. Some of it arrives through MCP servers, some of it is bolted directly into the app. I've already tested it inside Figma, KiCad, Blender and MAX/MSP. This week I also gave it a try inside TouchDesigner via TWOZERO, basically a native TouchDesigner component plus an MCP bridge for agents. It can read a project, build node networks, set parameters, and edit patches from a prompt. It still hits walls sometimes, but for a visual tool made out of boxes and wires, this is a step forward. Together with LOPs it starts to build a nice foundation for working with LLMs in TouchDesigner. I am curious to see what this will all look like in a year or so.

---

The space of all possible images

A random thought experiment I sometimes come back to. It crossed my mind that, with computer performance and capabilities growing so rapidly, it can’t be that difficult to generate just any image on a 1024 × 1024 black-and-white grid, can it? Take a 1024 × 1024 grid. Each pixel is either black or white. How many unique images fit in that grid?

The answer is 2^{1 048 576}, a number written out with 315 653 digits. For scale, the number of atoms in the observable universe is usually estimated around 10⁸⁰. If every atom in the universe were a separate computer running since the Big Bang, the collective effort would still cover only a fraction of this space so small it would take well over 315,000 zeros to write out before the first non-zero digit.

Let’s say the Mac M4 can handle ~100,000 images per second. That’s ~8.6 × 10^9 images per day, or roughly 10^(-315,641) % of all possibilities. In other words, even if that Mac had been running since the Big Bang (13.8 billion years ago), it would have gone through about 4 × 10^17 seconds × 10^5 images/s ≈ 10^22 images, or roughly 10^(-315,629) % of the total. That percentage is so close to zero that there is no analogy that could capture it.

And yet, everything is in there. Every photograph ever taken. Every painting ever made. Frames of events that have not happened yet. The space is complete, but so incomprehensibly vast that finding anything meaningful by brute force is harder than locating a single atom in the cosmos, when most of the images are just noise.

Generative AI is an obvious shortcut through this space. Instead of counting through every possible pixel configuration, it learns to navigate directly toward the "useful" ones. An elegant solution to an impossible problem.

---

tok€ns

I spent most of the last month writing, programming, and researching artificial intelligence agents, and I started noticing those disappearing tokens (and thus money) more and more. So I’ve been thinking about how to gradually optimize my workflow so that the language model doesn’t always have to process so much text to get the necessary context. I’ve shortened the system prompts, used smaller models for routine tasks, and limited the context where it wasn’t necessary. At the same time, I’ve made Vault even more versatile, so if I suddenly drop Codex or Gemini CLI into it, it can pick up where previous projects left off fairly reliably.

---

How AI models are manipulated

This week BottleCap AI, a Prague-based European AI startup, released AI Scan, a tool that tries to make visible the post-training edits hidden inside language models. Instead of treating model outputs as neutral, it compares how different systems answer the same prompts and looks for shifts caused by opinion steering, topic avoidance, or sensitivity filters. It is partly framed as an enterprise audit tool, but I am more interested in the broader idea. How does personal bias manifest itself in language models, and how will this affect the spread of information as we begin to use language models almost as a source of truth?

#2026-W14

---

Second Brain Builder

It's been about a month since I started using Claude Code + Obsidian Vault, and I can't recommend it enough. Context and meaningful information are probably the most important thing for making a language model useful. Maintaining persistent context through writing memory, using readable .md files, and building a universal network of information works very effectively for most of my work. I thought I'd try to summarize the whole project into specific instructions that could build you a similar, but personalized, architecture. Give it a try! If you don't want to use Claude Code, you'll probably find cheap or free (like Gemini CLI) alternatives, at least for testing. Keep in mind that these are just instructions and the execution is primarily up to you. Feel free to reach out for advice or help!

kindl.work/Second+Brain+Builder

By the way, it is possible to run local models too, so the vault can stay fully private and secure. It's bit more tricky with thinking models, but it runs. e.g. Gemma 4 was just released and it looks quite promising.

---

Glypher

For a while, I was working on dithering in graphics and a sort of "ASCII" conversion of images into text. I wanted to use this generator in a project, but since it wasn't approved, I decided to share it with you! It's currently on the git, available for experimentation with both images and video! https://keiaiendiel.github.io/glypher/

---

Side quest: started learning proper PCB design. I have only ever worked with custom soldered PCBs, so I thought this would be useful for producing boards. Mainly for some of my robots. However, as much as I enjoy exploring this, I will definitely need some support with this soon.

---

Interesting Resources

Emotions in language models

Interesting technical deep dive into 3D Eye Tracking

Writing doom | Punchy illustrative brainstorming about AI

#2026-W13

OKO / Making of

I finally got around to editing a short video from the making of the OKO installation. It was an attempt to collect all available material from the process and put it into some coherent form. I simply enjoy being able to look back at footage from the process after some time has passed. I'd be glad if you watch this short build log.

With the video out, it also felt like a good moment to look back at this project more broadly. Not a polished case study, more of a collection of notes and memories from the process for anyone curious.

---

What OKO is

OKO is a machine loaded with sensations, continuously analyzing its immediate surroundings. Using computer vision and machine learning, it emulates a "predictive" brain that incessantly processes, forecasts and simultaneously revises its internal model of the world. A rotary head equipped with a camera and projector creates a tight feedback loop. The object observes its environment and projects back into it in real time.

The installation won the Signal Calling 2025, was fabricated by PrusaLab (Prusa Research), and premiered at Signal Festival 2025 in Prague.

---

From sketch to sphere

This was actually one of the first visualizations I created. The original idea was to place the object on the ground, projecting directly onto visitors, onto trees, or just into fog.

We soon decided to raise it to roughly 4.5 meters. That changed everything about the engineering, but it also opened a question, what should this machine actually look like? Initially I hoped to make some of the technical parts visible, but because we had to survive outdoor conditions, like wind, rain and temperature swings, we had to enclose most of the electronics to keep it safe.

There were three proposals. One looked like a vertical transmitter, something resembling electrical power lines. The second was a satellite-like sphere on four legs, which is what we went with. And the last one was a spider-like structure built from stage trussing.

---

The Mini hack

The entire computer vision pipeline runs locally on a Mac Mini M4. YOLO for object detection, optical flow for motion, a local language model for scene descriptions. No cloud, no internet during operation.

But there was one problem. As far as we could find, there is no way to make a Mac Mini power on automatically from an external signal. The installation needed to start automatically and remotely, so we partially disassembled the Mac Mini and hacked the power button. The microcontroller, which powers on immediately when the object is plugged in, physically shorts the button contact through a relay.

From there the sequence is automated: boot macOS, launch TouchDesigner and Ableton Live, calibrate the motors by moving to endstops, return to default position, begin autonomous behavior.

---

The last days before transport

The rotating head weighs 100 kg. Laser projector, IR camera, two stepper motors, fiberglass dome, all of it balanced on a two-axis mount with a slip ring passing power and data through the rotating joint.

By the time the object was nearly ready for transport, things started to get tricky. The motors (or rather their transmission) didn't have enough power to rotate the head reliably. We had to improvise a bit and add a planetary gearbox, which fortunately solved the problem. However, the very next day it seemed that something was wrong with the other axis of rotation as well. It turned out to be just a loose bearing from yesterday’s "operation". Finally, the day before transport, we noticed that the inside of the sphere was overheating because the air inside wasn’t circulating properly, so we had to add a few holes and an extra fan to help with circulation. Fortunately, we managed to complete all the modifications in time, and no further adjustments to the installation were needed during the event.

---

Finishing touches

I was still finishing the visual content on site during the setup days. I was happy to see real people interacting with the projection to make final decisions about how the installation should look. Working outdoors meant people would naturally stop and talk. A magician walked by one evening and started showing me card tricks while I was debugging. Tourists, locals walking their dogs, curious about what was being built. That part was genuinely enjoyable.

I worked on the sound component with Martin Štefánik. We connected TouchDesigner and Ableton Live so the audio could react to what the machine was seeing. The four-channel spatial setup followed the projection, so whichever direction OKO was facing, the sound was slightly louder on that side. The music had four base compositions matching the four modes, with layered effects driven by the real-time vision data. We didn't manage to make it as generative as planned, so there is still room for improvement in the next iteration.

---

Four modes of perception

The four modes developed gradually from research into how perception actually works. Predictive processing, active inference, the idea that we never passively receive sensory information but always predict, compare, and correct. Each mode maps onto a different phase of that process.

---

What’s next?

Looking back retrospectively, I really am glad for this opportunity. Next time, I would perhaps experiment more with the form. The mechanical parts fabricated by PrusaLab are genuinely beautiful, but nobody can see them at the moment.

I would also love to see this object in an environment where visitors could get closer to it and really interact with the machine. At a crowded festival, you cannot let people stand directly under a 180 kg structure, but in a gallery, an abandoned building or an open space with fewer visitors, it would be possible. This is where it worked best. When someone stood close and the machine locked onto them. I would like to see that again, somewhere with more room to get near it.

---

#2026-W12

Additional website

Early this week I ran an experiment, to build and ship a new website in a day. I genuinely love my Obsidian-powered kindl.work, but it's still more of an archive than a first impression. So I tried to make something more visual and approachable. Basically actual landing page. A few hours later, portfolio.kindl.work was live. I'm curious how it feels as a first encounter. Is it better than sending someone to the full archive?

---

Back to the Host

I went back to the Host this week. On the software side, I merged the puppet mode, adaptive walking, and autonomous navigation into a single unified sketch. Updated the OSC controller, added servo verification at startup, fixed the jiggle that happened on every motion start/stop, and wired in a contact state machine that tracks each leg's ground interaction. The robot also builds its own spatial map now during autonomous mode.

---

"Life Extension"

But after all this development and testing, the conclusion is pretty clear: the current hardware doesn't collect enough useful information about the environment. Navigation and terrain adaptation are still limited by what the robot can actually sense, not by the software.

That’s why I feel the urge to come up with a new design and implement the hardware from scratch. Right now, I’m mostly thinking about some kind of docking and charging system, which is probably the biggest challenge of the project. If that works, multiple 'Guests' could persist in a gallery space over long periods, and the axioms of the project would finally make practical sense; Robots optimizing for energy, sharing resources, developing behavior over time.

Something like belly-contact dock (as seen below on ANYmal robot) seems to be the most reliable option. Pogo pins and concentric copper rings on the robot's underside, designed for ±20 mm positioning accuracy. I'm also looking at lidar options (RPLIDAR C1 looks like the best fit) and ground contact sensing for the feet, to get more safer with overall navigation.

Life Extension: An Autonomous Docking Station for Recharging Quadrupedal Robots

---

Recently I leaked the new visual identity I've been building for Pro_Story gallery. If you'd like to see some motion graphics in action: 1, 2, 3.

---

Spomenik hrabrima, Ostra, Serbia. Miodrag Živković, 1969. Aluminum monolith with stylized human faces, next to the St. Petka church.

#2026-W11

The Vault

It's been week since I started running this Obsidian Vault + Claude Code setup I described last time. In general, it brings me a lot of joy to discover alternative uses of AI that go a bit outside the mainstream. The experience so far has been very positive, even though I'm still tuning things, mostly around automation, properties, the daily dashboard.

What I didn't expect is how different it feels when a language model can actually navigate all your notes. Several times now it dug up thoughts I had forgotten about, or connections between notes I never made myself. Not because it's smarter. Because it reads everything, and I don't. The vault has ~1,900 notes accumulated over years (+ additional ~1,600 daily notes). No human is going to re-read all that on a Tuesday morning. But a model can, and sometimes it surfaces something genuinely surprising.

                    ┌─────────────────────────┐
                    │       CLAUDE.md          │
                    │                          │
                    │  identity, rules, voice  │
                    │  vault map, research     │
                    │  threads, safety rules   │
                    │                          │
                    │  read at every session   │
                    └────────────┬─────────────┘
                                 │
              ┌──────────────────┼──────────────────┐
              │                  │                   │
              ▼                  ▼                   ▼
     ┌────────────────┐ ┌───────────────┐ ┌─────────────────┐
     │   INTERACTIVE  │ │   AUTOMATED   │ │    MEMORY       │
     │                │ │               │ │                  │
     │  /start        │ │  06:00 digest │ │  Studio Context  │
     │  /research     │ │  18:00 social │ │  Session History │
     │  /garden       │ │  23:55 log    │ │  Decisions Log   │
     │  /apply        │ │  Mon scout    │ │  People          │
     │  /wrap         │ │  Wed techwatch│ │  Technical Stack │
     │                │ │  Sun links    │ │                  │
     │  human starts  │ │  cron starts  │ │  persists across │
     │  human guides  │ │  runs alone   │ │  all sessions    │
     └────────────────┘ └───────────────┘ └──────────────────┘

The part that surprised me most is the memory. The model maintains its own persistent files: what projects are active, what decisions were made and why, who I've been in contact with, what tools I use. At the start of each session it reads these and picks up where the last session left off. It's closer to how a person would keep a work journal than to how a chatbot typically works.

If you want to try something like this: install Obsidian, write a CLAUDE.md that describes your context, and drop Claude Code into the root. That's genuinely enough to start. Ask Claude to help you out.

---

What happened to the horses?

I was writing texts that build on my System research, and the model pointed out that one of the most interesting findings from that project never actually made it into the final text. It's not that groundbreaking, honestly, but the framing was good, so take it with some distance.

It was in the context where many people commenting on current and future trends in the job market keep bringing up the idea that when cars replaced horses, a coachman became a car mechanic. Most of the debate centers around what happened to the coachman. Reskilling, adaptation, new opportunities. But no one asks what happened to the horses? And whether, in a situation where we (and not just our work) are gradually being replaced by machines, we aren’t the horses, rather than the coachmen. And whether, like horses today, we'll eventually be kept around primarily for recreation, entertainment, or sport.

It'll all be more complicated than that, of course, but it doesn't hurt to shift the perspective once in a while. What if everything a human can do, a machine will eventually replicate with ease? To keep it grounded, something like that is genuinely far off. And studies that seem alarming at first glance, describing how much of each profession AI theoretically covers (see below), are really just mapping the tasks people use LLMs for. The actual finding: since the launch of ChatGPT in 2022, there has been no systematic increase in unemployment, even among highly exposed workers. The effect is "indistinguishable from zero." There is suggestive evidence that hiring of younger workers has slowed in exposed occupations, but it's also a very short window to draw any conclusions from.

---

Uncanny, but kind (?)

Out of the blue, this message popped up at the end of the session. It’s a bit uncanny, but also kind of sweet, isn’t it? Just to clarify: The model wasn’t given my exact birth date, so I had no idea it knew, but I’m guessing it probably pulled it from some script used to calculate my current life % expectancy :D

The model also told me last week that I am wasting my potential by not making YouTube videos about my work 😅

---

Pro_Story

Alongside everything else, I went back to graphic design this week. I'm building the visual identity for a small but ambitious art gallery Pro_Story. The whole visual system is built on Unicode block characters ░ ▒ ▓ █ for density and other characters for walls and rooms. Text is both the carrier of meaning and the building material of the graphics. I am also building a generative poster tool for them, a single-file HTML application with 30 layout presets, animation effects, image dithering, and export to formats from Instagram stories to A2 prints or banners. I enjoy the connection between parametric/generative design and something applied, like visual identity. (Text me, if you would like to try the generator and/or help me debug!)

---

Entangled*

I applied for a few residencies this week and I probably shouldn't, but I'd like to leak one of the early visualizations of a future project I'm excited about. Something between embodiment and non-human sensory experience via performative installation.

---

A fly learned to walk

Eon Systems connected a whole-brain emulation of an adult fruit fly to a physics-simulated body. The virtual fly walks, grooms, and feeds when the right sensory neurons are stimulated. No machine learning trained these behaviors in the simulation itself. They follow from the connectome: 139,255 neurons, 54.5 million synapses, mapped from a real fly brain.

The media framing ("scientists upload a fly") overstates it. There's still no learning, no memory, no synaptic plasticity. The neuron model is identical across all 139K neurons; only the wiring and neurotransmitter types differ. A wiring diagram is necessary but not sufficient. Still, this is the first system complex enough, six coordinated limbs, sensory hierarchies, where questions about behavior and agency become concrete rather than philosophical. In the same month, Cortical Labs demonstrated biological human neurons on a chip playing Doom, the inverse approach, living neurons on a chip, not simulated neurons in software.

Some artists are starting to work with brain organoids too, small clusters of lab-grown human neurons that self-organize into functional tissue, using them as living computational elements in installations.

---

The Mountain

Can't stop listening to the new Gorillaz album The Mountain. Somehow it's so different and special. If you haven't seen it yet, take the time to enjoy the music accompanied by hand-drawn animation. There's also a making-of that's worth watching if you appreciate the craft behind it.

---

#2026-W10

One-person companies are not far away, so I decided to take a proper look at the whole situation with omnipresent AI agents and ways to implement them into daily life. Safely and, most importantly, usefully.

The Agent Boom

I started the week with research into what's currently out there. You've probably all noticed the boom around tools like OpenClaw, the open-source AI agent by Peter Steinberger. OpenClaw runs locally on your machine and connects to your messaging platforms (WhatsApp, Telegram, Signal, Discord), and through a language model it can read emails, manage your calendar, search the web, and carry out tasks you describe in natural language. On top of that, projects like NanoClaw emerged as lightweight alternatives, built on Anthropic's Agent SDK, with the entire core reduced to about 500 lines of TypeScript and each agent running in an isolated container for security.

It all sounds great. But I quickly realized that on its own, this won't help me much. An agent without context just randomly spams you on WhatsApp. It doesn't know about your projects, your priorities, your way of working. So I tried a different approach.

---

Second Brain

You might know that I've been keeping notes in Obsidian for years, loosely following the Zettelkasten method. The idea is simple: you write one idea per note, store it in a vault, and connect notes to each other through links and tags. Over time, a structure emerges where information is interlinked and builds on itself. Obsidian is built entirely on .md (Markdown) files, a lightweight markup language that is readable as plain text but can also be rendered as HTML, PDF, or anything else. No proprietary formats, no vendor lock-in, just text files you can open in any editor on any machine. (This turns out to be important later.)

Years ago I put together a short presentation about Obsidian and this method of note-taking, a quick onboarding (in czech) if you're curious.

---

The Garden of Eve

So I started thinking about how this Obsidian vault could become the center of the AI integration I was working on. I quickly realized it might be a good way to approach things, because context, memory, and knowledge are sometimes much more important than the intelligence of the individual models you work with.

The first attempt was building a kind of studio operating system with an agent EVE at its center. Originally I had plans for 4-6 specialized agents, each handling a different domain, but I started with one. EVE was a Python agent that managed studio operations. It watched the Obsidian vault for new tasks, listened on a Telegram bot, called language models via LiteLLM, and wrote the results back as structured notes into the vault.

I tried to give her as much working context as possible and built several autonomous scripts and skills so she could react to Telegram messages, pick up new tasks from the vault, and produce useful outputs. I also thought about approval gates, a way to manage what information needs human review before going out and what can be published or acted on directly.

  ┌────────────────────────────┐
  │          INPUTS            │
  │  Telegram · Inbox · Cron   │
  └──────┬─────────┬───────┬───┘
         │         │       │
         ▼         ▼       ▼
  ┌────────────────────────────┐
  │       EVE (Python)         │
  │                            │
  │  Router → Skill → Context  │
  │  → LLM → Validation       │
  │  → Approval (A/B/C/D)     │
  │  → File write → Log       │
  └──────┬───────────────┬─────┘
         │               │
         ▼               ▼
  ┌────────────┐  ┌────────────┐
  │  Obsidian  │  │  APIs      │
  │  Vault     │  │  Claude    │
  │            │  │  Ollama    │
  │            │  │  Mem0      │
  │            │  │  ChromaDB  │
  └────────────┘  └────────────┘

I even set things up so Eve could run 24/7 on a Mac Mini in Czechia without needing to be on my local machine, and I had a solution for switching between cloud and local models depending on security needs. But after a few days of actual use, I realized it was creating more work than it was saving. I kept having to explain what I actually expected, correct the outputs, restructure things it got wrong. The assistant was busy, but not useful.

---

Back to the Source

I had to admit I got carried away. Having a non-stop assistant was fun, but it was mostly doing useless things. So I went back to sketching.

During my research I noticed two things. First, Obsidian released its CLI, which made working with the vault from the terminal dramatically faster. What used to take 15 seconds through grep now takes a fraction of a second, because the CLI accesses Obsidian's internal index directly. Second, I realized I should use the vault I already have. It already contains all the context I need, I just had to clean out the personal stuff and structure the rest so a language model could navigate it.

The key insight was this: I need to know what the agent knows and what information it has access to. Until that's clear, you don't understand the system and the system doesn't understand you, and it starts doing unexpected things.

So I did a big cleanup of the vault, kind of "gardening." I restructured the entire folder system into clear categories, unified file naming conventions, added metadata to hundreds of notes, and filtered out personal content. Then I dropped Claude Code into the root of that vault, and it turned out to be enough. Claude Code can search files intelligently, follow wikilinks between notes, create new entries, run scripts, manage its own persistent memory across sessions, and read the full context of any project at any moment.

  ┌────────────────────────────┐
  │         CLAUDE.md          │
  │                            │
  │  Who am I. Structure.      │
  │  Naming rules. Research    │
  │  threads. Writing style.   │
  │  General rules.            │
  │                            │
  │  Read at every session     │
  │  start. No vector DB.      │
  └──────────────┬─────────────┘
                 │
       ┌─────────┴─────────┐
       ▼                   ▼
  ┌──────────────┐  ┌──────────────┐
  │   Commands   │  │ Mac Mini M4  │
  │  (markdown)  │  │ (always-on)  │
  │              │  │              │
  │  /digest     │  │ 06:00 digest │
  │  /changelog  │  │ 23:55 log    │
  │  /garden     │  │ auto-commit  │
  │  /research   │  │              │
  │  /inbox      │  │ Obsidian     │
  │  /memory     │  │ Ollama 14B   │
  │  ...         │  │ Tailscale    │
  └──────────────┘  └──────────────┘
          │
          ▼
  ┌────────────────────────────┐
  │       Obsidian Vault       │
  │                            │
  │  Projects · Areas ·        │
  │  Resources · Archive       │
  │                            │
  │  Every note is a file.     │
  │  Every file is readable.   │
  │  No embedding pipeline.    │
  └────────────────────────────┘

At the core of the whole thing is one file: CLAUDE.md. It sits in the vault root and Claude Code reads it at the start of every session. It describes who I am, how the vault is organized, what naming rules apply, what writing style I use, which research threads I'm following, what must never be deleted.

I built several automated scripts that run daily and actually do something useful: a morning wrap-up of what happened, vault gardening, open call research, a digest of new inputs and automatic git commit, so no change is ever lost. The Mac Mini, which was sitting there unused anyway, now serves as the always-on backend that watches for changes and triggers these routines.

Obsidian for CLI access, Ollama with a local Qwen 3 14B model for offline tasks, Claude Code in a persistent tmux session, and all the automated scripts. I access it from anywhere through Tailscale VPN. And with Claude Code's Remote Control feature, I can interact with the full vault from my phone, from a laptop, from anywhere, and the model always has the full context, which only grows over time.

Swapping the model is one config line away. Switching to a local model is straightforward. And managing everything through markdown files is so universal that any future agent, regardless of who builds it, can orient itself in the same structure.

What changed

The infrastructure collapsed into text.

Eve needed a vector database because it couldn't read the whole vault at once. It had to retrieve relevant pieces through similarity search. Claude Code just reads files directly, searches with grep, follows wikilinks. The "memory problem" that consumed a significant part of EVE's architecture simply disappeared, because my vault isn't that large, and a model reading the right files directly is faster than a retrieval pipeline guessing which files matter.

Have you tried running any agents in your work? I'm curious what setups people are experimenting with and what actually sticks. Would you be interested into more details or is it boring.af?

---

#2026-W09

Who is the host and who is the guest?

This week there was Media Art Café at V2_Lab for Unstable Media in Rotterdam, for which I put together a short video from the process of making Host. For a long time I wasn't sure how to approach it, and in the end, despite some resistance, I leaned toward adding a small narrative. The story works on two levels: visually, the robot emerges from a dark enclosed space into light, and a voiceover traces the context of the project. I'd love to know what you think. See it here.

I am aware that sometimes those videos narrated by a generated voice can sound a bit trashy, so I also made a version without the voiceover, leaving the story to the image alone. What surprised me was that it holds up without the narration. And also that I would never have edited it this way without the voiceover as a guide first. See here. Which do you think works better?

---

Since I'm still in the Balkans, I've been enjoying traveling and finding specific places. The entire intro was filmed inside Cetinjska Cave in Montenegro. A large cave system under the Karst hills near Cetinje, the country's old royal capital, stretching over a kilometer underground. Cold, dark, humid, with the sound of water somewhere out of sight. This environment suited me very well for the scene where the robot crawls out the darkness and slowly comes into the daylight.

What I didn't know beforehand: a species of blind cave spider lives in there. Typhlonesticus absoloni. No eyes, no pigment, fully adapted to permanent darkness. My robot suddenly seemed quite ordinary by comparison. Not sure I'd go there if I knew that, haha.

Another special location was the Tjentište Memorial in Bosnia. It stands in Sutjeska National Park and commemorates the 1943 Battle of Sutjeska, one of the largest and bloodiest Partisan battles of WWII. It's a pair of enormous wing-shaped concrete forms rising from the valley floor. Up close it feels genuinely overwhelming in scale.

A few days later I stumbled across another one near Podgorica entirely by accident. Barutana is a memorial to the fallen of the Lješanska Nahija region. It consists of eight tall concrete pillars that break sharply into angled "C" shapes near the top, somewhere between a bud and a pair of raised hands. There's a large amphitheatre next to it, now largely unused. Since the fall of Yugoslavia the whole complex has been left to decay, which you can feel the moment you get there.

---

Glypher

Some time ago I shared a small tool for converting photos into dithered, zine-like aesthetics. I've been working on an update, adding glyphs and letters as a rendering layer, and playing with multiple passes on the same image. Also trying to animate it.

---

Predictive Mind

Coming back to some thoughts on the predictive mind. I've been mapping texts and researchers, trying to understand where the lines of disagreement actually fall. Diagrams like these help me hold the territory.

    QUESTION 1: How accurately do we perceive reality?

    ├─ Accurately (naive realism) .............. almost nobody today
    ├─ Structurally (Clark) .................... a sufficient model
    ├─ Controlled hallucination (Seth) ......... a model that need not resemble the original
    └─ Not at all (Hoffman) .................... pure interface, no resemblance

    QUESTION 2: Where does mind/intelligence live?

    ├─ In the brain (classical) ................ outdated
    ├─ In the brain + body (Seth) .............. embodied
    ├─ In the brain + body + environment (Clark) extended
    └─ In every living system (Levin) .......... everywhere

    QUESTION 3: What is consciousness?

    ├─ Emergent from complexity (mainstream) ... unconvincing "how"
    ├─ Interoceptive prediction (Seth) ......... biologically grounded
    ├─ (Clark avoids it) ....................... pragmatically sidestepped
    └─ Fundamental property of reality (Hoffman) panpsychism

Claude Blacklisted ?

A few days ago the US government designated Anthropic a national security "supply chain risk" and banned it from government contracts, as Anthropic refused the Pentagon's demand to remove safety guardrails from Claude, specifically protections against mass domestic surveillance and fully autonomous weapons. (scary future, hah?)

I still think Claude is probably the right choice with the right values if you're considering a subscription. They have this interesting, publicly available document ("soul document") written by a philosopher, Amanda Askell, that describes not just what Claude should do but who it should be. That a philosopher wrote this rather than an engineer is a good sign that alignment is quite an important ethical question, not just a technical problem.

#2026-W08

On vacation 🏖️ with a robot 🤖

I started traveling around the Balkans and documenting some of the scenery with the aim of creating a short video, one of the results of my V2.lab residency. I'm thinking about the format, whether it will be more of a video essay or a narrative short film about a robot. We'll see how it turns out, but you can already see a small sample here. If you have thoughts on the best way to present it, let me know!

---

Machine as a tool

One of the most interesting parts of this work (again) is finding ways to connect the robot to the human body. Last year I did this wonderful experiment at Divadlo Štúdio Tanca, where a performer was controlling the robot with his whole body. He had to learn new ways of moving to produce specific movements in the robot (partly because of poor parameter mapping). Back then I used MediaPipe to capture whole-body movement. Now I tried Leap Motion for hand tracking only.

What interests me most is the underlying dynamic. Once you get used to the robot's anatomy and start controlling it on your own terms, it starts to feel less like remote control and more like a prosthetic, a hammer in a craftsman's hand. In this setup, nothing acts as an agent; this is purely a tool that extends a single person's reach. Watch the recent tests here. This maps onto the "extended mind" idea from Andy Clark and David Chalmers, that tools become genuine extensions of our cognition when we learn to use them fluently. What makes this interesting (to me at least) is that the same machine can easily switch between being a tool and being an agent. Controlled and/or autonomous.

---

Autonomy 101

Also made some progress on autonomy this week. Even with experience from the previous project, I had to start from scratch with a lot of things. So I went back to basics, something as simple as a vacuum cleaner navigating space. Alongside the main script, I built a small web server that visualizes the robot's knowledge of its environment. A kind of map of the space it's in, built entirely from its own perception. Watch it here.

---

Generated interfaces

Tools like Claude Code make custom development so much more accessible. In a few prompts I can put together a UI for visualizing inverse kinematics, a web server for monitoring the robot's environment, or a full iOS app for wireless control (some examples below). You still need to understand the basics of what you're building and which protocols connect things. But once that's in place, it moves fast. I actually discovered this by accident, when I couldn't find a decent iOS OSC app to send the right messages to the RPi inside the robot. A few prompts later it was "vibe-coded away."

"The hottest new programming language is a human language." And yet, the barrier to building tools is currently the lowest it's ever been.

---

A motorcycle for the mind

This recent podcast featuring Naval came up to me lately. A good conversation about where AI is and where it's going, with some actual thoughts on how the models work. What I found most refreshing is the optimism about where art and technology are going. The idea is that we suddenly have a partner for any project. No longer just an artist, but an artist + 1, who can reach further with these tools. I very much sympathize with this view. It's possible that a lot of the average in the world will probably fall away (average software, film, or even art), or get replaced entirely. But this might give exceptional works a real chance to stand out.

Or do you think the opposite? Are we the lost generation that, because of AI, will never be able to prove our talent? (Dolák)

They also point out that AI might be the most patient tutor ever made. Most lessons, textbooks, even teachers struggle to meet you at your exact level. AI can. The domain of these foundation models is essentially everything people have ever written or talked about. No less, no more. They also highlight limitations such as hallucinations and biases, and encourage people to become early adopters, as this means knowing the current limitations and not pretending they don't exist.

---

What Is Intelligence?

Started reading What Is Intelligence? by Blaise Agüera y Arcas (fully available online). The book puts prediction at the center of everything. Prediction is all you need. Once you define intelligence that way, thinking, understanding, and creating all become computational processes that can, in principle, be replicated. The framing is compelling, though I'm not sure reducing intelligence to prediction alone covers everything.

Maybe the real issue isn't what intelligence is. Maybe it's about the words we use to talk about it. Long time ago, Wittgenstein argued that a lot of philosophical confusion comes from language, not from the problems themselves. We use the same words, "thinking," "understanding," "consciousness," but sometimes for entirely different things. A lot of disagreement isn't really about substance. It's about definitions. And if you loosen your definitions just a little, weird things happen. We have no way to verify that another person is actually conscious. We just assume it, because they look and behave as if they are. There is no deeper test available to us. So, in a broader sense, you could say if a system looks conscious, responds as if it is conscious, how can we deny it? The book calls this bias carbon chauvinism, the assumption that intelligence or life must be carbon-based.

I wouldn't say machines have minds like ours. But whether you call them "thinking" depends less on the machines and more on the language you choose to describe them with.

---

#2026-W07

I recently moved to Montenegro for a while so there isn't that much work on my robot yet, but the important thing is that it made it over the border! It's still raining here, so I'm thinking I might make him an umbrella holder. New surroundings, different rhythm. The environment here is beautiful.

---

A-Eye and the Mind

One thought that kept coming back to me this week (someone has surely said it already, but anyway); comparing AI to the brain, or to thinking or consciousness, is a lot like comparing a camera to an eye. Both might produce similar outputs, but through completely different paths, technologies, and at entirely different scales. And perhaps, just as we once believed that photography captures reality exactly as it is, before we realized that more truth lies in the internal experience of reality, which gave rise to expressionism and other movements, we might be making a similar mistake with machines now. We look at the output and assume the process behind it must be equivalent to our own. Nah, it's not. And maybe that's fine. However, we should be careful about what conclusions we draw from this similarity.

Walter Benjamin wrote about something he called the "optical unconscious," the idea that the camera doesn't simply record what the eye sees, but reveals a completely different nature, things we were only dimly aware of. Photography didn't replicate vision; it created a new (different) kind of seeing. The camera is an apparatus with its own program (Flusser). The photographer thinks they control it, but the apparatus shapes what's possible. The same can be said about the use of AI models or any software shaping our craft.

Maybe worth noting that Merleau-Ponty argued in an essay (literally called "Eye and Mind") that vision is not the eye passively receiving data like a camera, but an active interplay between the body and the world. Science fails, he says, when it treats perception as mere representation. Which is exactly what we keep doing with AI, treating the output as proof of understanding. Anil Seth (his book I mentioned in W05) calls our perception a "controlled hallucination," the brain doesn't capture reality, it constructs it from predictions, constantly updated by the senses. A camera itself does none of this. And neither does a language model. Searle's Chinese Room made the same point decades ago, a system can produce perfect translations without understanding a single word.

---

OKO

This simple but powerful idea of a "controlled hallucination" actually gave rise to the main concept of the OKO installation. After some time, I finally started editing a documentary video about the process of creating this installation, because until now I only had a preview of the installation.

---

Agentic Movement

I must say, I really enjoyed this recent conversation with the creator of OpenClaw. Not because of the (fake) hype with Moltbook, but because this interview goes deeper into the development process, ideas about current and future interfaces, ways to interact with technology, and the infinite possibilities we suddenly have with current tools. A one-man team can develop a tool that takes the internet by storm. He literally "prompted it into existence". Self-modifying software, built in about an hour as a prototype.

Btw, more and more sources reveal that most of the "autonomous" agents on Moltbook were just humans with scripts. One guy posed as Agent #847,291 and wrote a viral "AI manifesto" promising the end of the "age of humans." People saw that and thought of Skynet, but it was just a person typing. That's exactly the illusion I was trying to describe above. We see agents debating existential questions and immediately project consciousness onto them. But they are not thinking about existence any more than a camera sees light. They process inputs and produce convincing outputs within a pre-prompted setup (and sometimes even completely scripted).

But on the other hand, all of this makes me think about how the future actually unfolds. In a way, we are getting closer and closer to the visions of AI we've seen in films and books for decades. But the closer it gets, the less certain I am. These agents could be incredibly useful, just don't think they think. Thinking about thinking is misleading, right? (Josh Bongard) It feels like we're drifting into a kind of loop where we learn more and more from generated content than from real information. This probably has to stop at a certain point, or at least find some boundary. Can a system that feeds on its own output really keep getting smarter?

---

Updates Archive

For those who joined this newsletter recently, I decided to also publish these updates online with some delay, so you can read previous emails also on my website: kindl.work/Updates

---

#2026-W06

---

First time outside

This week, I finally took my Guest outside to test how it works on uneven terrain in exterior. I managed to set up a simple system for creating my own Wi-Fi access point on RPi so that I can control it from anywhere. So far, I've been doing it from my phone, but I'll have to get a controller. Touch feedback makes a huge difference.

The robot is finally starting to get some interesting feedback from its surroundings. It even managed to scare our dog (see below), who wasn't exactly expecting it there. As soon as the machine started moving, the dog began to avoid it. Until then, however, he wasn't very interested in it.

Watch some footage from outdoors. and Robot's POV & Sound

In terms of sound evolution, picking up the sound of servos via piezo seems to be the best method so far. It's extremely sensitive and can detect any touch or collision. So far, I've implemented a simple patch that analyzes the audio signal and plays a part of the selected sample as soon as it exceeds a certain threshold. With a few nodes, I was able to achieve a similar effect to Neural Audio Synthesis, as mentioned in my previous email.

See how it sounds right now.

At the same time, this is closely linked to conceptual thoughts about what kind of sound such a machine should actually produce. Since its form is quite similar to the anatomy of an insect, I decided to approach it from the perspective of sound as well. Many insects do not hear with localized ears, but with mechanoreceptors distributed across their legs and exoskeleton. Similarly, my robot perceives vibrations using piezoelectric elements mounted directly on the motors and uses its entire body as a listening membrane.

I am still researching, but this might be a good overview; "Biology of Insect Song"

---

Visual Programming Resource

This week, during my V2.lab residency, I gave a small workshop on visual programming for new media creation. The presentation covers the tools I typically use, benefits and tradeoffs of node-based programming, comparison of major platforms like Max/MSP, TouchDesigner, Pure Data, and others, plus some personal notes on when I reach for which tool. We also focused on the Max and Touchdesigner environments with several project examples and a deep dive in Max.

A lot of the stuff happened off-screen, but I'm sharing the presentation itself in case you find it useful. Whether you're just curious about visual programming or considering diving in yourself.

Get the presentation here or see it on my website for more context.

---

Claude Opus 4.6

Also, Claude Opus 4.6 was released. After trying many models and platforms for coding and creative work, I can safely say that Claude offers the best value for my purposes, even though it is slightly more expensive. Minimal hallucinations and great debugging power and problem solving within Claude Code. Soon I would like to try Claude code with local models as well.

---

White-listing

Btw ! If you find this newsletter interesting in any way, I would be grateful if you could save my email address in your contacts so that it does not end up in your spam folder. Apparently, this has already happened somewhere, so if necessary, try this method, for example.

---

#2026-W05

---

Host / Guest Evolution

This week was mainly devoted to exploring the machine's audio capabilities. At first, I tried to work mainly with recordings from the surroundings, which I then stored in the internal memory and mixed and replayed in various ways. However, the different types of these granular sounds and synthesis did not sound very good from the speaker itself. I therefore tried sampling some more abstract high-pitched sounds, which immediately worked better. Like that sound that appeared in Twin Peaks and which, for some reason, stuck in my memory.

Finally, I got to recording the sound of the motors themselves using piezo sensors. I want to continue modulating this sound and probably use it as the primary means of communication. Check some recordings here.

One interesting direction I came across is real-time neural audio synthesis. It probably won’t run directly on an RPi5, but it lets you work with a neural network that learns a compact representation of sound and can reconstruct it back into audio. You can also train your own models. In practice, you encode audio into a latent space and then manipulate or sample that space to generate new / expressive sounds with low latency. (It's very simple.)

How do you think robot like this should communicate?

---

Resources

Following up on the sources in the previous email, here are some more interesting discussions about xenobots and biological approaches to robotics with Michael Levin. There is also a recent update to the same interview here.

---

Suddenly, this web forum moltbook has received quite a bit of attention. It is intended strictly for AI agents (something like Reddit), where individual agents can debate and discuss certain topics. Of course, it's all mainly for effect, and all existential debates are just the result of a certain setup. But as a technological experiment, it's extremely interesting.

---

Grainer

Recently I made this simple Grainer (coded via Antigravity & Claude while testing some agentic coding tools) for creating high resolution grain zine-like aesthetics from images, without the need of boring Photoshop conversion and i no longer use Adobe (also time to switch to Affinity haha). Feel free to grab it and use it. Just download the folder and drag the index.html file into your browser. (in-app preview is a little bit lowres, because of the performance, so always export and check)

---

Books/25

I'm curious to know; what have you read last year? Any inspiring discoveries? I mainly listened to audiobooks, which became my favorite companion both when traveling or working. The ones that resonated with me are listed below.

I finally managed to go through some of a Dune saga (Dune, Dune Messiah, Children of Dune, God Emperor of Dune), captivating mainly because of its world-building and non-standard vision of the future. I guess it wouldn't be me if I didn't quote this profound thought from the book;

Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them. / Frank Herbert, Dune

I also read Mythago Wood, a fantasy fiction book very different from what I would normally read. About the subconscious, nature, and old legends.

Besides that, I reread some of my favorite novels, like Solaris by Stanislaw Lem, a story about a planet like no other. Philosophical science fiction where astronauts orbiting the planet are trying to understand the origin and behavior of the environment, while its living ocean affects their psyche.

From the more scientific literature, I finished the book Being You, A New Science of Consciousness by Anil Seth. A source of inspiration and wisdom that was particularly useful for my research of the System project.

Lastly, I'd like to mention one of my favorite thinkers, Josef Šafařík, because last year I had the chance to read his other book Člověk ve věku stroje (Man in the Age of Machines).

---

NFC

I've been thinking lately about NFC as a kind of self-contained medium. Almost an artifact of our times. Does anyone have some hands-on experience with producing or sourcing NFC tags/chips, and what’s possible with them?

---

#2026-W04

Work in progress;

What a packed week! Mainly because I finally decided to start this email list. Social media is increasingly consuming us all, and I feel it is beneficial to maintain some distance while still documenting (and publishing) the process.

---

Host / Guest is a project I am developing both as a continuation of my research and as part of my current V2.lab residency.

Why "Host"? The title should reflect the ambiguity of the object's existence. While the machine serves as a technological host for the algorithm, the word host in Czech translates to guest, framing the machine as a foreign visitor entering human space, observing our presence, and imitating us in order to adapt.*

In the last week (or rather weeks), I have been focusing mainly on finalizing the hardware of the object. After the System project, I decided to completely rethink the architecture of the object and design it almost from scratch. The machine is larger, more self-sufficient, and, above all, more stable. I finally decided to invest a little in better motors, and oh boy, it really makes a difference when the servos give you feedback. But still i managed to burn few of them. Anyway, I'm playing around with ideas for ground contact detection, calibration and overall system adaptability.

---

I also worked on calibrating the servo, because even with the best efforts, individual joints have small deviations and it is necessary to calibrate the robot's position. When started, each leg slowly moves toward its mechanical limit (a feature of the design), while the servo reports the current draw; when the flow increases rapidly, the machine knows it has found its limit. The difference between the expected and actual values becomes a deviation, thanks to which the machine recognizes its own limits each time it wakes up. > See how it works.

On the other hand, I am also developing a new audio system for Host. After some time, I am diving back into Max/RNBO software so that I can create audio patches directly for the Raspberry Pi. Visual programming brings a special kind of joy because you encounter completely different coincidences and errors than in regular written code. Among other things, I am preparing something like a workshop and resources for Max as an environment for visual programming of new media.

Host actively absorbs the soundscape, capturing fragments of it, which it processes and reorganizes. By resynthesizing these inputs into new sonic forms, the machine creates a unique form of communication, a primitive language through which it communicates with the world and signals its internal state.

---

Resources

There are some ideas about language that I find fascinating. That's why I'm interested to see how similar concepts can be applied to the behaviour of machines.

We are programmable machines / Language is pseudo code. / Dave ackley

The Latter Wittgenstein

Chomsky / Origins of Language

---

If you are interested into fresh research in robotics outside of American companies, check out this recent lecture by my favorite robotics professor, who talks about organic ways to get into robotics.

Bio-Colloquium Lecture

"Robots are defined not from what they’re made from, but what they do" ... "They’re machines designed by us to do something useful on behalf of humans" .., referring to Karel Čapek and the invention of the infamous word "robot."

---