#2026-W16 --- ### Host | Tier 1 Update A bit about [Host](https://kindl.work/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. ![[host-glyph-front.png]] --- ### Claude Design Anthropic this week released **[Claude Design](https://www.youtube.com/watch?v=t_LBECIQQqs)**, 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](https://www.graphicstandard.eu/), 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](https://kindl.work/2022/OSA+Visual+Identity)**, and the results are quite compelling [(see video)](https://vimeo.com/1184437424/0ceb615f95?share=copy&fl=sv&fe=ci). 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. ![[osa-kit.jpg]] *And someone is already starting to use it for [video editing](https://www.youtube.com/watch?v=ZNbgOhxhzXg) (motion graphics mostly).* --- ### Semantic Field This week, I worked further with [TWOZERO](https://twozero.ai/) 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](https://docs.dotsimulate.com/)),* 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](https://vimeo.com/1184443378/d04df34559?share=copy&fl=sv&fe=ci). ![[semantic-field.jpg]] --- > "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)*