The Quiet Wire is an exploration of Lower Cognitive Footprint AI: the idea that useful AI can ask less of us, not more. Fewer decisions, fewer interactions, less attention.
Most AI today asks the person to carry the system, to navigate, prompt, evaluate, and re-engage until an answer arrives, often drifting from what they set out to do. This series argues the opposite: treat constraints as the design, move the work of navigation from the person to the system, and ground it in a simple, highly-constrained substrate. The result is AI that recedes into the background. The measure of a good system becomes how easily it can be forgotten.
This is a personal portfolio exploring Human-AI interaction.
“The interactional cost of AI is one of the least regarded but most impacting constraints on its utility.”
Member of the Humanoid Robotics Association (Switzerland) (opens in a new tab).
Off the wire · 2026
Opportunities to connect: Zurich AI Festival, AI+X Summit, and NVIDIA GTC Berlin.
Attending
– · Zürich
A week of talks, workshops, and conversations across Zürich.
Explore the festival (opens in a new tab)
Attending
· StageOne Oerlikon, Zürich
I’ll be attending the AI+X Summit, part of the Zurich AI Festival. Let’s connect there.
Explore the summit (opens in a new tab)
Attending
– · Berlin
NVIDIA’s European developer conference, in Berlin.
Explore GTC (opens in a new tab)Going to one of these?Let’s discuss AI and grab a coffee.
Connect on LinkedIn (opens in a new tab)
Using LoRA on a single consumer GPU, character accuracy rose from 95.39% to 99.89% on the same 100 short test sequences, without an alphabet reference. Longer everyday sentences remained harder to decode reliably.
View experiment
Explore ideas through sound at your own pace, with tuning as part of the experience.
View project
Give BOB an architecture document. It identifies the connected actors and their interactions locally, then produces one narrated, standalone HTML Blueprint that can be inspected, shared or hosted without a generation service.
View project
A finite sequential game in which three AI agents hold different parts of a hidden code. Private messages, costly tests and shared resources shape the path each player takes toward a winning submission.
View project
Choose a subject once. A multi-agent system builds the curriculum, then researches the topic, writes a quiz, and produces a podcast each night. In the morning, it becomes a three-minute learning habit with no model latency on the daily interaction path.
View project
Each night, a local LangFlow workflow researches an event from exactly one hundred years earlier and writes a finished episode to iCloud. In the morning, Apple’s text-to-speech reads it live. A measured run uses 0.50 watt-hours.
View project
COLLABORATE preserves the common pool; DEFLECT improves the immediate private payoff while damaging it. Eight archived games preserve 135 turns and eight resource collapses for turn-by-turn comparison.
View projectCanonicalization, reconstruction and convergence show how a system can contain internal variation. The guarantee depends on a clear boundary and enough authoritative information to determine what crosses it.
Read articleA capable teacher model can create lessons and help construct the harness used to evaluate them. Fixed rules, deterministic gates and independent cases preserve the meaning of success while the deployed model remains replaceable.
Read articleAI can generate code faster than a person can review every mutation. This article argues that the useful place for human attention is higher in the system, where it can define boundaries, govern evidence, and limit the complexity that automated development is allowed to retain.
Read article