Facilitator: Venkatesh Rao
Participants: Rich McDowell, Patrick Nast, Timber, Seth Killian, Giovanni Merlino, plague_year, .unipuff, Ceciliyazi
Topic: Observability – understanding and measuring internal system states via external outputs, with applications from control theory to protocol design.
- Observability (Control Theory): Can you deduce the internal state of a system from its outputs?
- Origins: Mathematical analysis of dynamical systems (e.g., checking matrix rank).
- Expanded Usage: Now applied in computing, cloud systems, AI, and protocol governance.
- Historical Lesson (Cybernetics): Models built without observable data often fail due to lack of empirical grounding.
- Approach: Start protocol modeling from observability, not abstract system design.
- Examples:
- Chess vs. Card Games: Chess is fully observable; card games are partially observable.
- Constellations: Only 2D observable from Earth; third dimension hidden.
-
.unipuff: Asked whether Go is less observable than Chess. (Clarified: both are observable; the difference is in computability, not observability.)
-
Timber: Observability of a wild animal's intent in the woods—real-world unpredictability.
-
Seth Killian:
- In games, we log observable events but not player psychology.
- Observability sets boundaries for agency and decisions.
-
Rich McDowell:
- Wind turbine under load—same data can imply multiple hidden realities.
- Limits of observability → limits of control.
-
plague_year:
- Observability in parenting toddlers—overlapping adult attention as “sensors.”
- Broader metaphor for distributed observational coverage.
-
Patrick Nast:
- HVAC upgrade (adding a sensor) increased system observability and control.
- Introduced idea of engineered vs. practical observability.
-
Giovanni Merlino:
- Transparent: digital thesis submission (clear actor visibility).
- Opaque: university policy changes and opaque publication protocols.
-
Ceciliyazi:
- From a journalism/filming perspective: observers shape the observed.
- Directional cameras influence what becomes visible.
Prompt: How would you observe whether countries are complying with a global climate treaty?
- Venkatesh Rao:
- Low-level: ML-aided chainsaw sound detection in the Amazon.
- High-level: Analyze body language of summit delegates to infer true intentions.
-
Ceciliyazi: AI-generated satellite imagery using geolocation and keywords.
-
Giovanni Merlino: Analyze anomalies in financial instruments (e.g., green bonds) as market signals of compliance or risk.
-
Patrick Nast:
- Forecast emissions using diplomatic documents and communications.
- Use protocol structure (e.g., transparency reports) as a data source.
-
plague_year:
- Citizen science via home waste photo uploads.
- Instrument water, gas, and power usage for micro-level observability.
-
Rich McDowell:
- Adversarial observability: countries spy on each other to maximize detection of non-compliance.
- Indirect benefit: unexpected observables and even cooperation.
-
Seth Killian:
- Concerned with timescale mismatch in logs/metrics—limits of temporal observability.
- Observability is foundational for reliable models and governance structures.
- Design protocols with observability in mind, rather than treating it as an afterthought.
- Measurement is political and partial: what you observe depends on where you stand.
- Complex systems require layered, multi-scale observability strategies.
Transcription by SeaVoice • Discussion hosted on Discord voice channel.