Culture 3.14 · What we ask ourselves
What we ask ourselves
Questions that open a better way of understanding and building.
Open questions that cut across information, semantics, language, decisions, operations, agents, projects, Physical Intelligence and method. Each one states why it matters, what we know and what remains open.
A good question clarifies concepts, makes assumptions visible and makes it possible to design a test.
We gather questions that come up in several systems and that go on guiding decisions, experiments and conversations.
Each card distinguishes why it matters, what evidence we have, what we believe for now and what remains open. Some questions are answered by research; others only change shape when we try to build.
We publish questions, readings, systems and experiments that make it possible to follow how our judgement evolves.
Information
How do you represent a contradiction without hiding it inside an answer?
When two reliable sources say different things, summarising them into a single sentence destroys the most valuable information in the set.
How does a system know it has found enough evidence?
A system can retrieve ten relevant passages and still have nothing to support the claim it is about to write.
What should an organisational memory forget, and what should it keep?
A memory that keeps everything ends up retrieving expired decisions with the same authority as the ones still in force.
Language
How do voice and video change what informational quality means?
A text is reread; a narration is heard once and does not let the person receiving it go back to check.
How do you evaluate a shift in meaning that reads perfectly natural?
Automatic metrics reward fluency, and the most expensive error in a translation is precisely a fluent one.
What must stay invariant through a localisation?
Localising well requires deciding beforehand what cannot move, and that list is almost never written down.
Which part of a narrative should be generative and which deterministic?
The same piece contains data that admits no variation and explanations that need it.
Decisions
When does a confidence metric create a certainty the system does not yet have?
A figure that is well calibrated in aggregate can be badly calibrated for exactly the case somebody has in front of them.
When does a tabular model deliver more value than a language model?
On tabular data with labels and a known cost of error, gradient boosting still wins on accuracy, cost and the ability to analyse the result.
How can uncertainty be communicated without turning it into a misleading number?
87 % confidence reads as a promise, even when the model is only saying that it has seen similar cases.
What data was actually available when the decision was taken?
Reconstructing a decision with today's database produces an unfair judgement and an optimistic model.
When should a system recommend, decide, or do nothing but present evidence?
The three stances use the same model and distribute responsibility in three different ways.
What should a system optimise when time, cost, quality and risk compete with one another?
An operation rarely has a single objective; improving one metric can worsen another or shift the cost to a different phase.
Operations
Which responsibilities must stay deterministic in an agent system?
The agent can choose the route; permissions, calculations and writes should behave the same way in every execution.
What context does a bot need before acting on an ERP?
Most of an agent's work on an ERP happens before the first write, and consists of knowing what it cannot do.
How do you design an exception that no rule had anticipated?
The unforeseen cannot be enumerated, but you can decide in advance what happens to it.
What does it mean that a bot has finished a task?
Finishing the execution and producing the expected effect are two different facts, and usually only one of them gets recorded.
What should a process that may run for months remember?
A long-running process outlives deployments, changes of policy and sometimes the people who started it.
How can a system infer business logic without turning every historical habit into a rule?
Code, configurations, documents and traces show how the work has been done, but on their own they do not prove which logic ought to be kept.
Where should a person step in to contribute judgement without becoming a bottleneck?
Human intervention brings knowledge and authority, but an imprecise design can turn it into a queue of reviews without context.
How does a system learn from a resolved exception without generalising it too far?
An exception contains valuable operational knowledge, but it may depend on a context that disappears when it is turned into a rule too soon.
How is an operation shared between agents, deterministic software and people without losing a single point of responsibility?
Distributing tasks can improve the system, but it can also fragment the context and leave each component considering only its own work finished.
Physical Intelligence
What does robotics teach about silent software failures?
Robotics has spent decades assuming that a component can work and still produce no result at all; business software still discovers it case by case.
How should a system act when perception is not enough?
Stopping has a cost, carrying on with a poor estimate has another, and the choice has to be made before it happens.
Which decisions should stay local?
The link drops at the worst possible moment, and whatever is decided then was already decided beforehand.
Which signal shows that a physical action produced the expected result?
The command sent, the actuator moved and the object in place are three facts that often get merged into one.
Answers can close part of the problem and open a more precise question. We keep both so that the learning continues.