What Learns?

TL;DR

Learning is persistent change driven by feedback — not every change counts.
Find the loss function — what is the system implicitly optimizing?
Memory spans timescales — cache, habit, model weights, culture.

How to use this lens

Ask what gets updated after each outcome — weights, habits, policies, beliefs.
Identify the feedback signal — reward, error, surprise, social approval.
Locate memory — where is state stored and how long does it last?
Check exploration vs exploitation — is the system trying new actions or repeating winners?
Watch for stale learning — old data dominating new reality.
Name who or what learns — person, team, algorithm, ecosystem.

Core concepts

Takeaway: Optimization requires a target.

Even organisms without explicit math minimize prediction error or maximize inclusive fitness. Offices optimize revenue, morale, or survival — often simultaneously with tradeoffs unstated.

Takeaway: Learning rates matter.

Update too fast and noise dominates; too slow and regimes shift before you adapt. Statistical learning formalizes bias-variance tradeoffs that appear everywhere from sleep consolidation to model retraining.

Takeaway: Credit assignment is hard.

Which past action caused today’s outcome? Teams blame individuals; algorithms use backprop or eligibility traces. Misassigned credit teaches the wrong lesson.

Memory tiers

Tier	Example	Decay
Working	Current task context	Seconds–minutes
Episodic	Yesterday’s meeting	Days–years
Structural	Org policy, muscle memory	Months–forever
Parametric	Model weights	Until retrained

Mini-examples

Sleep: consolidates episodic memory into long-term storage; synaptic homeostasis resets learning capacity.
Office: retrospectives update process; CRM records customer history; ML models retrain on logs.
Blood: baroreceptors tune heart rate in seconds; cardiovascular fitness adapts over weeks of training.

Curriculum bridge

Mechanisms of learning in Chaitra:

Statistical Learning — generalization from data
RL Orchestration — agents acting to maximize reward

Try it now

What did you update after your last mistake — a habit, a checklist, nothing? Where is that memory stored (brain, notes, team wiki)?

Going deeper

Takeaway: Multi-agent learning creates games.

When many learners co-adapt, equilibria shift. A pricing bot and a competitor bot learn each other, not just the market. Office politics is multi-agent reinforcement learning without clear rewards.

Takeaway: Learning can be destructive.

Overfitting, burnout, and regulatory capture are learning failures. Good design includes forgetting — dropout, sleep, policy sunsets.

Takeaway: Orchestration schedules learning.

When to retrain, when to sleep, when to run an retrospective — timing is infrastructure. RL orchestration applies to machines and meetings alike.

If nothing learns, the system is open-loop — fine until the environment moves.