Letting the Team Learn Its Own Positions, Then Coaching From There

Opening Scene

Instead of assigning positions to a brand-new team on day one, a coach does something deliberate: lets the team scrimmage freely for the first two weeks, the same open, unstructured process from Article 2, and watches what natural roles emerge. Only after that does formal, targeted coaching begin — refining what the team organically discovered about itself rather than overriding it with a rigid, pre-decided structure. The unsupervised exploration didn’t replace coaching; it gave the coaching something real and specific to build on.

That two-phase approach — unstructured discovery first, targeted correction second — is exactly how many of today’s most capable AI systems are actually built, combining Article 2’s unsupervised learning and Article 1’s supervised learning into a single, deliberate pipeline rather than treating them as competing approaches.

In Plain English

Modern AI systems increasingly combine unsupervised (or self-supervised) learning and supervised learning in sequence: a model first learns broad structure from a large amount of unlabeled data, then gets refined with a smaller amount of labeled data for a specific task. This isn’t a compromise between the two approaches — it’s a genuinely more effective use of each one’s strengths, learning general structure where labels are scarce and expensive, then specializing precisely where labels are available.

The Old Way

Before this combined approach was standard, teams generally had to pick one lane:

A coach who only ever assigns rigid positions from day one, missing whatever natural talent didn’t fit the predetermined roles.
A coach who never moves past open scrimmage into any real structure, leaving real potential undeveloped for lack of targeted coaching.
A school that either lets students explore entirely freely, or drills a fixed curriculum with no room for discovered strengths — rarely both, in sequence, deliberately.

Choosing one extreme or the other left real value on the table either way — either genuine natural talent went unrefined, or genuine natural fit was overridden by a rigid predetermined structure.

What’s Changing (and Why AI Is the Reason)

This combined approach — broad unsupervised or self-supervised pretraining followed by targeted supervised fine-tuning — is now the default pattern behind many of the most capable modern AI systems, not a niche technique. It directly resolves the earlier articles’ tension between wanting real labels and having a limited labeling budget.
The proportion of unsupervised versus supervised effort is increasingly a deliberate design choice, tuned to how much labeled data is realistically available for the specific task, rather than an all-or-nothing decision between two separate approaches.
This pattern makes AI systems dramatically more data-efficient for any single new task, because most of the heavy lifting — the broad structure-learning — happened once, upstream, and gets reused across many different downstream, labeled fine-tuning efforts.

The Metaphor, Fully Extended

Team Development	Combined Learning Concept
Two weeks of open, unstructured scrimmage	Unsupervised or self-supervised pretraining
Natural roles the team discovered on its own	Broad structure learned without labels
Targeted coaching layered on top afterward	Supervised fine-tuning on labeled, task-specific data
A coach overriding natural talent with rigid assigned positions from day one	Skipping pretraining, relying purely on limited labeled data
A team that never gets targeted coaching after open scrimmage	Skipping fine-tuning, leaving raw structure unrefined
A well-coached team that still plays to its natural strengths	A fine-tuned model that retains useful pretrained structure

For Beginners: What to Actually Do

Recognize this combined pattern by name when you encounter it — “pretrained then fine-tuned” is describing exactly this two-phase process from this article.
Understand that the unsupervised phase isn’t a lesser substitute for labeled data — it’s doing genuinely different, complementary work that labeled data alone can’t easily replicate.
When facing a limited labeling budget, consider whether a relevant pretrained starting point exists before assuming you need to label a large dataset from scratch.

For Practitioners and Leaders: The Deeper Layer

Evaluate whether your team’s real bottleneck is genuinely a lack of any labeled data, or specifically a lack of a strong pretrained starting point to fine-tune — the right investment differs depending on which is actually true.
The quality of the unsupervised pretraining phase still matters enormously; a weak or mismatched pretrained foundation limits how much fine-tuning alone can fix.
This combined approach doesn’t remove the need for the data quality discipline covered throughout this series — labels still need to be accurate, and pretraining data still needs to be reasonably representative of the eventual task.

Quick Recap

Combining unsupervised or self-supervised pretraining with supervised fine-tuning is the default pattern behind many capable modern AI systems.
It mirrors letting a team discover its own natural structure first, then layering targeted coaching on top of what’s real.
This pattern makes AI systems far more data-efficient for any single new task, because broad learning happens once and gets reused.
The proportion of unsupervised versus supervised effort is a deliberate, tunable design choice, not an all-or-nothing decision.

Where This Fits in the Series

Article 18 covered reusing knowledge across genuinely different tasks; this article covered combining the series’ two central approaches into a single, sequential pipeline. Article 20 closes the series, bringing supervised and unsupervised learning together as one connected picture of how a team — and a model — actually learns.