Every loud AI claim sounds persuasive in isolation.
"RAG is dead." "Agents replace all developers." "Prompting is obsolete." "Model moats are over." The sentence format is always the same: declarative, absolute, and emotionally efficient.
Your job is to slow it down.
In this course, we will use a four-question read. Run all four before the claim is allowed into planning.
- Who said it?
- What is their operating position?
- What is their time horizon?
- What is the reversibility of acting on it?
This method extends Learning in the Step-Change: discourse can be useful, but only after epistemic hygiene.
Let us unpack each question.
Who said it?
This is provenance, not personality worship. You are asking whether the claimant has direct contact with the problem they describe.
Someone shipping retrieval systems at scale discussing RAG (retrieval-augmented generation, where models fetch external context at runtime) has different evidentiary weight than someone reacting to screenshots. Someone running a coding team in production has different signal than someone reviewing demo clips.
Do not confuse polished communication with evidentiary depth.
What is their operating position?
Position determines what they can see and what they cannot.
A model vendor sees aggregate usage patterns and benchmark trajectories.
An application operator sees user behavior, support burden, and workflow failure points.
A consultant sees many client patterns but often with limited longitudinal ownership.
A commentator sees second-order narratives.
None of these positions is useless. But each has blind spots. If you do not account for position, you will over-generalize local truth.
What is their time horizon?
Many claim failures are horizon failures. A statement can be directionally right over five years and operationally wrong for next quarter decisions.
"RAG becomes less central as context windows grow" might be a long-run trend claim.
"Stop investing in retrieval infrastructure this quarter" is an immediate allocation claim.
Those are not the same sentence.
Force the horizon into explicit terms: this quarter, next year, three years.
What is reversibility?
This is the grown-up question most discourse ignores. If you act on the claim and it is wrong, how expensive is undo?
Adopting a pilot workflow with rollback is reversible.
Replatforming architecture and deprecating critical systems is less reversible.
Hiring plans based on "developers are obsolete" are painfully irreversible at human level.
Reversibility tells you how much evidence you need before action.
Now let us run this on one live-style claim: "Agents replace all developers."
Who said it?
Imagine the claimant is a creator who demos autonomous code agents on curated tasks. Strong demo skill, unclear production ownership. That matters.
Operating position?
They are close to tool capability but may be far from enterprise delivery constraints: legacy systems, unclear requirements, security review, integration tax, and organizational coordination.
Time horizon?
Most such claims blend horizons. Near-term demo success is used to imply near-term labor replacement. That leap is usually unsupported.
Reversibility?
If you respond by reducing junior hiring, cutting mentorship, and redesigning team structure around assumed full autonomy, reversal is expensive and culture-damaging.
So what is the decision-quality output from the four-question read?
Not "ignore agents."
Rather: "Treat agents as force multipliers for scoped tasks; test where supervision burden is low; preserve human architecture and review capability until evidence shows sustained autonomy in your own environment."
That is a strategic sentence. It has shape, boundaries, and rollback.
Now bring in GitHub Copilot's adoption curve, because it is the perfect warning against lazy dismissals. Early critics framed Copilot as "just autocomplete," which was technically accurate and strategically wrong. Why wrong? Because they failed the same four-question read in reverse.
Who was speaking?
Many observers were outside day-to-day coding workflows.
Operating position?
They judged by novelty of model behavior instead of compounding time-saved inside existing developer flow.
Time horizon?
They evaluated immediate wow factor and missed medium-term habit formation.
Reversibility?
Teams that experimented lightly had low downside and discovered strong productivity gains; teams that dismissed outright lost learning time.
The lesson is symmetrical. Do not swallow hype. Do not dismiss emerging tools with snark either. Use structure.
A practical template for team discussions:
Claim statement: one sentence.
Provenance notes: source, direct experience, evidence quality.
Position notes: vendor/operator/researcher/commentator, with blind spots.
Horizon notes: immediate, annual, structural.
Reversibility notes: what action is being implied, and the rollback cost.
Decision: monitor, pilot, adopt, or reject for now.
When you use this template repeatedly, two things happen.
First, your internal debates get calmer because people argue on dimensions instead of vibe.
Second, your learning speed improves because you run better pilots instead of binary takes.
Do not skip the reversibility step. It is the bridge between discourse and execution.
A PM in Bengaluru deciding whether to sunset a search UI after seeing "search is dead" threads should immediately ask reversibility. A full sunset is costly to undo. A dual-surface experiment with clear cohort metrics is cheap to undo. Same claim, different action quality.
Apply lever, risk, rollback explicitly.
Lever: The four-question read turns noisy opinion streams into structured decision inputs.
Risk: Bureaucratic overuse can slow action if every minor claim gets full ritual treatment.
Rollback: Use lightweight mode for low-stakes claims; full mode only when the implied action affects architecture, hiring, pricing, or trust.
One final point: provenance is not seniority theater. A junior engineer with direct production evidence can outrank a famous commentator on this framework. Likewise, a famous operator can still be wrong outside their domain. Respect evidence, not aura.
In the next lesson, we expand this from single-claim analysis to portfolio design: building your personal frontier-watcher's dozen so the upstream input quality improves before claims even reach your team.
Rules from this lesson
- Never let a claim enter planning without answering who said it, from what position, on what horizon, with what reversibility.
- Separate directionally true long-run statements from near-term operational decisions.
- Demand higher evidence when the implied action is hard to reverse.
- Use structured pilot decisions instead of binary hype-or-dismiss reactions.
- Trust provenance and operating evidence over audience size or rhetorical confidence.