PO: What If AI Became More Useful by Knowing Less?
From context accumulation to context intelligence.
Generative View Lateral Thinking Context Engineering AI AgentsAIYING — From AI Awareness to Accomplishment
1. The Conventional View
We often believe that an AI system becomes more useful when we give it more information.
More documents. More conversation history. More instructions. More examples. More memory.
This belief has a reasonable foundation. AI systems frequently need relevant context to answer accurately and perform tasks.
But consider a different possibility.
What if the problem is that the AI is carrying too much irrelevant, outdated, conflicting, or distracting context?
A system can have access to abundant information and still fail to identify what matters for the task at hand.
2. The Hidden Assumption
Behind the familiar idea of "give AI more context" lies an assumption:
Sometimes this is true. But it is not a universal rule.
Information can become stale. Documents can contradict one another. Long histories can bury the latest decision. Retrieved text can be irrelevant or malicious.
Context engineering therefore involves not only adding information but also selecting, prioritizing, separating, and sometimes removing it.
3. The PO Provocation
The AI must forget information to become more useful.
This statement is deliberately provocative.
It reverses the normal direction of thought: instead of asking what more information the AI should receive, we ask what information it should not carry forward.
In Edward de Bono's Lateral Thinking, a PO provocation is not a factual claim that must be defended.
It is a device for moving away from an established pattern toward new possibilities.
We are not proposing that AI systems discard essential facts, safety constraints, audit records, or legally required information.
The provocation is about rethinking what enters the model's active working context. Records retained for accountability may be managed separately.
4. Movement: What Can We Discover?
We now move forward from the provocation rather than arguing whether it is literally correct.
If forgetting can improve performance, perhaps relevance matters more than quantity.
Discovery: Supply only the context needed for the current decision.
A human expert does not reread every document before every decision.
Discovery: Retrieve targeted evidence when needed instead of loading everything in advance.
At what point should old context stop influencing a new task?
Discovery: Introduce context checkpoints between workflow stages.
What if an AI agent began each major task with a clean working space?
Discovery: Carry forward verified decisions, required constraints, and a concise task state rather than every message.
5. Five Possibilities for AI Engineering
Possibility 1 — Just-in-Time Context
Instead of supplying an entire knowledge base, retrieve relevant material at the moment it is needed.
This is one useful design direction for retrieval-based AI systems.
Possibility 2 — Context Pruning
Remove repeated, outdated, or low-relevance material from the active context.
The challenge is to prune without losing essential evidence.
Possibility 3 — Task-Specific Memory
An AI assistant may need different information for different tasks.
A coding task may need source files and tests. A customer support task may need product documentation and the customer's current issue.
A single undifferentiated memory is not always the best design.
Possibility 4 — Fresh-Context Verification
After generating an answer, a separate verification step can examine the result against authoritative evidence.
Separating generation from verification can help expose unsupported assumptions. It does not guarantee that every error will be detected.
Possibility 5 — Context Boundaries
Not every document or tool result should have the authority to instruct the AI system.
Untrusted retrieved content should be treated as information to assess, not as a replacement for trusted system or developer instructions.
This matters for prompt-injection resistance and secure agent design.
6. An Example: The AI Project Assistant
The Situation
An AI project assistant has access to months of meeting notes, emails, project plans, issue logs, and status reports.
A manager asks:
"What are the three most important risks for this week's release?"
Conventional Approach
Load as much project history as possible and ask the AI to identify the risks.
Generative View
Begin with the current release, its latest verified status, unresolved critical issues, and relevant recent decisions.
Retrieve older records only when they are necessary to explain or validate a risk.
What Changes?
The design shifts from maximum context toward minimum sufficient, trustworthy context.
7. From Provocation to Practical Design
| Conventional Pattern | Generative Possibility |
|---|---|
| Keep all conversation history | Carry forward a verified task summary and retrieve details when needed |
| Load entire documents | Retrieve relevant sections with source references |
| Use one memory for everything | Separate memory by task, role, or workflow |
| Trust the accumulated context | Check freshness, provenance, and contradictions |
| Use the same context for every step | Assemble context appropriate to each stage |
8. Try the Provocation With Generative AI
Tool: ChatGPT
Choose a real but non-sensitive task that involves a lot of information.
For example: preparing a report, summarizing a project, analyzing a document, or planning a learning activity.
Use the following prompt.
You are my Context Engineering Design Partner. My task is: [Describe the task] Available information: [List the documents, facts, instructions, and history] Use this Lateral Thinking provocation: PO: The AI must forget information to become more useful. Do not treat PO as a literal recommendation. Follow these steps: 1. Identify the conventional assumption about context. 2. Separate essential, useful, optional, outdated, and potentially untrusted information. 3. Use deliberate movement from the PO provocation to generate five alternative context designs. 4. For each design, explain what information is included, excluded, or retrieved later. 5. Identify risks of removing too much context. 6. Recommend one design for a small controlled test. 7. Propose three evaluation criteria to compare it with the original approach. Do not invent facts. Make assumptions explicit.
Your Small Accomplishment
Produce two versions of the same task:
- Version A: Use all available relevant context without deliberate pruning.
- Version B: Use a carefully selected context package.
Compare them for:
- Accuracy and factual support.
- Relevance to the task.
- Completeness of essential details.
- Unsupported assumptions.
- Effort and time required.
Do not assume Version B will win. The purpose is to discover when selective context helps, when it hurts, and why.
9. The Larger Discovery
It is also the engineering of what information deserves the AI's attention, at which moment, for which purpose, and under whose authority.
This is the transition from simply accumulating context to intelligently managing context.
The PO provocation has helped us discover a different design question:
What is the smallest trustworthy context sufficient for this task?
The answer will depend on the task, the risks, and the evidence required.
10. The AIYING Five-Format Journey
This is the second post in our repeating AIYING publishing sequence.
- V1.0 — Regular View: Understand Context Engineering.
- V2.0 — Generative View: Challenge context accumulation through PO and movement.
- V3.0 — A(IQ): Test your judgment about conflicting AI evidence.
- V4.0 — AI Lab: Experiment with a real AI workflow.
- V5.0 — AI in Action: Examine real-world AI outcomes.
Next: AIYING V3.0 — A(IQ) #002
When three AI models agree, are they necessarily right?
No comments:
Post a Comment