● Home Knowledge Action Journal Homework Labs Study Try Case Make Training
Encode Cards Study History Free Recall Quizzes Digests Tasks Conjectures Belief Gym Audio Timeline Knowledge Items People Templates Outcomes DMXs Situation Analysis Council Debate Parable Branch Learning Story Engine Content Engine Explainer Rehearsal Bench Import Kindle Notes

📐 Encode

Every card for one document, both halves open. Nothing here is graded, scheduled or due — tick the ones that are still abstract and queue an example for each.

20 cards · 0 with examples
Getting the most out of superintelligent ai models
Q When interacting with a superintelligent AI, utility transfer is capped by the bounded receiver's three scarce resources. What are they?
A Specification fidelity, comprehension bandwidth, and verification capacity
new ai-interaction scalable-oversight cognitive-limits
Q In AI safety research, what is 'scalable oversight'?
A The problem of how a weaker agent reliably extracts work from a stronger one it can't fully verify.
new ai-safety scalable-oversight superintelligent-ai verification
Q When working with a superintelligent AI, why should you demand outputs in verifiable form (tests, worked examples, falsifiable predictions) rather than free-form answers?
A Verification is cheaper than generation — you force the model to spend its surplus intelligence making your cheap check possible, so you can confirm work you could never produce yourself.
learning ai-interaction verification p-vs-np protocol-design
Q Why can a human reliably rank AI outputs they couldn't have generated themselves?
A Comparison is cheaper than construction — the cognitive cost of ordering options is far lower than the cost of generating them.
new ai-interaction cognitive-asymmetry rlhf spaced-repetition
Q Why is asking an AI for alternatives + tradeoffs more effective than asking it to generate a solution from scratch?
A Ranking/choosing is cheaper than generating — your preference signal stays high-value even when your generative capacity is fully outmatched.
new ai-interaction cognitive-asymmetry prompt-design decision-making
Q Why is cross-examining two AI models against each other easier than auditing one model's answer alone?
A Each model surfaces the other's weaknesses for you — you only need to judge the disagreement, not independently verify the content.
new ai-workflow verification cross-examination scalable-oversight
Q Why are artifacts a better output format than advice when working with a smarter AI system?
A Advice is capped by your comprehension; artifacts (code, spreadsheets, documents) are capped only by your testing infrastructure.
new ai-workflow scalable-oversight verification artifacts-vs-advice
Q When you can't verify every AI output, what is the practical alternative to per-instance verification?
A Check a random sample, compute an error rate per domain, and calibrate your reliance on that measured rate — not on how confident the output sounds.
new ai-workflow verification statistical-trust epistemics
Q When using a superintelligent AI model, why does comprehension difficulty depend on representation rather than being fixed?
A A great explanation is a search over representation-space for one that fits your priors — the same content can be compressed into your ceiling if the model targets your vocabulary and prior knowledge.
new ai-interaction comprehension representation scalable-oversight
Q When a claim from an AI model can't be verified by checking, cross-examination, or comparison, what is the fallback strategy?
A Convert it into a cheap, reversible experiment (pilot, A/B test, small bet) — outcomes are an intelligence-free verifier.
new ai-workflow verification epistemics experimentation
Q Why are artifacts more valuable than advice when working with a superintelligent AI model?
A Advice is capped by your comprehension; artifacts are capped only by your ability to test them.
new ai-interaction productivity scalable-oversight verification
Q When prompting an AI, why should you demand 'checkable form' rather than accepting a bare answer?
A Verification is cheaper than generation — you can confirm work you couldn't produce, so forcing the model to include tests, worked examples, or falsifiable predictions lets you spend minimal effort checking rather than trusting blindly.
new ai-interaction verification prompt-design scalable-oversight
Q When working with a superintelligent AI, why is asking for multiple alternatives better than specifying a single solution upfront?
A Ranking/comparing outputs is cheaper than generating them — your ability to choose good options outlasts your ability to specify or create them.
new ai-interaction prompt-design cognitive-asymmetry rlhf
Q When working with a superintelligent AI, why should you start hard tasks with 'Ask me clarifying questions until you have everything you need, then proceed'?
A It moves the specification burden to the smarter side — the model figures out what it needs rather than you trying to perfectly articulate it upfront.
new ai-workflow prompt-engineering scalable-oversight
Q Why is cross-examining an AI answer (e.g., asking it to argue against itself) more effective than directly auditing it?
A Judging a fight is easier than auditing a monologue — adversarial critique surfaces weaknesses for you, so you spend less verification effort than checking a single answer directly.
new ai-interaction verification critical-thinking scalable-oversight
Q Why does reversibility allow you to extend trust to a system you don't fully understand?
A Because it bounds your worst case — if something goes wrong, you can undo it, so incomprehension no longer implies unbounded risk.
new ai-workflow delegation risk-management reversibility
Q When calibrating how much to trust an AI model's outputs in a given domain, what should you base your trust on?
A The measured hit rate from spot-checked claims — not how confident the output sounds.
new ai-workflow epistemics verification calibration
Q When an AI's claim is unverifiable by reasoning or cross-examination, what is the fallback strategy?
A Convert it into a small, cheap, reversible experiment (pilot, backtest, A/B) — reality is an intelligence-free verifier.
new ai-workflow verification epistemics experimentation
Q When asking an AI to explain something complex, whose job is it to compress the explanation to your level of understanding?
A The model's — compression is its job, not yours. Instruct it to use your strongest domain as a reference point and break it into checkable steps.
new ai-interaction learning protocol-design comprehension
Q Why is synchronous chat a bottleneck when working with AI on large tasks?
A Throughput is capped by your reading and typing speed; async delegation removes your latency from the loop and lets you review artifacts instead.
new ai-workflow productivity async-delegation