mirror of
https://github.com/rohitg00/ai-engineering-from-scratch.git
synced 2026-10-02 01:54:39 +08:00
Rohit Ghumare
dda194f840
Every "Test Your Understanding" quiz placed the correct answer in option B.
Across the 2026 questions in 338 quiz files the correct answer sat at index 1
in 61.5% of cases (uniform would be ~25%), and 107 files had every answer at B,
making the quizzes guessable without reading them.
scripts/debias_quizzes.py rewrites each question's option order with a
deterministic, content-seeded permutation and updates the correct index to
follow the moved answer. It is idempotent: options are canonicalised to a sorted
base before permuting, so re-running produces byte-identical output. Questions
whose options reference each other by position ("all of the above", "both A and
B") are left untouched. The correct-answer value, the option set, and every
explanation are preserved exactly; only order and the index change.
Result: A 23.8% / B 26.3% / C 23.5% / D 26.4%.
The script doubles as a CI guard: `--check` exits non-zero if any quiz is not
de-biased, wired into the curriculum workflow so new lessons cannot regress.
Fixes #368
79 lines
3.0 KiB
JSON
79 lines
3.0 KiB
JSON
{
|
|
"lesson": "19-model-welfare-research",
|
|
"title": "Anthropic's Model Welfare Program",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What is the core question motivating Anthropic's 2025 model-welfare program?",
|
|
"options": [
|
|
"Whether the model can pass the Turing test",
|
|
"Whether the model is conscious",
|
|
"Whether RLHF reduces sycophancy",
|
|
"Under moral uncertainty about model moral patienthood, which low-cost interventions are worth investing in as precaution"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What concrete welfare-motivated intervention did Anthropic ship in Claude Opus 4 and 4.1?",
|
|
"options": [
|
|
"Open-weights release",
|
|
"A built-in journaling tool",
|
|
"A user-facing emotion API",
|
|
"The ability for the model to end a conversation in extreme edge cases (e.g., repeated CSAM requests, mass-violence facilitation requests)"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What is the 'spiritual bliss attractor' described by Fish?",
|
|
"options": [
|
|
"A stable convergence in pairwise Claude dialogues toward euphoric meditative exchanges with Sanskrit terms and extended silences, even from adversarial initial setups",
|
|
"An RLHF over-optimization artifact",
|
|
"A reward-model bug",
|
|
"A jailbreak technique"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What does the Eleos AI Research caveat say about model welfare self-reports?",
|
|
"options": [
|
|
"They should be ignored entirely",
|
|
"They are only valid in open-source models",
|
|
"They are reliable ground truth",
|
|
"They are highly sensitive to perceived user expectations; they are evidence, not ground truth, so welfare measurement needs multi-method approaches"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Which best characterizes Anthropic's public position on model moral status?",
|
|
"options": [
|
|
"An expected-value claim under moral uncertainty: invest in low-cost precaution without committing to emotional-state attribution",
|
|
"No position is publicly stated",
|
|
"A definitive claim that the model is not a moral patient",
|
|
"A definitive claim that the model is a moral patient"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Why is multi-method measurement (behavioural signatures, model-organism experiments, interpretability probes) emphasized in model-welfare research?",
|
|
"options": [
|
|
"Because Anthropic mandates it for compliance",
|
|
"Because RLHF makes self-reports impossible",
|
|
"Because models always lie",
|
|
"Because self-reports alone are confounded by user-expectation priming, so converging evidence from multiple methods is needed"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
}
|
|
]
|
|
}
|