rohitg00--ai-engineering-from-scratch
wehub-resource-sync
91 行
3.4 KiB
JSON
91 行
3.4 KiB
JSON
{
|
|
"lesson": "11-planning-htn-and-evolutionary",
|
|
"title": "Planning with HTN and Evolutionary Search",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What does an HTN add over a free-form LLM plan?",
|
|
"options": [
|
|
"Provable correctness when operator preconditions and effects are enforced",
|
|
"Cheaper inference",
|
|
"Shorter prompts",
|
|
"Better embeddings"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "HTN's symbolic operators with preconditions and effects guarantee soundness by construction."
|
|
},
|
|
{
|
|
"stage": "pre",
|
|
"question": "Which problem class is AlphaEvolve built for?",
|
|
"options": [
|
|
"Free-form prose generation",
|
|
"Optimizations with a machine-checkable, deterministic fitness function",
|
|
"Multi-turn chat memory",
|
|
"Vector search ranking"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Evolutionary search needs a deterministic evaluator; AlphaEvolve targets domains where one exists."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "How does ChatHTN preserve plan soundness while using an LLM?",
|
|
"options": [
|
|
"It does not; soundness is best-effort",
|
|
"LLM suggestions only enter as candidate decompositions, validated against the operator schema; the symbolic layer owns correctness",
|
|
"It fine-tunes the LLM on HTN traces",
|
|
"It uses a vector database"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "The LLM expands the method library but cannot bypass operator preconditions and effects."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Which AlphaEvolve result does the lesson cite?",
|
|
"options": [
|
|
"First improvement over Strassen for 4x4 complex matrix multiplication in 56 years",
|
|
"10x speedup of inference on Gemini",
|
|
"Beating GPT-4 on HumanEval",
|
|
"First proof of P=NP"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "AlphaEvolve found 48 scalar multiplications for 4x4 complex matmul, the first improvement on Strassen in 56 years."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Which element of an HTN is a primitive directly-executable action with preconditions and effects?",
|
|
"options": [
|
|
"Task",
|
|
"Method",
|
|
"Operator",
|
|
"State"
|
|
],
|
|
"correct": 2,
|
|
"explanation": "Operators are the primitives; methods decompose compound tasks; state is a set of facts."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What is the lesson's warning about AlphaEvolve without a real evaluator?",
|
|
"options": [
|
|
"It is slow",
|
|
"Asking an LLM whether the code is better is not a fitness function; the evaluator must be deterministic and fast",
|
|
"It violates Apache 2.0",
|
|
"It cannot run on GPUs"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Without a deterministic evaluator the search has no signal to converge on."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "When should you reach for ReAct or ReWOO instead of HTN or AlphaEvolve?",
|
|
"options": [
|
|
"Never; HTN is strictly better",
|
|
"When you do not need formal soundness or a machine-checkable fitness; most agent tasks land here",
|
|
"When you have a GPU cluster available",
|
|
"When latency is below 100 ms"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "The lesson explicitly warns against over-engineering: most tasks do not need formal planning or evolutionary search."
|
|
}
|
|
]
|
|
}
|