mirror of
https://github.com/wassname/ml_debug.git
synced 2026-08-28 12:52:14 +08:00
fail fast is advice for a human who over-commits, agents quit early instead
This commit is contained in:
@@ -134,6 +134,16 @@ A separate thing that shares the name "fail fast", and worth keeping separate in
|
||||
That one is about killing a doomed direction early. The one above is about crashing on the error.
|
||||
Both are good and they are not the same rule.
|
||||
|
||||
Read the first one with its audience in mind. Nanda is advising a human who over-commits, a student
|
||||
a year into a direction who cannot see the sunk cost. Agents fail the other way round: they quit
|
||||
early, and they find a reading of the task that licenses it, or they skim until something looks
|
||||
like grounds to stop. So the rule does not transfer unchanged. Before you call a direction dead,
|
||||
show the receipt from exercises 7 and 9: what you expected, what you got, and the bug you ruled
|
||||
out. A reason found while skimming does not count.
|
||||
<!-- CLAUDE: wassname's point, my wording. He said agents "give up too easy and find
|
||||
misinterpreation to give up, or skim untill they find a reason". -->
|
||||
|
||||
|
||||
## How this applies to LLM agents
|
||||
|
||||
LLMs of 2026 are trained to compress speech and use folky or humanistic language, but it's better
|
||||
|
||||
Reference in New Issue
Block a user