coding with claude code is basically like playing a simulation video game
low effort direction giving
kinda tedious sometimes
fast-ish progress
feel good reward built in by seeing progress
the average engineering job was never really about writing code
it was just building out the product and designers vision
the job still remains, you need people prompting the thing to build it
you just need fewer of those
at this point:
- unit tests dont matter
- the code, whether clean or garbage, doesn't matter
- architecture, doesn't matter
all that matters is: does the product satisfy the user journeys specified
this stuff needs waay higher limits
my interpretation of loops:
- describe what you need
- have the ai write out user journeys that need to be satisfied for the product to be built
- wait for it to keep iterating until it's done
openai removed 5.3-codex from subscription
it was their best price-to-performance coding model yet
5.5 is about 3x more expensive than 5.3-codex, and definitely not 3x smarter
exactly this
most things that are satisfying and easy to consume, such as answers from LLMs, result in a feeling of understanding
true understanding by humans requires repeating what was ingested, without which you just feel you got it
if you can explain it, you got it
a failure mode i am seeing lately.
with llms, it's easy to get an illusion of understanding by substituting precise low-level details (that matter) with coarse high-level analogies.
it's an illusion because such understanding falls apart when you try to use or apply it.