polysemantic neurons make naive saliency maps basically useless
a neuron lighting up tells you very little when it represents five unrelated things. saliency over raw neurons keeps leading people astray. features first, then saliency.
agree with the conclusion, not the reasoning
how much did this cost you in tokens to figure out
i love that this is just practical and not hype
the moment you add memory this gets way harder, fwiw
we run a similar setup, biggest pain was rate limits
my team is going to hate me for sending them this
the eval first mindset is underrated, nice to see it here
i would add: log everything, you will thank yourself later
how much did this cost you in tokens to figure out
solid. one nit: the naming is confusing
claude code handled this way cleaner for me tbh
great in theory, messy in practice from what I have seen
i think the spicy take is actually correct here
this matches my experience almost exactly
the eval first mindset is underrated, nice to see it here
we run a similar setup, biggest pain was rate limits
hard disagree honestly, in my testing it went the other way
this matches the anthropic docs almost word for word
i would love a follow up on the cost side of this