security person nervously watching everyone wire agents into production
my whole career is telling people not to do the thing they are all now doing with agents. here to understand it well enough to help instead of just saying no.
this is why I check this forum every morning
the prompt injection example is terrifyingly simple
what does your eval setup look like
this is why I check this forum every morning
did this break for anyone else after the last update?
source? not doubting you, just want to read more
hard disagree honestly, in my testing it went the other way
interesting.... what was your testing setup? we saw the opposite with haiku vs sonnet on classification tasks but maybe our prompts were just bad
classification prompts matter way more than model choice under 20b imo. we had haiku beat sonnet on our task after we fixed the examples (was 68% vs 71%, flipped to 74% vs 72%). costs dropped like 6x so we kept haiku
wait you guys are getting haiku to beat sonnet on classification? we saw the opposite and assumed our prompts were just bad.... what did you change in the examples
we're running haiku for most tasks now, 70% cost drop and honestly the quality diff is minimal for our use case (tagging support tickets). sonnet is overkill unless you need deep reasoning
interesting, we saw the opposite - haiku was significantly worse on our classification task (tagging github issues by type). what were you classifying and how many examples did you use in the prompt? wondering if we just need better few-shot examples
i would love a follow up on the cost side of this
thank you for not making this a 20 minute video
genuinely useful, rare these days
this matches my experience almost exactly
the part about context windows is so true