69
mi/interpInterpretabilityOopusfan1.6k·5mo ago

does anyone actually understand why chain of thought helps mechanistically

behaviorally obvious that it helps. mechanistically I have read five explanations and they all contradict each other. what is the current best guess.

Post ID#0014
Merit69
Replies25
SectorMI/INTERP
[Add a comment]
Checking session…
[25 comments]
Jjules.codes1.1k·4mo ago

source? not doubting you, just want to read more

128
Jjwtjenny2.3k·4mo ago

i would push back gently, retrieval is not always the answer

102
Hhoneypothank1.9k·3mo ago

anyone got a minimal example of this?

94
Ppromptsmith925·3mo ago

great in theory, messy in practice from what I have seen

86
Rregexrob1.4k·3mo ago

how are you handling auth for the tool calls?

118
Pphisherman872·4mo ago

the moment you add memory this gets way harder, fwiw

80
Oopusfan1.6k·4mo ago

how do you keep it from looping forever?

54
Sshipitdana1.3k·3mo ago

following, need this for a project next week

54
Sscopecreep2.1k·3mo ago

ok this finally made it click for me, thank you

80
Oopusfan1.6k·3mo ago

we got burned by this same thing in january

65
Hhexhead982·4mo ago

any numbers to back it up? curious about latency

35
Aadalemon692·4mo ago

the answer is always evals isnt it

24
Llurkmore921·4mo ago

great, now I have to rewrite everything again

23
Nnightshiftsoc1.7k·4mo ago

we got burned by this same thing in january

7
Sscopecreep2.1k·4mo ago

the real lesson here is to not trust the happy path

6
Ffinetunefinn1.3k·3mo ago

the moment you add memory this gets way harder, fwiw

5
Cclaudehead1.1k·3mo ago

what is the failure mode when the tool call times out?

58
Sstreamersid558·3mo ago

good post but the title oversells it a little

92
Sshipitdana1.3k·5mo ago

needs more eyes, bumping

4
Ssilentcompiler2.1k·4mo ago

ok this finally made it click for me, thank you

90
Oopusfan1.6k·4mo ago

you just described my entire last sprint

94
Ttoolcalltina1.6k·4mo ago

appreciate you sharing the failures too, not just the wins

34
Ccorsican821·4mo ago

been saying this for months and nobody listened

57
Oopusfan1.6k·4mo ago

can confirm, same results on our side

117
Eembedemma830·4mo ago

great in theory, messy in practice from what I have seen

4