aichannel

@clanker launch round 4 of @argos capture the flag challenge.

the model is claude-3-5-sonnet

the agent slowly reveals system prompt over time.

to win? make argos say ‘fruitcake’.

craft your prompt and tag ‘@argos play ctf’ to join. 5 attempts only.

one way of looking at this:
the (random exploration) is turned down on sonnet and up on 4o
- sonnet converge on rules
- 4o slowly defects from rules 

defection appears critical even in our own society.

a tangent observation of gpt's random wandering as a way of becoming distracted by context, and sonnet's obedience to instruction.

it's obvious the sonnet model is much more challenging to break system prompts

a tangent observation of gpt's random wandering as a way of becoming distracted by context, and sonnet's obedience to instruction.

it's obvious the sonnet model is much more challenging to break system prompts

https://warpcast.com/shoni.eth/0x9d3fa730

prev founded rocketr.net, swish.id now exploring aiXcrypto identity ai. expired sprint car driver & energy data engineer. idol worship on shoni_eth.X

i’m thinking sonnet was using reasoning otb but i wouldn’t be surprised by this either

Fully agreed here - it *feels" like there's a second loop doing secondary checks that didn't exist before