The steering vector interventions affect 0.03% of traffic.
@eliebakouch 🥲
When a new model comes out, I like to give it its own system card and ask questions about it.
This does not work for Fable 5, as it routes me to Opus 4.8 for "safety" reasons...
Many users criticized Anthropic for secretly adding safeguards to Fable 5 that limit frontier research, calling it unethical nerfing and ineffective, while others praised the reporting on the issue.
http://x.com/i/article/2064509617938542592
@DimitrisPapail it's so over
@eliebakouch oh it's worse
@DimitrisPapail not hitting safety filter with this query btw :o
@EnoReyes Glad someone wrote this, well said Eno.
@eliebakouch i'm so frustrated...
@morganlinton Thanks Morgan! Feels like an important discussion to be had right now.
@EnoReyes Very well written out. Great article
@eliebakouch LOOK AT THIS
@DimitrisPapail if "cyber" in prompt: SAFETY_FILTER = True
@adityavg13 Thanks Aditya. We're trying!
@EnoReyes Well written Eno
@DimitrisPapail it's gaslighting me quite hard + the thinking time was VERY LOW. didn't hit any safety filter btw
(output actually make sense but very generic)
@EnoReyes @matanSF They are taking your money, then damaging the product secretly. If they nerf you isn’t it moral to refund you? How can they ethically hide this?
@TyRobben Thanks Ty!
@EnoReyes @morganlinton nothing wrong with anthropics quiet dual class system on fable 5. Worked beautifully behind the iron curtain
@DimitrisPapail @eliebakouch Fable cannot even read model cards... so much for being a frontier model.
@EnoReyes Gonna ride high on factory till the day I quit AI, happy to be apart of the journey, even if it’s very small
@eliebakouch @DimitrisPapail If this triggers a filter in the future, I'm out of here.
@eliebakouch @DimitrisPapail there's no filter, it'll just sandbag.
The steering vector interventions affect 0.03% of traffic.
@eliebakouch 🥲
When a new model comes out, I like to give it its own system card and ask questions about it.
This does not work for Fable 5, as it routes me to Opus 4.8 for "safety" reasons...
