Anthropic silently limits LLM capabilities for 0.03% of traffic.
Can you imagine the safety disaster if i speed up my input pipeline 2x???
Joke's on them my input pipeline is already prefect.
When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT.
Anthropic estimated that this would affect approximately 0.03% of traffic.
Users criticized Anthropic for imposing silent capability limits on frontier LLMs without notification because the changes hinder legitimate infrastructure work.
@giffmana don’t flex that hard pls:(
Can you imagine the safety disaster if i speed up my input pipeline 2x???
Joke's on them my input pipeline is already prefect.
@xeophon I actually love our new codebase so much, it's my magnum opus
@arcjax7 Homegrown. tfdata and tfrecord are the devil in prada
@arcjax7 You may not
@giffmana What’s your input pipeline? I’m using http://tf.data but tf records are annoying to use
@giffmana May I assume it’s webdataset based?
@giffmana Same
@giffmana silent capability limits are rough for legitimate infra work
@giffmana prefect" — this guy pipelining.
@giffmana Is it based upon spdl: https://github.com/facebookresearch/spdl
@giffmana The illusion of a perfect pipeline
Speed just exposes flaws faster. Context retention is where the real work happens.
Anthropic silently limits LLM capabilities for 0.03% of traffic.
Can you imagine the safety disaster if i speed up my input pipeline 2x???
Joke's on them my input pipeline is already prefect.
When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT.
Anthropic estimated that this would affect approximately 0.03% of traffic.