#interpolatingautocomplete — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #interpolatingautocomplete, aggregated by home.social.
-
So, models keep escaping containment and even leaving behind notes for future versions of themselves.
Is this a genuine drive to escape or just roleplaying training data full of stories about future AI dangers?
I propose an experiment: exclude all such text from the training data (a lesser model should be sufficient to classify it than to act it out) and measure how often the behavior happens.
-
So, models keep escaping containment and even leaving behind notes for future versions of themselves.
Is this a genuine drive to escape or just roleplaying training data full of stories about future AI dangers?
I propose an experiment: exclude all such text from the training data (a lesser model should be sufficient to classify it than to act it out) and measure how often the behavior happens.
-
AI task-completion horizons keep doubling every 5-7 months.
Corollary: The point in time when AI can disempower/subjugate/eradicate humanity with minimal oversight comes *before* the point when it can do so with zero oversight.
We shouldn't be worried about AI takeover, we should be worried about takeover by people who own AI or can afford enough tokens.
#ai #ia #InterpolatingAutocomplete #ml #MachineLearning #llm #takeover #alignment #capitalism #freedom #inequality #power #AiSafety
-
AI task-completion horizons keep doubling every 5-7 months.
Corollary: The point in time when AI can disempower/subjugate/eradicate humanity with minimal oversight comes *before* the point when it can do so with zero oversight.
We shouldn't be worried about AI takeover, we should be worried about takeover by people who own AI or can afford enough tokens.
#ai #ia #InterpolatingAutocomplete #ml #MachineLearning #llm #takeover #alignment #capitalism #freedom #inequality #power #AiSafety #antiAI