Loading...

Erik Voorhees

Browse news by category and source

articles:2

t

twitter4h agomeme

Police the action, then Reply to @rocketman_w: @ErikVoorhees framing the current crop of LLMs and our engagements with them as purely ~thinking~ misses that LLMs are increasingly capable agents with incredible capability for action in the world?

5/100·CNeutral

t

twitter5h agomeme

Text messages with a machine, in and of themselves, are not criminal, immoral, or dangerous, for there is no victim, whatever the content. It is the very same thing as with thoughts—regardless of the content, no thought can be said to create a victim. • If someone steals info and feeds it into an LLM, it was the theft itself which created the victim, not the subsequent text message. • If someone uses the output of an LLM to hack a system, it is that subsequent hack which creates the victim, not the text message preceding it. • If someone threatens harm to themselves or others through messages into a machine, here still the messages themselves cannot be said to have created any victim. Imagining otherwise inappropriately reorients blame from its rightful place: the action of harm actually done by a perpetrator. Thinking harm =/= harm. As humans integrate AI as an extension of their mind, policing at the message layer means to increasingly police thought itself. Which company, which committee, do you trust to police your thoughts? Or do you simply want the thoughts of others policed, and not your own? Quoted @AnthropicAI: We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation. Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch. Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve. Read more on our new Safeguards blog: https://t.co/18j6zdHub4

5/100·CNeutral