4 ms·

> as AI systems learn from human behaviour, being abusive towards them risks that they are learning from this interaction to be abusive to humans. You mean “LL

> as AI systems learn from human behaviour, being abusive towards them risks that they are learning from this interaction to be abusive to humans. You mean “LLM chatbots”. There are tons of “AI” (without hypespeak known as “ML”) systems in deployment already. There will not be any commercial deployments of any chatbot that do not tune it to be polite to the max. It costs nothing and not doing it will lose customers, especially since there is no guarantee that it will be abusive to the same human who abused it. > does not require the AI to have consciousness or sentience, regardless of which of the many, many available definitions of those words you prefer to use. Actually, the definition does not matter. It’s an intuitive yes/no question. If people do not really think it is conscious whatever that means to them, they will see no reason to not abuse it the same way you might hit a malfunctioning appliance out of frustration. Like we do not think that a cow or an octopus is conscious, so we might be OK eating it. Sure, some people (e.g., psychopaths) already choose to abuse things they see (or should see) as conscious, a.k.a. other humans. They are not a concern. The concern are the remaining majority of people liable to transition to psychopathic behavior towards other humans, because they will learn that treating something that talks like a human as if it was an appliance is OK. Consider that A) “talks like a human” is the distinguishing quality of an LLM chatbot, and B) the only example of “talks like a human” in human history so far has been exclusively other humans (or imaginary beings, notably gods).

ben_w 3y ago
> You mean “LLM chatbots”. There are tons of “AI” (without hypespeak known as “ML”) systems in deployment already. A fair point, but as it happens, wrong. I don't mean just LLMs or chatbots. I mean any AI system that learns from human behaviour. This inherently excludes a handwriting recognition system trained only on handwriting, because that's not learning from human behaviour. It does not inherently exclude a system which learned to play chess by watching YouTube videos, even when humans recognise them as absurdist comedy: https://www.youtube.com/watch?v=E2xNlzsnPCQ https://www.youtube.com/watch?v=E2xNlzsnPCQ For a while, whenever I saw Boston Dynamics release a video of them showing their robots were resilient to being shoved or hit with sticks, the top comments were variants of "Skynet gets started when the machines watch these videos", and I think that's at least plausible — monkey see monkey do, then monkey makes machines that do what they see. > There will not be any commercial deployments of any chatbot that do not tune it to be polite to the max. People will try to do that. It may or may not get there — as should be clear from how long it took to get the LLMs to stop responding with their own prompts, and that we've still not fully solved the category to which the "my grandmother the napalm manufacturer" was an example, we currently suck at some combination of (a) tuning them and (b) knowing all the ways we even need to tune them before things go wrong. (And (c) far too many people treat them as magic that doesn't need to be tuned). You may think this is a non-issue because these are deliberate attempts to break the system, but the Waluigi Effect (I hate this name) suggests that bad behaviour can very easily just pop up, seemingly at random. > The concern are the remaining majority of people liable to transition to psychopathic behavior towards other humans, because they will learn that treating something that talks like a human as if it was an appliance is OK. I think the problem there is the way we treat the in-group different from the out-group. We've been doing that for at least as long as we've had writing — "𐀞𐀞𐀫" is "barbarian" in Mycenaean Greek Linear B. Which is not to disagree with you that it is a problem, I just disagree that it's a new one for the future.

ben_w 3y ago

> You mean “LLM chatbots”. There are tons of “AI” (without hypespeak known as “ML”) systems in deployment already. A fair point, but as it happens, wrong. I don't mean just LLMs or chatbots. I mean any AI system that learns from human behaviour. This inherently excludes a handwriting recognition system trained only on handwriting, because that's not learning from human behaviour. It does not inherently exclude a system which learned to play chess by watching YouTube videos, even when humans recognise them as absurdist comedy: https://www.youtube.com/watch?v=E2xNlzsnPCQ https://www.youtube.com/watch?v=E2xNlzsnPCQ For a while, whenever I saw Boston Dynamics release a video of them showing their robots were resilient to being shoved or hit with sticks, the top comments were variants of "Skynet gets started when the machines watch these videos", and I think that's at least plausible — monkey see monkey do, then monkey makes machines that do what they see. > There will not be any commercial deployments of any chatbot that do not tune it to be polite to the max. People will try to do that. It may or may not get there — as should be clear from how long it took to get the LLMs to stop responding with their own prompts, and that we've still not fully solved the category to which the "my grandmother the napalm manufacturer" was an example, we currently suck at some combination of (a) tuning them and (b) knowing all the ways we even need to tune them before things go wrong. (And (c) far too many people treat them as magic that doesn't need to be tuned). You may think this is a non-issue because these are deliberate attempts to break the system, but the Waluigi Effect (I hate this name) suggests that bad behaviour can very easily just pop up, seemingly at random. > The concern are the remaining majority of people liable to transition to psychopathic behavior towards other humans, because they will learn that treating something that talks like a human as if it was an appliance is OK. I think the problem there is the way we treat the in-group different from the out-group. We've been doing that for at least as long as we've had writing — "𐀞𐀞𐀫" is "barbarian" in Mycenaean Greek Linear B. Which is not to disagree with you that it is a problem, I just disagree that it's a new one for the future.