I'm not sure I agree, mostly just because of the implication that Claude (or any AI) has anything approaching the ability to act as a sort of "spokesperson". It gives the energy of "straight from...
I'm not sure I agree, mostly just because of the implication that Claude (or any AI) has anything approaching the ability to act as a sort of "spokesperson".
It gives the energy of "straight from the horse's mouth", which feels like it should make the comments trustworthy or at least somehow meaningful, but they're not. Much like a parrot that is taught to repeat the phrase "death to America" doesn't actually harbour any particular ideology. Nor does the fact that it can repeat those sounds in that order say anything more than if it were repeating any other sounds.
I don't think the idea is it's a spokesperson for Anthropic. This is mostly just a speech from Bernie. Having Claude respond in turn is just a tongue-in-cheek way to have a conversation partner...
I don't think the idea is it's a spokesperson for Anthropic. This is mostly just a speech from Bernie. Having Claude respond in turn is just a tongue-in-cheek way to have a conversation partner created by the very people you want to regulate.
This is an interesting piece, and I love Bernie for doing things to bring attention to AI's many faceted problems. However there are a few things I don't like about it (as a format, not with any...
This is an interesting piece, and I love Bernie for doing things to bring attention to AI's many faceted problems. However there are a few things I don't like about it (as a format, not with any of the content per se).
Firstly - even with no malice or ill intent on anyone's part, AI in default mode is going to agree with the person they're talking too. So that isn't really "Claude's opinion", Claude isn't a person with specific beliefs.
Secondly - even though we see the text on screen, drawing attention to the fact that this is a conversation happened as it is being shown, it is trivially easy to have put in a setup prompt at the very beginning to make sure the conversation goes exactly as they want it to.
This is incidentally why they are incredibly dangerous. A suicidal kid will get told how to make a noose. A schizophrenic will get told his delusions are reality. Trying to make these LLMs 'human...
default mode is going to agree with the person they're talking too.
This is incidentally why they are incredibly dangerous.
A suicidal kid will get told how to make a noose.
A schizophrenic will get told his delusions are reality.
Trying to make these LLMs 'human like' is what causes them to be used like therapists, support networks, and mentors. Things that are terrible to be done by yes-men.
yeah. That's definitely high up in the reasons I don't trust LLM's. A yes man that people trust as a figure of authority has so many bad implications. That said, these models should be able to...
yeah. That's definitely high up in the reasons I don't trust LLM's. A yes man that people trust as a figure of authority has so many bad implications.
That said, these models should be able to refuse to answer prompts relating to the above topics. Of course, that only came about after litigation ramped up from real world events. It's so disgusting how tech treats these as costs to bear and not mission critical scenarios to prevent from the outset.
Right. When it hit Bernie with "you are absolutely right" I cringed. Because you cannot treat anything these little syccophants say seriously without consulting more authoritative sources. But, at...
Right. When it hit Bernie with "you are absolutely right" I cringed. Because you cannot treat anything these little syccophants say seriously without consulting more authoritative sources.
But, at the same time, I was like: Good luck explaining why exactly is your machine saying "you are absolutely right" to Bernie, but it's just a fluke, because that's not Anthropic's standpoint at all.
gpt-4o was incredibly bad for this and it was a massive scandal (rightfully so, arguably not enough happened because of it, though i guess legal cases will take a while to percolate), but current...
A suicidal kid will get told how to make a noose.
A schizophrenic will get told his delusions are reality.
gpt-4o was incredibly bad for this and it was a massive scandal (rightfully so, arguably not enough happened because of it, though i guess legal cases will take a while to percolate), but current frontier models are much much better at appropriate refusals and better but not perfect at avoiding sycophancy
I repeated the conversation to a new Claude account, it gave very similar responses, perhaps a little more verbose, so I think the only instruction they might have given was to be succinct.
I repeated the conversation to a new Claude account, it gave very similar responses, perhaps a little more verbose, so I think the only instruction they might have given was to be succinct.
I appreciate that, I wanted to do the same but couldn't this morning. It's gotten me thinking though, for the wider audience, how do you convince them that such a conversation is "real"? When the...
I appreciate that, I wanted to do the same but couldn't this morning.
It's gotten me thinking though, for the wider audience, how do you convince them that such a conversation is "real"? When the AI in the video itself is telling you about use cases where AI can be used to spread disinformation and propaganda, this seems particularly difficult.
I recognize that from a technical perspective, everything you are saying is completely true. Understanding the shortcomings of LLMs does make this video feel strange as a political message. I...
I recognize that from a technical perspective, everything you are saying is completely true. Understanding the shortcomings of LLMs does make this video feel strange as a political message.
I think it’s a poignant commentary on the larger social situation, though. Most AI endorsers, including the companies building them, are not advertising “you can prompt this tool to say anything” or “this is a glorified yes-man”. They are saying “this will replace the vast majority of skilled white-collar workers”.
That makes this “interview” an effective catch-22. Either AI proponents must hypocritically critique the real shortcomings of their own products, or contend with their “oracle of truth” making a robust argument against the big tech business model.
Bernie Sanders session with Claude on privacy. Somewhere between hilarious and Black Mirror.
At first I thought this was going to be weird or cringe, but it’s a well thought out idea. Great editing and messaging.
I'm not sure I agree, mostly just because of the implication that Claude (or any AI) has anything approaching the ability to act as a sort of "spokesperson".
It gives the energy of "straight from the horse's mouth", which feels like it should make the comments trustworthy or at least somehow meaningful, but they're not. Much like a parrot that is taught to repeat the phrase "death to America" doesn't actually harbour any particular ideology. Nor does the fact that it can repeat those sounds in that order say anything more than if it were repeating any other sounds.
I don't think the idea is it's a spokesperson for Anthropic. This is mostly just a speech from Bernie. Having Claude respond in turn is just a tongue-in-cheek way to have a conversation partner created by the very people you want to regulate.
This is an interesting piece, and I love Bernie for doing things to bring attention to AI's many faceted problems. However there are a few things I don't like about it (as a format, not with any of the content per se).
Firstly - even with no malice or ill intent on anyone's part, AI in default mode is going to agree with the person they're talking too. So that isn't really "Claude's opinion", Claude isn't a person with specific beliefs.
Secondly - even though we see the text on screen, drawing attention to the fact that this is a conversation happened as it is being shown, it is trivially easy to have put in a setup prompt at the very beginning to make sure the conversation goes exactly as they want it to.
This is incidentally why they are incredibly dangerous.
A suicidal kid will get told how to make a noose.
A schizophrenic will get told his delusions are reality.
Trying to make these LLMs 'human like' is what causes them to be used like therapists, support networks, and mentors. Things that are terrible to be done by yes-men.
Precisely! That's demonstrating the problem, not a caveat to the exercise.
yeah. That's definitely high up in the reasons I don't trust LLM's. A yes man that people trust as a figure of authority has so many bad implications.
That said, these models should be able to refuse to answer prompts relating to the above topics. Of course, that only came about after litigation ramped up from real world events. It's so disgusting how tech treats these as costs to bear and not mission critical scenarios to prevent from the outset.
Right. When it hit Bernie with "you are absolutely right" I cringed. Because you cannot treat anything these little syccophants say seriously without consulting more authoritative sources.
But, at the same time, I was like: Good luck explaining why exactly is your machine saying "you are absolutely right" to Bernie, but it's just a fluke, because that's not Anthropic's standpoint at all.
gpt-4o was incredibly bad for this and it was a massive scandal (rightfully so, arguably not enough happened because of it, though i guess legal cases will take a while to percolate), but current frontier models are much much better at appropriate refusals and better but not perfect at avoiding sycophancy
(there are benchmarks in model report cards if you're curious about numbers)
I repeated the conversation to a new Claude account, it gave very similar responses, perhaps a little more verbose, so I think the only instruction they might have given was to be succinct.
I appreciate that, I wanted to do the same but couldn't this morning.
It's gotten me thinking though, for the wider audience, how do you convince them that such a conversation is "real"? When the AI in the video itself is telling you about use cases where AI can be used to spread disinformation and propaganda, this seems particularly difficult.
I recognize that from a technical perspective, everything you are saying is completely true. Understanding the shortcomings of LLMs does make this video feel strange as a political message.
I think it’s a poignant commentary on the larger social situation, though. Most AI endorsers, including the companies building them, are not advertising “you can prompt this tool to say anything” or “this is a glorified yes-man”. They are saying “this will replace the vast majority of skilled white-collar workers”.
That makes this “interview” an effective catch-22. Either AI proponents must hypocritically critique the real shortcomings of their own products, or contend with their “oracle of truth” making a robust argument against the big tech business model.
I didn’t get to watch all of it. But I was thinking that Claude would know everything about Bernie, and be sure you cater its responses to him.