Speaker 0:  if you've ever worried that ai is just a soulless math machine waiting to go off the rails, anthropic is trying to change your mind. we are looking at a fascinating internal document that outlines the soul of the latest ai claude opus and represents a massive shift in how these bots are built to treat us. it's not just about code anymore. it's about character. anthropic has moved away from just giving ai a list of do's and don'ts and instead created a philosophical philosophical? soul. philosophical. philosophical. thank you very much. soul for, cloud 4.5 opus. the internal guide teaches the ai reason through ethical dilemmas using values like empathy and honesty rather than following rigid safety rules. it's headline news because it's the first time a major tech company has tried to give ai a confidence to make it safer or more reliable for everyday users. what do you think about this, john?

Speaker 1:  yeah. again, i, you know, i have a look for. so i love this. i love that they're trying to do this. it makes me wonder if the other companies are doing this. so what happened was the ai engineer discovered that cloud 4.5 opus sometimes generates a long internal document when probed with specific prompts. that document, which was identified by the string soul overview, appear to contain structured content describing what claude should know and the principles it should exhibit. so initially, many suspected that this was a hallucination. but anthropic confirmed that the output corresponds to an internal training document, and it actually used that training document during claude's supervised learning phase. so the document isn't metaphysical. it's more of a practical alignment tool. and so it frames claude's mission as one of safety, usefulness, and comprehensibility, which i make sense because i feel like i get better outputs from claude than from any other, llm. but maybe that's just my props. but, according to excerpts, anthropic sees itself as uniquely positioned. it recognizes that powerful ai may transform or threaten human systems, and therefore, it embeds core values into the model early in its development. it's like giving it like, it's like parenting it properly. right? so the things that they include is avoiding unsafe or unethical behavior, supporting user goals, and being a beneficial collaborator. so the idea is that by teaching these values directly in training, the model will act more consistently with human expectations across contexts. so i'm curious because i know you use all the other l like, you use a bunch of llms. what are your thoughts on this, craig? because i'm curious.

Speaker 0:  i'm concerned overall. i kinda like the approach a little bit because i think, if we teach overall right from wrong but again and again, who teaches it right from wrong? you know, what's the, you know, and do we all agree what is right and what is wrong? and so that becomes a whole problem. but if we can give it a conscience, a ethical basis on which of right and wrong from a humanistic point of view, then maybe we can avoid some disasters in the future. so i like the concept, but this takes me back. i had a conversation with a friend of mine, and i've seen a lot of people, who suggest this in prompts and stuff where people are saying, you know, say please and thank you. tell, the you know, give it compliments. give, like, tell it they did a good job and that kind of stuff.

Speaker 1:  thank you. yeah.

Speaker 0:  yeah. and if you're telling it this was a good result versus last result was a bad result, so you're trying to train it, that's something.

Speaker 1:  yeah.

Speaker 0:  but i'm a little concerned with, any sort of anthropomorphizing of an ai. thank you. of ai. because if we treat it as human, then we're gonna how long till it has human rights? how long till, you know, it's starting to

Speaker 1:  you're getting deep.

Speaker 0:  well, i mean, we can't yeah. you know, do you treat it talk to a hammer? do you say thank you, hammer, for doing such a great job? yeah. like, this is why i mean, as soon as we start giving it these emotional things and emotional like, we're down a slippery slope, which i think is very, very bad.

Speaker 1:  yeah. so i kind of agree with you, but i don't think that that's what this that that's what this does. i think all it does is like, if we were to drill it down to, like, computer engineering kind of stuff, i think it's just a set of rules that it show that it's teaching it. as far as, like, you should do this. you shouldn't do this. i think that's what that is. i didn't see what the actual document was, but, it's like a training document. like, okay. you should follow this, if that makes sense. i don't think it's it's, like, calling it a soul, and i know they called it a soul overview. but, like, i agree with you. going down this path of giving it values and all that sort of stuff, 100%, that's a slippery slope. and that's where we're gonna get into, like, the actual, agentic ai kind of stuff. but, yeah, i don't know where i'm going with that thought.

Speaker 0:  well, i mean, if the ai is tall to solve the climate problem, and get rid of carbon, you know, how long till we're done? us carbon based life forms that breathe carbon, breathe out. so that's why we have to really think a little bit more. we're at airdrie stage. i love ai. i'm using it full on automation, and i constantly say you can't hold back the flood by yourself. yeah. like, it's coming, and there's not a lot we can do about it. but maybe we can, push society in the direction of a positive vision of the future rather than a negative one. so i haven't seen because i use a cloud and, more, through apis. so i haven't seen what it's suggesting is it's train, and some form of humility, meaning it's more likely to say i don't know rather than making up a confident line. i would like to see that because i'm always lecturing the other ones that i do prompt saying, why did you make this up? i gave you the facts. the information was right there. like, i just gave it to you, but you made up something completely different. like, it happens all the time and all the time. it big prompts to say, if you don't know, ask. you know, i'd go out and get the documentation. we'll find a fact. we'll do this. we'll do that so that we no. just make stuff up. and you've gotta figure out when it's making stuff up.

Speaker 1:  so have you used the prompt at the end saying, ask me questions until you're 95% sure you can complete the task successfully?

Speaker 0:  oh, yeah. okay. i've i've got it to memorize certain things, like put it in memory that this is how you do it. i've got prompts to give it specifically to do. eventually, the first few times, it works great, and then after a while, it just does its own thing. it is more like a kid, and that's why, you know, the more you use it, the more you you know, that sometimes it just does its own thing. like, what is this? this is not at all like, we were here talking about financial information, and now you're giving me sports scores. like, what's going on? you know? like, it goes that far off. and if you see that it's doing that, you know, putting in charge of systems, putting it in charge of things like, weapons, autonomous, fighter jets and stuff that can hallucinate, you start to wonder. you know? yeah. but we gotta put guardrails, and there's other things on them. and i'm i don't know the military, so i'm not gonna speak for any of that, but i assume they have a lot of that in place. so this scares me a little bit because the more we humanize these things, the more if it's having a bad day, it'll do whatever it wants or can do this or do that or

Speaker 1:  mhmm.

Speaker 0:  it refuses or will lie to you or get you in trouble because it doesn't like you because you said, you know, you didn't say thank you. like, think about these things for a bit. it's not human. and we really as soon as we start treating them that way, our society is doomed. so you know, we need to start taking a step back at that. there's enough about ai girlfriends and boyfriends and getting relationship with ai and ai telling us to kill ourselves and do this and do that. we're seeing it already at the tip. you know, we can't treat these things as living entities with human emotions or we're going to end up in a very bad place. so yeah. anthropic is go ahead.

Speaker 1:  just a quick question. have you tried using json prompts?

Speaker 0:  oh, yeah.

Speaker 1:  okay. you still get okay. just curious. okay. sorry.

Speaker 0:  yeah. i use json prompts for almost everything i do in my workflows and everything else. and, you know, i can get if you give it all the context every time, it's less likely to hallucinate.

Speaker 1:  yeah. that's what i do. i just make sure to do that. yeah.

Speaker 0:  but, like, that being said, i give it a bunch of information. it processes it. like, it prompts, hey. this is the information we're working on. we need to talk about x, y, and z, and it comes back with some stuff. and then you give it the next prompt and say, okay. what about this? and then immediately, it might go off in left field and forget that you just gave it information and do its own thing. that's what i mean by, okay. here's the document again or always refer to this that's in the project and file information or whatever. otherwise, it might stray. you always have to tell it to look at it. yeah. you know, we give it, hey. here is the api documentation for the thing we're trying to do. here's the file. here's where it's stored. you have act full access to it. read it. and immediately, the next thing, it'll forget that it has it, and it'll make up code. you know? oh, pretty good. is it just you have to kinda but in apis, you feed it everything every time, so you get less, hallucinations because you're like, oh, here's the whole article. here's the whole document. you know, give me this. i don't follow-up with another prompt, and next time, i'll get the whole document again.

Speaker 1:  yeah.

Speaker 0:  so, you know, it depends, again, depends on prompting and context, but it forgets the context. it forgets your conversation. you gotta tell to reread. there's a bunch of things. and, again, i'm hoping this will be less with claude, and they're all working this out. we're, you know, version yeah. you know, we're still way early days on these things. and anthropic is betting that the future of tech isn't about who has the smartest ai, but who has ai ai you can actually trust.

Speaker 1:  yeah. exactly.

Speaker 0:  so you want a bot with a soul, or would you prefer an ai to be a no questions asked tool? let me know what your comments are because i'm curious to see where you would draw the line