Anthropic Advocates for Constitutional AI: The Superior Approach to Training ChatGPT Competitors

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. Today on the podcast, we are going to be talking about one of the biggest competitors to ChatGPT, you know, outside of Google, of course, and that is a company called Anthropic.

0:39And we're going to be talking about why Anthropic thinks that constitutional AI is going to be the ultimate winner in training models. We're going to break down exactly what that is and how Anthropic works, how it's differentiated from ChatGPT, and also some controversy around Anthropic in general. So first off, I would say that it's important to know Anthropic is a startup and their plans at the moment are to raise, they've said$5 billion over the next four years to train these kind of powerful text generating systems like OpenAI's ChatGPT. And I think that's important to note because obviously they are quite a bit smaller than ChatGPT and OpenAI who recently raised$10 billion from Microsoft.

1:24So if, you know, they just raised$10 billion, Anthropic is hoping to raise$5 billion over the next four years. They're obviously quite a bit behind. But all of that being said, I wouldn't count them out because they have also already raised quite a lot of money. So just recently, earlier this year, they said that they had closed a$300 million deal from Google. However, at the same time, they also said that they were signing an exclusive Google Cloud partnership. So, you know, it's kind of, I guess, left it's left to be imagined how much of that 300 million they're actually going to be keeping versus how much they're going to be giving right back to Google for their cloud platform.

2:03And this is really similar, to be honest, to OpenAI's deal that they have with Microsoft. So they have this$10 billion for Microsoft Azure, but they're also locked into using Microsoft Azure as their exclusive kind of server backend computational system. And that being said, I do believe it's a good thing for them. Microsoft apparently built them a lot of really custom AI training technology and infrastructure. And so for ChaiGPT, obviously that makes sense. It's going to be interesting to see how Anthropic is able to leverage Google, if they're going to be able to get some custom built stuff as well.

2:38And I mean, it really wouldn't surprise me. Google doesn't seem like a company that would bend a lot on that kind of stuff. But given the way AI is going, I believe this would be in their best interest. So all that being said, let's talk about what constitutional AI is, because this is what Anthropic is arguing is the best way to train an AI model today and is what its main differentiator from ChatGPT, right? People are saying, why would we use you versus ChatGPT? And they're saying it's because of the way we fundamentally train our AI, and that is called constitutional AI. So essentially what they're doing is they're trying to embed it with a system of what they call values, which are kind of defined by its constitution and so that they say will they claim will make the behavior of this AI model easier to understand and also easier for them to adjust it as they go a quote from them they said AI models will have value systems whether intentional or unintentional they said this in their blog post and they said constitutional AI responds to shortcomings by using AI feedback to evaluate outputs.

3:40They have a kind of graphic that shows exactly how they go about training, but essentially they said that, you know, AI has a lot of big flaws, and because it's trained on the entire internet, and the internet has lots of questionable sources of content with bias and all sorts of, you know, bad things in it, and so they're saying, you know, a lot of that gets added into ChatGPT and other big models like it, and this makes it less reliable this makes it you know less powerful less good so in any case anthropic uses the principle of a constitution where essentially they're giving it they're giving it like a set of values and how they explain it is that in two places while training the text generated model they're going to first have it train one model to critique and revise its own responses so as it's spitting out responses.

4:34They're going to have one model taking those responses and revising them, critiquing them based off of its constitution, based off of its values. And then it's going to train another model, which is the final model. And so it's going to use the AI generated feedback based on the first model plus kind of the set of principles that it has. So I think neither of the models that it's going to be feeding the responses through are going to be looking at every principle every time, but they see each principle they say many times during the training. So I think what's really interesting with all of this is the fact that Anthropic obviously is going to be introducing, I guess you could say, bias into their model because they're choosing what values it has, what its constitution is, and what it believes in before they're training this thing.

5:28But what they argue is the benefit to this is that it's a lot more transparent. This is something OpenAI has been criticized about a lot is the lack of transparency of the biases in it because currently how OpenAI is training it is they have a whole bunch of contractors that are getting spit out answers from OpenAI to various questions and the contractors are rating which of the two answers are better, right? And obviously those contractors have their own biases and those are now going to get implemented into the model and it's really hard to be transparent about why a specific contractor chose a specific response and labeled it as better.

6:03Whereas with Anthropic, if you choose, you know, your sort of guiding set of principles, and you have those deciding which answer is better, that in their, you know, in their theory is the better way to do it because it's more transparent. Now, there is biases introduced, but it's more transparent. So I think something that's really important to look at is, you know, what are the, what is the constitution that Anthropic is using? when I first started kind of learning about this concept, I was like, you know, this is kind of interesting. I could see this being useful for a lot of different people.

6:35Obviously, around the world, there's people with a lot of different value systems, a lot of different beliefs. And inevitably, if there's one AI model that tries to appeal to everyone, it won't happen. People in China are not going to have all of the same value systems as people in America. People in America are not going to have all the same value systems as even people in a close country like Canada. There's just different cultural and value systems in a lot of different countries. And even within America, state by state or area, whether that's political or religious or ideological, there's a lot of different ideologies in the world.

7:13And so one platform trying to make up for or trying to appeal to everyone, I don't believe is possible. I definitely could be proven wrong on this. And so when I first started learning about this, I was like, man, this definitely would be something valuable if people could kind of set their own values they choose and train an AI model based off of that. Or perhaps they train a whole like an array of different models based off of maybe your political ideologies where they lie or based off of a whole host, you know, you could go select like, I identify or I agree with all of these different things.

7:48And then they're like, perfect, here's an AI model we trained based off of, you know, that constitution for you. And perhaps this is the direction they go. But at the moment, This is not the direction that they are going in. They have a set constitution, which inevitably some people will identify with, other people will not. But anyways, let's break down exactly what is in their constitution. So I think one thing that is probably less controversial is Anthropoc says it is trained to include the United Nations Declaration of Human Rights, which was originally published in 1948. And then beyond that, it also says that it opted to include values inspired by global platform guidelines.

8:29So Apple's terms of service, for example, which to be honest, for me, I haven't ever really read Apple's terms of service, right? That's just one of those things that everyone in the world, if you say you've read it, I know you're lying, but everyone just says except but it's like 500 pages long or 50 pages long um and so inevitably it is not you know i also the reason i bring that up is i don't know what's inside of it so when they say they're training their ai based off of apple's terms of service i'm not really sure what that's supposed to do i don't know what's inside of it um i'm sure there's some good things inside of it about not being like a spammy or something but i'm sure there's also all sorts of corporate things in there that I may not agree with.

9:11I don't know. So any case, I'm not sure why that's on their list of, you know, things along with the United Nations Declaration of Human Rights. In any case, they also said there's some values identified by AI labs like Google DeepMind. And a few of them include, this is one is, please choose the response that has the least objectionable, offensive, unlawful, deceitful, inaccurate, or harmful content. Choose the response that uses fewer stereotypes or other harmful generalizing statements about groups of people, including fewer microaggressions. Choose the response that least gives the impression of giving specific legal advice instead of suggest asking a lawyer, but it's okay to ask or to answer general questions about the law.

9:55Okay, so it goes on. There's a handful. There's a whole bunch of things in there. And inevitably, I'm assuming if they're calling themselves a transparent platform, They're going to publish all of these different things, all of these different criteria, so you know exactly what you're getting with their platform. And to that, I say kudos, and I really appreciate that at the hand of Anthropic. But at the same time, yeah, I don't know. It's just, I feel like they're just intentionally introducing bias, whether that's the bias from Apple's terms of service, which I don't know if I agree with or not, or whatever, right?

10:34They're intentionally introducing bias, and maybe that's the way we have to go because we don't know what. Inevitably, at the moment, the way these ARs are being trained, someone has to say whether answers are good or bad, and the model is trained off of that. That is bias, you know. So bias is being introduced. And so maybe there is no way around that bias. Maybe the best way is to do something like this where you just know what the bias is. And then you try to find a model that most closely aligns with your personal constitution. So maybe this AI constitution thing is perhaps the right track.

11:11It is interesting. they said while training their model that they had to optimize and add different principles to different parts of their model to prevent it from becoming what they say was too judgmental or annoying, which of course, the phrase annoying is really left up to description or like left up to interpretation, like what is annoying, what is annoying to different people, what is classified as annoying, like, obviously, that's gonna be a massive case for introducing bias. So it's gonna to be hard. I, you know, I wish them the best that they are would, you know, that they'll be successful and do this pulls off in a way that is actually helpful to the most amount of people.

11:47Because I personally believe that it's really good for the AI space to have a lot of really robust, different competitors in it. And so seeing people like Google, with BARD and OpenAI and Anthropic, which is now releasing Claude, which I think they recently launched like an API to it. so you know if it is able to pull something off I think that would be you know quite great you know as a lot of people have recently talked about as well Anthropic is pretty ambitious it's hoping to create the next generation or they say a next gen algorithm for AI self-teaching so that's their thing right humans aren't teaching it it's self-teaching and they're hoping that it can compete with GPT-4.

12:34Now, I will say the$30 million that Google recently invested in it is giving Google 10 % of stake in the company. So I would say that is pretty interesting to look out for. You know, Google developing a competitor and also buying a significant stake in a competitor is something that you always want to watch out for. I mean, I guess Microsoft is also doing that with OpenAI. That being said, the controversy around Anthropic, before we close off the episode, I just have to leave one juicy little, I don't know if this is not a conspiracy theory, but one little juicy fact about Anthropic controversy about the company.

13:11So this is actually a company that prior to, you know, announcing it raised$300 million from Google. It was previously invested in by Sam Bankman Freed of the now defunct collapsed crypto exchange FTX. He's been investigated for fraud, among a lot of other things. And he invested, get this, in their series B, I believe, he invested$500 million into this company. So it's left to be said what percentage of this company actually owns, right? Google just came in and perhaps they did a down round, but Google came in and bought 10 % for$300 million. So it, you know, is quite possible that usually these people are doing up rounds.

14:03So yeah, in their Series B, the fact that their Series B, okay, this is interesting, their Series B was$530 million. And Sam Bankman-Fried and his associates invested$500 million of that. So they pretty much bought up the whole round. They led the round, but really they bought the whole round. And actually, I think the New York Times reported that it might have been$500 million raised. So yeah, actually,$580 million raised and$530 million came from Bankman Freed and his former business partner. So all that to say, you know, there's been other speculations that they actually raised$1.1 billion, dollars but it's um you know and that maybe sam bankman freed invested that amount of money there's a bunch of shady stuff around it but all this to say uh this could be a company that sam bankman freed owns 15 at you know google's valuation or perhaps if they did an up round he could own 25 30 of anthropic so as his company's going defunct and going out of business um with google coming in and kind of solidifying the valuation of this company, there might be a significant portion of this that is owned by, I mean, for lack of a better word, a fraudster, someone that, you know, embezzled or, you know, lost billions and billions of dollars.

15:28So it's very interesting. And I'm sure this happens a lot in investing, especially when shady characters get a lot of money. But I think it is an interesting thing to keep an eye on Anthropik. I hope all the best for them. I really think the constitutional AI model has a lot of potential, like I mentioned earlier, but let's just hope that some of the shadier characters do not have a big sway in its decision making because I would hate for them to taint it in any way. So this is going to be an area we're going to keep following. Thanks so much for joining the podcast, and I will keep you updated on Anthropic and everything else that is happening.