Anthropic's Swift Demise: ChatGPT's Last Stand

0:00The chat GPT competitor AI tool Anthropic has just announced something massive today we're going to break down what it means for the industry of AI and what some of the big piece of AI they announced today is that they are increasing their they're pretty much expanding their context windows from about 40 ,000 words or characters to 100 ,000 characters so the reason or I guess tokens so the reason that this is important I guess I'll cover the tokens thing real quick so a token is essentially a piece of a word like a long word might be multiple tokens something like machine might be two tokens and it's essentially just the way that these language models are predicting the they're not really predicting the next word that comes they're predicting the next like few letters or chunks of words that's just how their models work so it's um i guess what that kind of corresponds to essentially is um now that they're doing a hundred thousand tokens that corresponds to around 75 000 words but anyways the reason why this is massively important and is making some really big changes is because now this means that anthropic um can essentially analyze not just, you know, like a chunk of text, you get this with a chat GPT, where like I've implemented it into a bunch of different software tools now.

1:17And if the input or the output is too big, chat GPT is like, sorry, like you just ran out of space, pretty much. And the problem is with a lot of these requests, sometimes your request and its response, so what you say to it and what it responds is all combined. And so if you have a really long request and someone and it's supposed to generate a really long response it just caps out it can't do it so what's really interesting with this is the fact that um chat gpt and gpt4 specifically announced that they're going to expand to 40 000 tokens and that was like oh my gosh it's huge that's amazing um this is still in beta it's you know you gotta get on wait list to get it no one really has access to this but um you know and so that like that's one thing but this is literally more than double of what this waitlisted feature is.

2:03This is insanity. I'm really literally this means that businesses now can submit like hundreds of pages of materials for Claude. And it can look through analyze. And the conversations you have the Claude can go on for hours or even days. And it doesn't miss a beat. It's able to use all of the context and all this data, which is like absolutely phenomenal. And so when I first heard about this, I was like, Oh, awesome, like it can write longer things, I guess. I think honestly what this is going to be used for more than like creating a 75 ,000 word book is going to be digesting content and analyzing things.

2:39And I think people are not thinking enough about what the implications of this are. So I'm going to go over some of those implications because I think they're absolutely massive. First off, I think it's important to note that like, so your average person could probably read about 75 ,000 words of text in five hours or more. And so if you had someone that took five hours to read that text, obviously you're going to have to digest it, you're going to have to remember it, you're going to have to analyze it. There's a lot of stuff that goes into reading 75 ,000 words of text. What's amazing is Claude can do all of that in less than a minute.

3:12So it's time to digest and all that takes less than a minute. An example that the Anthropic team gave in their demo is that they essentially loaded up the entire text of the great Gatsby into Claude. And, and I think that equates to around 72 ,000 tokens. And they modified one line somewhere in the text to say, you know, a software engineer that works on machine learning, tooling, and anthropic. And then they put the entire book in there pretty much and asked it to spot, you know, what was different between this document and that document and it responded with the correct answer in 22 seconds reading 72 ,000 tokens between the two doc and then comparing that and finding the issue so honestly this is insanity but the implications are not just finding random easter eggs and books obviously I think the implications are going to be number one finance this is massive they give another demo where they essentially were able to upload Netflix's 10k document.

4:16And they said from the consolidated balance sheet, please highlight the most important items up to a potential investor and explain their significance. First, make the markdown formatted table or tables with the selected results and then provide a summary and an analyst of the results. And it did an amazing job in a very small amount of time. They uploaded Netflix's 10k document, which is, I believe, like 40 pages long. And from that, it created like a summary. It was like Netflix has a strong cash position with $5.2 billion in cash. Content assets are, you know, this many billion, blah, blah, blah.

4:51Like it gave a summary of what was going on, what an investor needed to know. It had, it created cells of data around, you know, some of the changes in their finances and some things that were relevant. really, really amazing. And so I think it's pretty cool because you can drop multiple documents or even a book into a prompt, like multiple documents. That's insane. First off dropping documents and I can't even do a chat GPT. Um, so that's going to be pretty cool. Um, you can ask Claude different questions that essentially require the synthesis of knowledge, the synthesis of knowledge. So across different parts of text.

5:27So like, this is really, really amazing. If you think about it, because it isn't just two questions and you're saying like, you know, like write an article about X, Y, and Z and this tone and this thing. And it's like, oh man, it's got like three different concepts it's got to work with. Like you're, you're asking those questions for like, for like full books, like a hundred page document. So I think this is pretty amazing. Um, it's great following your instructions in the demo. I haven't actually been able to get my hands on that. So I will have to, you know, spare or hold back judgment until I can actually get that.

6:00But they're saying it does better than a human assistant would. And like, honestly, a financial analyst, you got to learn how to use these tools to really become more effective at your job and to really put yourself out there because and I think, OK, here's the thing, you know, people are like, oh, man, it's going to kill financial analysts. Not necessarily, definitely not because, you know, me as someone that is not a professional financial analyst, I don't always know the questions to ask. And I think this is true of every industry where you may not know the questions to ask and the prompt to use.

6:33Now, that being said, there definitely are tools that are there and people that are out there helping share the right prompts and right questions to ask. And so it's kind of like saying, you know, X, Y, and Z, like a librarian's job isn't going to get replaced because not everyone knows the right questions to ask on Google. Like eventually we kind of figure out a lot of these questions, but that being said, I think there's still, um, you know, if you were a financial analyst and you're worried about your job being displaced because now in a couple seconds, these tools can like analyze multiple things.

7:00I think there's still a strong use case, uh, for your job position. As long as you're able to just do way more analyze way more stuff, you know, the questions to ask, you know, the right pieces to put together. Um, and you make a really customizable, um, You know response but that being said like definitely this is going to be Disrupting a lot in the finance industry, but a ton of other industries which I want to go over so The another demo that they gave was for one of their partners, I guess called assembly AI so they said that a hundred characters of Or like tokens essentially translates into around six hours of audio you know if you listen to an audiobook or something and so they did a demo with assembly AI where essentially they transcribed a really long podcast into about 58 ,000 tokens and then they used Claude to summarize and do questions and answers about that podcast.

7:55Here are a few other use cases I'm thinking that are going to be quite revolutionary. The first one is obviously just to be able to digest and summarize and explain really dense documents, financial statements, of research papers, like we were talking about before. Another is to its ability to analyze strategic risks and opportunities for a company based on its annual reports. So again, this kind of goes into the whole financial analyst piece, but being able to take an annual report and do a lot of robust calculations on that. It's also going to be able to assess the pros and cons of a piece of legislation, which I think is really actually going to be interesting to see how this impacts, right?

8:33I'm not sure how this works in every country in the world. But here in America, on both sides of the aisle, a very common strategy is that they're they say, hey, we're going to vote on a piece of legislation, they drop the piece of legislation, or the bill, which is going to be like 1000 page document, you know, like a day before and they say we're gonna vote on this tomorrow. And really, what they're trying to do is they tried it's usually with like budgets and stuff like that, where they try to cram a whole bunch of little special interest things in from all the different senators. The senator is like, hey, I'll vote and, you know, agree with the bill if you do these special interests for people in my, like, state or jurisdiction.

9:09And they, like, a lot of people vote on the thing so that just, like, to get it passed and, you know, because they're like, oh, we don't want, like, the government to shut down. We got to pass our budgets, blah, blah, blah. And, you know, in fact, I believe that there was a quote I heard somewhere. So you go validate this. But someone said, like, we have to pass this bill to find out what's inside of it. And it's like, anyways, it's kind of just like a meme in America where you got to pass a bill to find out what's inside of it because there are huge, dense documents and they get dropped. You know, everyone contributes like a whole bunch of things they want to see in it.

9:42And the whole thing gets dropped like a couple of days or a day before the vote. And there's no way people are going to be able to analyze the entire thing and make like a very informed decision. And so consequently, what they do now is they hire like a team of 100 people and they all take a chunk and they have to like read it and summarize and try to figure out like the loopholes that are being added from it. And people come out with reports based off of stuff like that. Anyways, it seems ridiculous. And I think that something like Claude is going to really shift this where, you know, a bill might be proposed.

10:12And if they made it public, the American public could literally take it, stick it in, ask any questions they had about it, find any loopholes, find all the places that money is getting spent or things that are happening. And, you know, if they don't make it available to the general public, at least senators and Congress, people will be able to do this. but I think it's going to be really powerful to upload that massive document and really assess the pros and cons of that piece of legislation. So I think that's going to be a really big change. Another one is just the ability to identify risks and themes and different forms of arguments across legal documents.

10:45So this is going to be massive for law. I think this is going to be really big just to be able to read through hundreds of pages of developer documentation and surface answers to technical questions. They did a demo where they said they uploaded, first off, they said, hey, what's LangChain? And it said, I don't know what LangChain is. And they said, okay, here is LangChain's API documentation. It's a PDF. They uploaded it as a long PDF. They said, based off this documentation, create a simple demo of LangChain that uses Anthropics language model. Think step by step. And it went and it created that documentation, or it created that simple demo of Langchain right inside of Anthropic had all the code.

11:24So I think this can be really powerful for code, right? You're going to be able to rapidly prototype by dropping an entire code base into the context, and you'll be able to intelligently build on or modify that. So really big implications for developers, for legal, for governments, for financial analysts, a lot of very big industries that previously felt like they were a little untouchable they had a little bit of a mo they're a little further back I feel like are going to have some massive disruption because here's the thing if Claude is doing a hundred thousand tokens Chad Gbt is going to have to do a hundred thousand tokens and if this becomes mainstream and eventually it's just like okay you get unlimited tokens use whatever you want and just pay for it it's going to be insane people will be able to upload insane amounts of data insane amounts of content and just get really robust analysis on this content.

12:19So I think this is actually a massive and very major game changer in the AI space.