How I use LLMs

Hi everyone. So, in this video, I'd like to continue our general audience series on large language models like ChatGPT.

In the previous video, a deep dive into LLMs that you can find on my YouTube channel, we went into a lot of the under-the-hood fundamentals of how these models are trained and how you should think about their cognition or psychology. Now, in this video, I want to go into more practical applications of these tools. I want to show you lots of examples. I want to take you through all the different settings that are available, and I want to show you how I use these tools and how you can also use them in your own life and work. So let's dive in.

Okay, so first of all, the webpage that I have pulled up here is chat.openai.com. Now, as you might know, ChatGPT was developed by OpenAI and deployed in 2022. So this was the first time that people could actually just kind of like talk to a large language model through a text interface, and this went viral all over the place on the internet, and this was huge. Now, since then, though, the ecosystem has grown a lot. So I'm going to be showing you a lot of examples of ChatGPT specifically, but now in 2025, there are many other apps that are kind of like ChatGPT-like, and this is now a much bigger and richer ecosystem. So, in particular, I think ChatGPT by OpenAI is this original gangster incumbent; it's most popular and most feature-rich, also because it's been around the longest, but there are many other kind of clones available. I would say I don't think it's too unfair to say, but in some cases, there are kind of like unique experiences that are not found in ChatGPT, and we're going to see examples of those. For example, big tech has followed with a lot of kind of ChatGPT-like experiences. So, for example, Gemini, from Google; Meta; and Copilot, from Microsoft, respectively. And there's also a number of startups. For example, Anthropic has Claude, which is kind of like a ChatGPT equivalent; xAI, which is Elon's company, has Grok; and there are many others. So all of these here are from United States companies, basically. DeepSeek is a Chinese company, and LLaMA is a French company, Mistral.

Now, where can you find these and how can you keep track of them? Well, number one, on the internet somewhere, but there are some leaderboards. In the previous video, I showed you; Chatbot Arena is one of them. So here you can come to some ranking of different models, and you can see sort of their strength or ELO score. And so this is one place where you can keep track of them. I would say like another place maybe is this Scale leaderboard from Scale AI. So here you can also see different kinds of eval and different kinds of models and how well they rank, and you can also come here to see which models are currently performing the best on a wide variety of tasks. So understand that the ecosystem is fairly rich, but for now, I'm going to start with OpenAI because it is the incumbent and is most feature-rich, but I'm going to show you others over time as well. So let's start with ChatGPT. What is this textbox, textbox, and what do we put in here?

Okay, so the most basic form of interaction with the language model is that we give it text, and then we get some type text back in response. So, as an example, we can ask to get a haiku about what it's like to be a large language model. So this is a good kind of example ask for a language model because these models are really good at writing; so writing haikus or poems or cover letters or resumés or email replies, they're just good at writing. So when we ask for something like this, what happens looks as follows: the model basically responds: "Words flow like a stream, endless echo, never mind. Ghost of thought unseen." Okay, it's pretty dramatic, but what we're seeing here in ChatGPT is something that looks a bit like a conversation that you would have with a friend; these are kind of like chat bubbles. Now, we saw in the previous video is that what's going on under the hood here is that this is what we call a user query, this piece of text, and this piece of text, and also the response from the model, this piece of text, is chopped up into little text chunks that we call tokens. So these, this sequence of text, is under the hood a token sequence, one-dimensional token sequence. Now, the way we can see those tokens is we can use an app like, for example, Tiktoken. So making sure that GPT-4 is selected, I can paste my text here, and this is actually what the model sees under the hood. My piece of text to the model looks like a sequence of exactly 15 tokens, and these are the little text chunks that the model sees. Now there's a vocabulary here of 200,000 roughly of possible tokens, and then these are the token IDs corresponding to all these little text chunks that are part of my query, and you can play with this and update, and you can see that, for example, this is case-sensitive; you would get different tokens, and you can kind of edit it and see live how the token sequence changes. So our query was 15 tokens, and then the model response is right here, and it responded back to us with a sequence of exactly 19 tokens. So that haiku is this sequence of 19 tokens. Now, so we said 15 tokens, and it said 19 tokens back. Now, because this is a conversation, and we want to actually maintain a lot of the metadata that actually makes up a conversation object, this is not all that's going on under under the hood, and we saw in the previous video a little bit about the conversation format. So it gets a little bit more complicated in that we have to take our user query, and we have to actually use this a chat format. So let me delete the system message; I don't think it's very important for the purposes of understanding what's going on. Let me paste my message as the user, and then let me paste the model response as an assistant, and then let me crop it here properly; the tool doesn't do that properly. So here we have it as it actually happens under the hood. There are all these special tokens that basically begin a message from the user, and then the user says, and this is the content of what we said, and then the user ends, and then the assistant begins and says this, etc. Now, the precise details of the conversation format are not important. What I want to get across here is that what looks to you and I as little chat bubbles going back and forth, under the hood we are collaborating with the model, and we're both writing into a token stream, and these two bubbles back and forth were in a sequence of exactly 42 tokens under the hood. I contributed some of the first tokens, and then the model continued the sequence of tokens with its response, and we could alternate and continue adding tokens here, and together we are building out a token window, a one-dimensional tokens, one-dimensional sequence of tokens.

Okay, so let's come back to ChatGPT. Now, what we are seeing here is kind of like little bubbles going back and forth between us and the model. Under the hood, we are building out a one-dimensional token sequence. When I click "New chat" here, that wipes the token window; that resets the tokens to basically zero again and restarts the conversation from scratch. Now, the cartoon diagram that I have in my mind when I'm speaking to a model looks something like this: when we click "New chat," we begin a token sequence; so this is a one-dimensional sequence of tokens. The user, we can write tokens into this stream, and then when we hit enter, we transfer control over to the language model, and the language model responds with its own token streams, and then the language model has a special token that basically says something along the lines of "I'm done." So when it emits that token, the ChatGPT application transfers control back to us, and we can take turns. Together, we are building out the token, the token stream, which we also call the context window. So the context window is kind of like this working memory of tokens, and anything that is inside this context window is kind of like in the working memory of this conversation and is very directly accessible by the model.

Now, what is this entity here that we are talking to, and how should we think about it? Well, this language model, here we saw that the way it is trained in the previous video; we saw there are two major stages: the pre-training stage and the post-training stage. The pre-training stage is kind of like taking all of the internet, chopping it up into tokens, and then compressing it into a single kind of like zip file. But the zip file is not exact; the zip file is a lossy and probabilistic zip file because we can't possibly represent all of the internet in just one, one sort of like, say, terabyte of zip file, because there's just way too much information. So we just kind of get the gist or the vibes inside this zip file. Now, what's actually inside the zip file are the parameters of a neural network. And so, for example, a one-terabyte zip file would correspond to roughly, say, one trillion parameters inside this neural network. And when this neural network is trying to do is it's trying to basically take tokens, and it's trying to predict the next token in a sequence, but it's doing that on internet documents. So it's kind of like this internet document generator, right? And in the process of predicting the next token on a sequence on the internet, the neural network gains a huge amount of knowledge about the world, and this knowledge is all represented and stuffed and compressed inside the one trillion parameters, roughly, of this language model. Now, this pre-training stage, also we saw, is fairly costly. So this can be many tens of millions of dollars, say like three months of training and so on. So this is a costly, long phase; for that reason, this phase is not done that often. So, for example, GPT-4, this model was pre-trained probably many months ago, maybe like even a year ago by now. And so that's why these models are a little bit out of date; they have what's called a knowledge cutoff because that knowledge cutoff corresponds to when the model was pre-trained, and its knowledge only goes up to that point. Now, some knowledge can come into the model through the post-training phase, which we'll talk about in a second, but roughly speaking, you should think of these models as kind of like a little bit out of date because pre-training is way too expensive and happens infrequently. So any kind of recent information, like if you wanted to talk to your model about something that happened last week or so on, we're going to need other ways of providing that information to the model, because it's not stored in the knowledge of the model. So we're going to have various tool uses to give that information to the model.

Now, after pre-training, there's a second stage; it goes post-training, and the post-training stage is really attaching a smiley face to this zip file because we don't want to generate internet documents; we want this thing to take on the persona of an assistant that responds to user queries, and that's done in a process of post-training where we swap out the dataset for a dataset of conversations that are built out by humans. So this is basically where the model takes on this persona and that actually, so that we can like ask questions, and it responds with answers. So it takes on the style of an assistant; that's post-training, but it has the knowledge of all of the internet, and that's by pre-training. So these two are combined in this artifact. Now, the important thing to understand here, I think for this section, is that what you are talking to is a fully self-contained entity. By default, this language model—think of it as a one-terabyte file on a disk secretly that represents one trillion parameters and their precise settings inside the neural network that's trying to give you the next token in the sequence—but this is the fully self-contained entity. There's no calculator, there's no computer and Python interpreter, there's no worldwide web browsing, there's none of that; there's no tool use yet in what we've talked about so far. You're talking to a zip file; if you stream tokens to it, it will respond with tokens back, and this zip file has the knowledge from pre-training, and it has the style and form from post-training. And so that's roughly how you can think about this entity.

Okay, so if I had to summarize what we talked about so far, I would probably do it in the form of an introduction of ChatGPT in a way that I think you should think about it. So the introduction would be: "Hi, I'm ChatGPT. I am a one-tab zip file. My knowledge comes from the internet, which I read in its entirety about six months ago, and I only remember vaguely. Okay? And my winning personality was programmed by example by human labelers at OpenAI." So the personality is programmed in post-training, and the knowledge comes from compressing the internet during pre-training, and this knowledge is a little bit out of date, and it's probabilistic and slightly vague. Some of the things that are probably mentioned very frequently on the internet, I will have a lot better, better recollection of than some of the things that are discussed very rarely, very similar to what you might expect with a human. So let's not talk about some of the repercussions of this entity and how we can talk to it and what kinds of things we can expect from it. Now, I'd like to use real examples when we actually go through this. So, for example, this morning, I asked ChatGPT the following: "How much caffeine is in one shot of Americano?" And I was curious because I was comparing it to matcha. Now, ChatGPT will tell me that this is roughly 63 mg of caffeine or so. Now, the reason I'm asking ChatGPT this question is that I think this is okay is number one, I'm not asking about any knowledge that is very recent, so I do expect that the model has sort of read about how much caffeine there is in one shot; this I don't think this information has changed too much. And number two, I think this information is extremely frequent on the internet; this kind of a question and this kind of information has occurred all over the place on the internet, and because there were so many mentions of it, I expect a model to have good memory of it in its knowledge. So there's no tool use, and the model, the zip file, responded that there's roughly 63 mg. Now, I'm not guaranteed that this is the correct answer; this is just its vague recollection of the internet, but I can go to primary sources, and maybe I can look up, okay, caffeine and Americano, and I could verify that, yeah, it looks to be about 63 is roughly right, and you can look at primary sources to decide if this is true or not. So I'm not strictly speaking guaranteed that this is true, but I think probably this is the kind of thing that ChatGPT would know. Here's an example of a conversation I had two days ago, actually, and there's another example of a knowledge-based conversation and things that I'm comfortable asking of ChatGPT with some caveats. So I'm a bit sick; I have a runny nose, and I want to get meds that help with that. So it told me a bunch of stuff, and I want my nose to not be runny, so I gave it a clarification based on what it said, and then it kind of gave me some of the things that might be helpful with that, and then I looked at some of the meds that I have at home and I said, "Does DayQuil or NyQuil work?" And it went off, and it kind of like went over the ingredients of DayQuil and NyQuil and whether or not they helped mitigate runny nose. Now, when these ingredients are coming here again, remember we are talking to a zip file that has a recollection of the internet; I'm not guaranteed that these ingredients are correct, and in fact, I actually took out the box and I looked at the ingredients and I made sure that NyQuil's ingredients are exactly these ingredients, and I'm doing that because I don't always fully trust what's coming out here, right? This is just a probabilistic statistical recollection of the internet, but that said, conversations of DayQuil and NyQuil, these are very common meds; probably there's tons of information about a lot of this on the internet, and this is the kind of things that the model has pretty good recollection of. So actually these were all correct, and then I said, "Okay, well, I have NyQuil; how far, how fast would it act, roughly?" And it kind of tells me, and then is a basically a tal and says yes. So this is a good example of how ChatGPT was useful to me; it is a knowledge-based query; this knowledge sort of isn't recent knowledge; this is all coming from the knowledge of the model. I think this is common information; this is not a high-stakes situation; I'm checking ChatGPT a little bit, but also this is not a high-stakes situation, so no big deal. So I popped an NyQuil, and indeed it helped. But that's roughly how I'm thinking about what's going back here.

Okay, so at this point, I want to make two notes. The first note I want to make is that naturally, as you interact with these models, you'll see that your conversations are growing longer, right? Anytime you are switching topics, I encourage you to always start a new chat. When you start a new chat, as we talked about, you are wiping the context window of tokens and resetting it back to zero. If it is the case that those tokens are not any more useful to your next query, I encourage you to do this because these tokens in this window are expensive, and they're expensive in kind of like two ways: number one, if you have lots of tokens here, then the model can actually find it a little bit distracting. So if this was a lot of tokens, the model might—this is kind of like the working memory of the model—the model might be distracted by all the tokens in the past when it is trying to sample tokens much later on, so it could be distracting, and it could actually decrease the accuracy of the model and of its performance. And number two, the more tokens are in the window, the more expensive it is by a little bit—not by too much, but by a little bit—to sample the next token in the sequence. So your model is actually slightly slowing down; it's becoming more expensive to calculate the next token, and the more tokens there are here. And so think of the tokens in the context window as a precious resource; think of that as the working memory of the model, and don't overload it with irrelevant information, and keep it as short as you can, and you can expect that to work faster and slightly better. Of course, if the information actually is related to your task, you may want to keep it in there, but I encourage you to, as often as you can, basically start a new chat whenever you are switching topics. The second thing is that I always encourage you to keep in mind what model you are actually using. So here in the top left, we can drop down, and we can see that we are currently using GPT-4. Now, there are many different models of many different flavors, and there are too many, actually, but we'll go through some of these over time. So we are using GPT-4 right now, and in everything that I've shown you, this is GPT-4. Now, when I open a new incognito window—so if I go to chat.openai.com and I'm not logged in—the model that I'm talking to here—so if I just say "hello"—the model that I'm talking to here might not be GPT-4; it might be a smaller version. Now, unfortunately, OpenAI does not tell me when I'm not logged in what model I'm using, which is kind of unfortunate, but it's possible that you are using a smaller, kind of dumber model. So if we go to the ChatGPT pricing page here, we see that they have three basic tiers for individuals: the free, plus, and pro. And in the free tier, you have access to what's called GPT-4 mini, and this is a smaller version of GPT-4; it is a smaller model with a smaller number of parameters; it's not going to be as creative; like its writing might not be as good; its knowledge is not going to be as good; it's going to probably hallucinate a bit more, etc. But it is kind of like the free offering; the free tier; they do say that you have limited access to GPT-4 and GPT-3.5 mini, but I'm not actually 100% sure; like it didn't tell us which model we were using, so we just fundamentally don't know. Now, when you pay for $20 per month—even though it doesn't say this, I think basically like they're screwing up on how they're describing this—but if you go to fine print, limits apply; we can see that the plus users get 80 messages every 3 hours for GPT-4. So that's the flagship, biggest model that's currently available as of today, that's available, and that's what we want to be using. So if you pay $20 per month, you have that with some limits, and then if you pay for $200 per month, you get the pro, and there's a bunch of additional goodies as well as unlimited GPT-4. And we're going to go into some of this because I do pay for a pro subscription. Now, the whole takeaway I want you to get from this is be mindful of the models that you're using. Typically with these companies, the bigger models are more expensive to calculate, and so therefore the companies charge more for the bigger models, and so make those tradeoffs for yourself depending on your usage of LLMs. Have a look at—you can get away with the cheaper offerings, and if the intelligence is not good enough for you, and you're using this professionally, you may really want to consider paying for the top-tier models that are available from these companies. In my case, in my professional work, I do a lot of coding and a lot of things like that, and this is still very cheap for me, so I pay this very gladly because I get access to some really powerful models that I'll show you in a bit. So yeah, keep track of what model you're using and make those decisions for yourself. I also want to show you that all the other LLM providers will all have different pricing tiers with different models at different tiers that you can pay for. So, for example, if we go to Claude from Anthropic, you'll see that I am paying for the professional plan, and that gives me access to Claude 3.5, Sonnet. And if you are not paying for a pro plan, then probably you only have access to maybe a haiku or something like that. And so use the most powerful model that kind of like works for you. Here's an example of me using Claude a while back; I was asking for just travel advice. So I was asking for a cool city to go to, and Claude told me that Zermatt in Switzerland is really cool. So I ended up going there for a New Year's break following Claude's advice. But this is just an example of another thing that I find these models pretty useful for is travel advice and ideation and getting pointers that you can research further. Here we also have an example of Gemini.google.com. So this is from Google; I got Gemini's opinion on the matter, and I asked it for a cool city to go to, and it also recommended Zermatt. So that was nice. So I like to go between different models and asking them similar questions and seeing what they think about. And for Gemini, also on the top left, we also have a model selector, so you can pay for the more advanced tiers and use those models. Same thing goes for Grok; just released; we don't want to be asking Grok 2 questions because we know that Grok 3 is the most advanced model, so I want to make sure that I pay enough such that I have Grok 3 access. So for all these different providers, find the one that works best for you; experiment with different providers; experiment with different pricing tiers for the problems that you are working on, and that's kind of—and often I end up personally just paying for a lot of them and then asking all of them the same question, and I kind of refer to all these models as my LLM council. So they're kind of like the council of language models; if I'm trying to figure out where to go on vacation, I will ask all of them. And so you can also do that for yourself if that works for you.

Okay, the next topic I want to now turn to is that of "thinking models," quote unquote. So we saw in the previous video that there are multiple stages of training: pre-training goes to supervised fine-tuning goes to reinforcement learning, and reinforcement learning is where the model gets to practice on a large collection of problems that resemble the practice problems in the textbook, and it gets to practice on a lot of math and code problems. And in the process of reinforcement learning, the model discovers thinking strategies that lead to good outcomes, and these thinking strategies, when you look at them, they very much resemble kind of the inner monologue you have when you go through problem-solving. So the model will try out different ideas; it will backtrack; it will revisit assumptions; and it will do things like that. Now, a lot of these strategies are very difficult to hardcode as a human labeler because it's not clear what the thinking process should be; it's only in the reinforcement learning that the model can try out lots of stuff, and it can find the thinking process that works for it with its knowledge and its capabilities. So this is the third stage of training these models; this stage is relatively recent, so only a year or two ago, and all of the different LLM labs have been experimenting with these models over the last year, and this is kind of like seen as a large breakthrough recently. And here we looked at the paper from DeepSeek that was the first to basically talk about it publicly, and they had a nice paper about incentivizing reasoning capabilities in LLMs via reinforcement learning. So that's the paper that we looked at in the previous video. So we now have to adjust our cartoon a little bit because basically what it looks like is our emoji now has this optional thinking bubble, and when you are using a thinking model, which will do additional thinking, you are using the model that has been additionally tuned with reinforcement learning. And qualitatively, what does this look like? Well, qualitatively, the model will do a lot more thinking, and what you can expect is that you will get higher accuracies, especially on problems that are, for example, math, code, and things that require a lot of thinking; things that are very simple like might not actually benefit from this, but things that are actually deep and hard might benefit a lot. And so but basically what you're paying for is that the models will do thinking, and that can sometimes take multiple minutes because the models will emit tons and tons of tokens over a period of many minutes, and you have to wait because the model is thinking just like a human would think. But in situations where you have very difficult problems, this might translate to higher accuracy. So let's take a look at some examples. So here's a concrete example: when I was stuck on a programming problem recently, so something called the gradient check fails, and I'm not sure why, and I copy-pasted the model, my code. So the details of the code are not important, but this is basically an optimization of a multilayer perceptron, and details are not important; it's a bunch of code that I wrote, and there was a bug because my gradient check didn't work, and I was just asking for advice, and GPT-4, which is the flagship, most powerful model for OpenAI, but without thinking, just kind of like went into a bunch of things that it thought were issues or that I should double-check but actually didn't really solve the problem; like all of the things that it gave me here are not the core issue of the problem. So the model didn't really solve the issue, and it tells me about how to debug it and so on. But then what I did was here in the dropdown I turned to one of the thinking models. Now, for OpenAI, all of these models that start with "O" are thinking models: O1, O3 mini, O3 mini high, and O1 Pro, promote are all thinking models, and they're not very good at naming their models, but that is the case. And so here they will say something like "uses advanced reasoning" or "good at code and logic" and stuff like that, but these are basically all tuned with reinforcement learning, and because I am paying for $200 per month, I have access to O1 Pro, which is best at reasoning, but you might want to try some of the other ones if depending on your pricing tier. And when I gave the same model the same prompt to O1 Pro, which is the best at reasoning model, and you have to pay $200 per month for this one, then the exact same prompt, it went off, and it thought for 1 minute, and it went through a sequence of thoughts, and OpenAI doesn't fully show you the exact thoughts; they just kind of give you little summaries of the thoughts, but it thought about the code for a while, and then it actually came to get, came back with the correct solution; it noticed that the parameters are mismatched and how I pack and unpack them and etc. So this actually solved my problem, and I tried out giving the exact same prompt to a bunch of other LLMs. So, for example, Claude; I gave Claude the same problem, and it actually noticed the correct issue and solved it, and it did that even with Sonnet, which is not a thinking model. So Claude 3.5, Sonnet, to my knowledge, is not a thinking model, and to my knowledge, Anthropic, as of today, doesn't have a thinking model deployed, but this might change by the time you watch this video, but even without thinking, this model actually solved the issue. When I went to Gemini, I asked it, and it also solved the issue, even though I also could have tried the a thinking model, but it wasn't necessary. I also gave it to Grok; Grok 3 in this case; and Grok 3 also solved the problem after a bunch of stuff. So it also solved the issue. And then finally, I went to Perplexity.ai, and the reason I like Perplexity is because when you go to the model dropdown, one of the models that they host is this DeepSeek R1. So this has the reasoning with the DeepSeek R1 model, which is the model that we saw over here; this is the paper. So Perplexity just hosts it and makes it very easy to use. So I copy-pasted it there, and I ran it, and I think they render, they like really render it terribly, but down here you can see the raw thoughts of the model, even though you have to expand them, but you see like, okay, the user is having trouble with the gradient check, and then it tries out a bunch of stuff, and then it says, "But wait, when they accumulate the gradients, they're doing the thing incorrectly; let's check the order; the parameters are packed as this," and then it notices the issue, and then it kind of like says that's a critical mistake, and so it kind of like thinks through it, and you have to wait a few minutes, and then also comes up with the correct answer. So basically, long story short, what do I want to show you? There exist a class of models that we call thinking models; all the different providers may or may not have a thinking model; these models are most effective for difficult problems in math and code and things like that, and in those kinds of cases, they can push up the accuracy of your performance. In many cases, like if you're asking for travel advice or something like that, you're not going to benefit out of a thinking model; there's no need to wait for one minute for it to think about some destinations that you might want to go to. So for myself, I usually try out the non-thinking models because their responses are really fast, but when I suspect the response is not as good as it could have been, and I want to give the opportunity to the model to think a bit longer about it, I will change it to a thinking model, depending on whichever one you have available to you. Now, when you go to Grok, for example, when I start a new conversation with Grok, when you put the question here like "hello," you should put something important here; you see here "Think," so let the model take its time; so turn on "Think," and then click "Go," and when you click "Think," Grok under the hood switches to the thinking model, and all the different LLM providers will kind of like have some kind of a selector for whether or not you want the model to think or whether it's okay to just like go with the previous kind of generation of the models.

Okay, now the next section I want to continue to is tool use. So far, we've only talked to the language model through text, and this language model is again this zip file in a folder; it's inert; it's closed off; it's got no tools; it's just a neural network that can emit tokens. So what we want to do now, though, is we want to go beyond that, and we want to give the model the ability to use a bunch of tools, and one of the most useful tools is an internet search. And so let's take a look at how we can make models use internet search. So, for example, again using concrete examples from my own life, a few days ago I was watching White Lotus season 3, and I watched the first episode, and I love this TV show, by the way, and I was curious when episode two was coming out. And so in the old world, you would imagine you go to Google or something like that; you put in like "new episodes of White Lotus season 3," and then you start clicking on these links, and maybe open a few of them or something like that, right? And you start like searching through it and trying to figure it out, and sometimes you luck out and you get a schedule, but many times you might get really crazy ads; there's a bunch of random stuff going on, and it's just kind of like an unpleasant experience, right? So wouldn't it be great if a model could do this kind of a search for you, visit all the web pages, and then take all those web pages, take all their content, and stuff it into the context window, and then basically give you the response? And that's what we're going to do now. Basically, we haven't a mechanism or a way; we introduce a mechanism for the model to emit a special token that is some kind of a searchy internet token, and when the model emits the searchy internet token, the ChatGPT application or whatever LLM application it is you're using will stop sampling from the model, and it will take the query that the model gave it; goes off; it does a search; it visits web pages; it takes all of their text; and it puts everything into the context window. So now you have this internet search tool that itself can also contribute tokens into our context window, and in this case, it would be like lots of internet web pages, and maybe there's 10 of them, and maybe it just puts it all together, and this could be thousands of tokens coming from these web pages just as we were looking at them ourselves. And then after it has inserted all those web pages into the context window, it will reference back to your question as to, "Hey, what, when is this, when is this season getting released?" And it will be able to reference the text and give you the correct answer. And notice that this is a really good example of why we would need internet search; without the internet search, this model has no chance to actually give us the correct answer because, like I mentioned, this model was trained a few months ago; the schedule probably was not known back then. And so when White Lotus season 3 is coming out is not part of the real knowledge of the model, and it's not in the zip file most likely because this is something that was presumably decided on in the last few weeks. And so the model has to basically go off and do internet search to learn this knowledge, and it learns it from the web pages just like you and I would, without it, and then it can answer the question once that information is

The button says "Search the web," so we're going to take a look at that in a second. For now: when are new episodes of White Lotus season 3 getting released? You can just ask, and instead of having to do the work manually, we just hit enter. The model will visit these web pages, create all the queries, and then give you the answer. So it just kind of did a ton of the work for you. Usually, there will be citations so you can actually visit those web pages yourself and make sure these are not hallucinations from the model. You can double-check that this is actually correct because it's not, in principle, guaranteed; it's just something that may or may not work.

If we take this, we can also go to, for example, ChatGPT, say the same thing. But now, when we put this question in without actually selecting "search," I'm not actually 100% sure what the model will do. In some cases, the model will actually know that this is recent knowledge and that it probably doesn't know, and it will create a search. In some cases, we have to declare that we want to do the search. In my own personal use, I would know that the model doesn't know, and so I would just select "search." But let's see first; let's see what happens. Okay, searching the web, and then it prints stuff and then it cites. The model actually detected itself that it needs to search the web because it understands that this is some kind of recent information, etc. So this was correct. Alternatively, if I create a new conversation, I could have also selected "search" because I know I need to search. Enter, and then it does the same thing: searching the web, and that's the result.

Basically, when you're using these LLMs, look for this—for example, Grok. Excuse me, let's try Grok without it, without selecting "search." Okay, so the model does some searching, just knowing that it needs to search, and gives you the answer. Basically, let's see what Claude does. You see, Claude does actually have the search tool available, so it will say, "As of my last update in April 2024..." This last update is when the model went through pre-training. So Claude is just saying, as of my last update, the knowledge cut-off of April 2024, it was announced, but it doesn't know. Claude doesn't have the internet search integrated as an option and will not give you the answer. I expect that this is something that Anthropic might be working on. Let's try Gemini and let's see what it says. Unfortunately, no official release date for White Lotus season 3 yet. So Gemini 2.0 Pro experimental does not have access to internet search and doesn't know. We could try some of the other ones, like 2.0 Flash. Let me try that. Okay, so this model seems to know, but it doesn't give citations. Oh wait, okay, there we go: sources and related content. So we see how 2.0 Flash actually has the internet search tool, but I'm guessing that the 2.0 Pro, which is the most powerful model that they have, this one actually does not have access, and in here it actually tells us: "2.0 Pro experimental lacks access to real-time info and some Gemini features." So this model is not fully wired with internet search.

Long story short, we can get models to perform Google searches for us, visit the web page, just pull in the information to the context window, and answer questions. This is a very, very cool feature, but different models, possibly different apps, have different amounts of integration of this capability, and so you have to be kind of on the lookout for that. Sometimes the model will automatically detect that they need to do a search, and sometimes you're better off telling the model that you want it to do the search. So when I'm doing GPT-4 and I know that this requires a search, you probably will not tick that box. So that's search tools. I wanted to show you a few more examples of how I use the search tool in my own work. So what are the kinds of queries that I use? This is fairly easy for me to do because usually for these kinds of cases I go to Perplexity, just out of habit, even though ChatGPT today can do this kind of stuff as well, as do probably many other services as well, but I happen to use Perplexity for these kinds of search queries. Whenever I expect that the answer can be achieved by doing basically something like a Google search and visiting a few of the top links, and the answer is somewhere in those top links, whenever that is the case, I expect to use the search tool, and I come to Perplexity.

Here are some examples: Is the market open today? This was President's Day; I wasn't 100% sure. So Perplexity understands what today is; it will do the search and figure out that it's President's Day, this was closed. Where's White Lotus season 3 filmed? Again, this is something that I wasn't sure that a model would know in its knowledge; this is something niche, so maybe there aren't that many mentions of it on the internet, and also this is more recent, so I don't expect a model to know by default. So this was a good fit for the search tool. Does Vercel offer PostgreSQL database? This was a good example of this because this kind of stuff changes over time, and the offerings of Vercel, which is a company, may change over time, and I want the latest. Whenever something is latest or something changes, I prefer to use the search tool, so I come to Perplexity. When is—what does Apple launch tomorrow, and what are some of the rumors? Again, this is something recent. Where is the Singles Inferno season 4 cast? Must know. So this is again a good example because this is very fresh information. Why is the Palantir stock going up? What is driving the enthusiasm? When is Civilization 7 coming out? Exactly. This is an example also like: Has Brian Johnson talked about his toothpaste uses? And I was curious; basically, I like what Brian does, and again it has two features: number one, it's a little bit esoteric, so I'm not 100% sure if this is at scale on the internet and would be part of like the knowledge of a model; and number two, this might change over time, so I want to know what toothpaste he uses most recently. So this is a good fit again for a search tool. Is it safe to travel to Vietnam? This can potentially change over time. And then I saw a bunch of stuff on Twitter about a USAID, and I wanted to know kind of like what's the deal, so I searched about that. And then you can kind of like dive in in a bunch of ways here, but this use case here is kind of along the lines of: I see something trending and I'm kind of curious what's happening, like what is the gist of it, and so I very often just quickly bring up a search of like what's happening and then get a model to kind of just give me a gist of roughly what happened, because a lot of the individual tweets or posts might not have the full context just by itself. These are examples of how I use a search tool.

Okay, next up, I would like to tell you about this capability called "Deep research." This is fairly recent—only as of like a month or two ago—but I think it's incredibly cool and really interesting and kind of went under the radar for a lot of people, even though I think it shouldn't have. When we go to ChatGPT pricing here, we notice that deep research is listed here under Pro, so it currently requires $200 per month. This is the top tier; however, I think it's incredibly cool. Let me show you, by example, in what kinds of scenarios you might want to use it. Roughly speaking, deep research is a combination of internet search and thinking, rolled out for a long time. The model will go off and spend tens of minutes doing what deep research. A first sort of company that announced this was ChatGPT as part of its Pro offering, very recently, like a month ago.

Here's an example: Recently, I was on the internet buying supplements—which I know is kind of crazy—but Brian Johnson has this starter pack, and I was kind of curious about it. There's this thing called Longevity Mix, right, and it's got a bunch of health actives, and I want to know what these things are, right? Of course, like, so like, CAKG, like, what the hell is this? "Boost energy production for sustained vitality"—like, what does that mean? One thing you could, of course, do is you could open up Google search and look at the Wikipedia page or something like that and do everything you're kind of used to, but deep research allows you to basically take an alternate route, and it kind of like processes a lot of this information for you and explains it a lot better. As an example, we can do something like this: This is my example prompt: "CAKG is one health—one of the health actives in Brian Johnson's Blueprint at 2.5 grams per serving. Can you do research on CAKG? Tell me why—tell me about why it might be found in the Longevity Mix; its possible efficacy in humans or animal models; its potential mechanism of action; any potential concerns or toxicity or anything like that." Now, here I have this button available to me, and you won't, unless you pay $200 per month right now, but I can turn on deep research. Let me copy-paste this and hit "go." And now the model will say, "Okay, I'm going to research this," and then sometimes it likes to ask clarifying questions before it goes off. So a focus on human clinical studies? Animal models are both? So let's say both. Specific sources? All of all sources? I don't know. Comparison to other longevity compounds? Not needed. Comparison, just CAKG. We can be pretty brief. The model understands, and we hit "go." And then, "Okay, I'll research CAKG. Starting research." And so now we have to wait for probably about 10 minutes or so. If you'd like to click on it, you can get a bunch of previews of what the model is doing on a high level. This will go off and do a combination of, like I said, thinking and internet search, but it will issue many internet searches; it will go through lots of papers; it will look at papers, and it will think, and it will come back 10 minutes from now. So this will run for a while. Meanwhile, while this is running, I'd like to show you equivalents of it in the industry. Inspired by this, a lot of people were interested in cloning it, and so one example is, for example, Perplexity. Perplexity, when you go to the model drop-down, has something called "Deep research," and so you can issue the same queries here, and we can give this to Perplexity, and then Grok as well has something called "Deep search" instead of "deep research," but I think that Grok's deep search is kind of like deep research, but I'm not 100% sure. So we can issue Grok deep search as well: Grok 3 deep search, go. And this model is going to go off as well. Now I think—where is my ChatGPT? So ChatGPT is kind of like maybe a quarter done. Perplexity is going to be done soon. Okay, still thinking, and Grok is still going as well. I like Grok's interface the most; it seems like—okay. So basically, it's looking up all kinds of papers, WebMD browsing results, and it's kind of just getting all this. Now, while this is all going on, of course, it's accumulating a giant context window and it's processing all that information, trying to kind of create a report for us. So key points: What is CAKG and why is it in Longevity Mix? How is it associated with longevity, etc.? And so it will do citations and it will kind of like tell you all about it. And so this is not a simple and short response; this is a kind of like almost like a custom research paper on any topic you would like, and so this is really cool, and it gives a lot of references, potentially for you to go off and do some of your own reading and maybe ask some clarifying questions afterward, but it's actually really incredible that it gives you all these different citations and processes the information for you a little bit. Let's see if Perplexity finished. Okay, Perplexity is still—still researching, and ChatGPT is also researching. So let's briefly pause the video, and I'll come back when this is done.

Okay, so Perplexity finished, and we can see some of the report that it wrote up. So there's some references here and some basically description, and then ChatGPT also finished, and it also thought for 5 minutes, looked at 27 sources, and produced a report. So here it talked about research in worms, drosophila, in mice, and in human trials that are ongoing, and then a proposed mechanism of action and some safety and potential concerns and references, which you can dive deeper into. Usually, in my own work right now, I've only used this maybe for like 10 to 20 queries so far, something like that. Usually, I find that the ChatGPT offering is currently the best; it is the most thorough; it reads the best; it is the longest; it makes the most sense when I read it; and I think the Perplexity and the Grok are a little bit—a little bit shorter and a little bit briefer and don't quite get into the same detail as the deep research from Google—from ChatGPT. Right now, I will say that everything that is given to you here—again, keep in mind that even though it is doing research and it's pulling in—there are no guarantees that there are no hallucinations here; any of this can be hallucinated at any point in time; it can be totally made up, fabricated, misunderstood by the model. So that's why these citations are really important. Treat this as your first draft; treat this as papers to look at, but don't take this as definitely true. So here what I would do now is I would actually go into these papers and I would try to understand—is the ChatGPT understanding it correctly?—and maybe I have some follow-up questions, etc. So you can do all that, but still incredibly useful to see these reports once in a while to get a bunch of sources that you might want to descend into afterward.

Okay, so just like before, I wanted to show a few brief examples of how I've used deep research. For example, I was trying to change browsers because Chrome was not—Chrome upset me, and so it deleted all my tabs. So I was looking at either Brave or Arc, and I—I was most interested in which one is more private, and basically ChatGPT compiled this report for me, and I—this was actually quite helpful, and I went into some of the sources, and I sort of understood why Brave is—basically, TL;DR—significantly better, and that's why, for example, here I'm using Brave because I switched to it now. And so this is an example of basically researching different kinds of products and comparing them. I think that's a good fit for deep research. Here, I wanted to know about life extension in mice. So it kind of gave me a very long reading, but basically, mice are an animal model for longevity, and different labs have tried to extend it with various techniques. And then here I wanted to explore LLM labs in the USA, and I wanted a table of how large they are, how much funding they've had, etc. So this is the table that it produced. Now, this table is basically hit and miss, unfortunately. So I wanted to show it as an example of a failure. I think some of these numbers—I didn't fully check them—but they don't seem way too wrong. Some of this looks wrong, but the bigger omission I definitely see is that xAI is not here, which I think is a really major omission. And then also, conversely, Hugging Face should probably not be here because I asked specifically about LLM labs in the USA. Also, a Luther AI, I don't think should count as a major LLM lab, due mostly to its resources. And so I think it's kind of a hit and miss; things are missing; I don't fully trust these numbers; I have to actually look at them. So again, use it as a first draft; don't fully trust it; still very helpful. That's it.

So what's really happening here that is interesting is that we are providing the LLM with additional concrete documents that it can reference inside its context window. So the model is not just relying on the hazy knowledge of the world through its parameters and what it knows in its brain; we're actually giving it concrete documents. It's as if you and I reference specific documents, like on the internet or something like that, while we are kind of producing some answer for some question. Now we can do that through an internet search or like a tool like this, but we can also provide these LLMs with concrete documents ourselves through a file upload, and I find this functionality pretty helpful in many ways. As an example, let's look at Claude because they just released Claude 3.7 while I was filming this video. This is a new Claude model that is now the state-of-the-art, and notice here that we have "thinking mode" now as of 3.7. So "normal" is what we looked at so far, but they just released "extended," best for math and coding challenges. And what they're not saying, but is actually true under the hood, probably most likely, is that this was trained with reinforcement learning in a similar way that all the other thinking models were produced. So what we can do now is we can upload documents that we wanted to reference inside its context window. As an example, there's this paper that came out that I was kind of interested in; it's from the Arc Institute, and it's basically a language model trained on DNA. And so I was kind of curious—I mean, I'm not from biology—but I was kind of curious what this is, and this is a perfect example of what LLMs are extremely good for because you can upload these documents to the LLM, and you can load this PDF into the context window and then ask questions about it and basically read the document together with an LLM and ask questions off it. The way you do that is you basically just drag and drop. So we can take that PDF and just drop it here. This is about 30 megabytes. Now, when Claude gets this document, it is very likely that they actually discard a lot of the images and that kind of information. I don't actually know exactly what they do under the hood, and they don't really talk about it, but it's likely that the images are thrown away, or if they are there, they may not be as—as well understood as you and I would understand them potentially. And it's very likely that what's happening under the hood is that this PDF is basically converted to a text file, and that text file is loaded into the token window, and once it's in the token window, it's in the working memory, and we can ask questions of it. Typically, when I start reading papers together with any of these LLMs, I just ask for: "Can you give me a summary—summary of this paper?" Let's see what Claude 3.7 says. Okay, "I'm exceeding the length limit of this chat." Oh god, really? Oh damn. Okay, well, let's try ChatGPT: "Can you summarize this paper?" And we're using GPT-4, and we're not using "thinking," which is okay; we don't—we can start by not thinking. "Reading documents. Summary of the paper: Genome modeling and design across all domains of life. So this paper introduces Evo 2, a large-scale biological foundation model, and then key features and so on." So I personally find this pretty helpful, and then we can kind of go back and forth, and as I'm reading through the abstract and the introduction, etc., I am asking questions of the LLM, and it's kind of like making it easier for me to understand the paper.

Another way that I like to use this functionality extensively is when I'm reading books. It is rarely ever the case anymore that I read books just by myself; I always involve an LLM to help me read a book. A good example of that recently is *The Wealth of Nations*, which I was reading recently. It is a book from 1776, written by Adam Smith, and it's kind of like the foundation of classical economics, and it's a really good book, and it's kind of just very interesting to me that it was written so long ago, but it has a lot of modern-day kind of like—it's just got a lot of insights that I think are very timely even today. The way I read books now, as an example, is you basically pull up the book, and you have to get access to like the raw content of that information. In the case of *The Wealth of Nations*, this is easy because it is from 1776, so you can just find it on Project Gutenberg, as an example, and then basically find the chapter that you are currently reading. As an example, let's read this chapter from Book One, and this chapter I was reading recently, and it kind of goes into the division of labor and how it is limited by the extent of the market. Roughly speaking, if your market is very small, then people can't specialize, and specialization is what—is basically huge; specialization is extremely important for wealth creation because you can have experts who specialize in their simple little task, but you can only do that at scale because without the scale you don't have a large enough market to sell to your specialization. So what we do is we copy-paste this book—this chapter, at least—this is how I like to do it. We go to, say, Claude, and we say something like: "We are reading *The Wealth of Nations*." Now remember, Claude kind of has knowledge of *The Wealth of Nations*, but probably doesn't remember exactly the content of this chapter, so it wouldn't make sense to ask Claude questions about this chapter directly because it probably doesn't remember—remember what this chapter is about, but we can remind Claude by loading this into the context window. So, "We're reading *The Wealth of Nations*. Please summarize this chapter to start." And then what I do here is I copy-paste. Now, in Claude, when you copy-paste, they don't actually show all the text inside the text box; they create a little text attachment when it is over some size, and so we can click enter, and we just kind of like start off. Usually, I like to start off with a summary of what this chapter is about, just so I have a rough idea, and then I go in and I start reading the chapter, and any point we have any questions, then we just come in and just ask our question. And I find that basically going hand-in-hand with LLMs dramatically increases my retention, my understanding of these chapters, and I find that this is especially the case when you're reading, for example, documents from other fields, like, for example, biology, or, for example, documents from a long time ago, like 1776, where you sort of need a little bit of help even understanding what the basics of the language are. For example, I would feel a lot more courage approaching a very old text that is outside of my area of expertise; maybe I'm reading Shakespeare, or I'm reading things like that. I feel like LLMs make a lot of reading very dramatically more accessible than it used to be before because you're not just right away confused; you can actually kind of go slowly through it and figure it out together with the LLM in hand. So I use this extensively, and I think it's extremely helpful. I'm not aware of tools, unfortunately, that make this very easy for you today. I do this clunky back and forth; so literally, I will find the book somewhere, and I will copy-paste stuff around, and I'm going back and forth, and it's extremely awkward and clunky, and unfortunately, I'm not aware of a tool that makes this very easy for you, but obviously what you want is, as you're reading a book, you just want to highlight the passage and ask questions about it. This currently, as far as I know, does not exist, but this is extremely helpful. I encourage you to experiment with it, and don't read books alone.

Okay, the next very powerful tool that I now want to turn to is the use of a Python interpreter, or basically giving the ability to the LLM to use and write computer programs. So instead of the LLM giving you an answer directly, it has the ability now to write a computer program and to emit special tokens that the ChatGPT application recognizes as: "Hey, this is not for the human; this is"—basically saying that whatever I output here is actually a computer program; please go off and run it and give me the result of running that computer program. So it is the integration of the language model with a programming language here, like Python. So this is extremely powerful. Let's see the simplest example of where this would be used and what this would look like. So if I go—go to ChatGPT and I give it some kind of a multiplication problem—problem—let's say 30 * 9 or something like that—then this is a fairly simple multiplication, and you and I can probably do something like this in our head, right? Like 30 * 9, you can just come up with the result of 270, right? So let's see what happens. Okay, so the LLM did exactly what I just did; it calculated the result of this multiplication to be 270, but it's actually not really doing math; it's actually more like almost memory work, but it's easy enough to do in your head. So there was no tool use involved here; all that happened here was just the zip file doing next token prediction and gave the correct result here in its head. The problem now is: What if we want something more—more complicated? So what is this times this? And now, of course, this—if I asked you to calculate this, you would give up instantly because you know that you can't possibly do this in your head, and you would be looking for a calculator, and that's exactly what the LLM does now too. OpenAI has trained ChatGPT to recognize problems that it cannot do in its head and to rely on tools instead. So what I expect ChatGPT to do for this kind of a query is to turn to tool use. Let's see what it looks like. Okay, there we go. So what's opened up here is what's called the Python interpreter, and Python is basically a little programming language. Instead of the LLM telling you directly what the result is, the LLM writes a program, and then—not shown here are special tokens that tell the ChatGPT application to please run the program—and then the LLM pauses execution; instead, the Python program runs, creates a result, and then passes this—this result back to the language model as text, and the language model takes over and tells you that the result of this is that. So this is tool use—incredibly powerful—and OpenAI has trained ChatGPT to kind of like know in what situations to use tools, and they've taught it to do that by example. So human labelers are involved in curating datasets that kind of tell the model by example in what kinds of situations it should lean on tools and how. But basically, we have a Python interpreter, and this is just an example of multiplication, but this is significantly more powerful. Let's see what we can actually do inside programming languages.

Before we move on, I just wanted to make the point that unfortunately you have to kind of keep track of which LLMs that you're talking to have different kinds of tools available to them because different LLMs might not have all the same tools, and in particular, LLMs that do not have access to the Python interpreter or programming language or are unwilling to use it might not give you correct results in some of these harder problems. As an example, here we saw that ChatGPT correctly used a programming language and didn't do this in its head. Grok 3 actually, I believe, does not have access to a programming language like a Python interpreter, and here it actually does this in its head and gets remarkably close, but if you actually look closely at it, it gets it wrong; this should be 1120 instead of 060. So Grok 3 will just hallucinate through this multiplication and do it in its head and get it wrong, but actually like remarkably close. Then I tried Claude, and Claude actually wrote—in this case, not Python code, but it wrote JavaScript code, but JavaScript is also a programming language, and gets the correct result. Then I came to Gemini, and I asked 2.0 Pro, and Gemini did not seem to be using any tools; there's no indication of that, and yet it gave me what I think is the correct result, which actually kind of surprised me. So Gemini, I think, actually calculated this in its head correctly, and the way we can tell that this is—which is kind of incredible—the way we can tell that it's not using tools is we can just try something harder: What is...? We have to make it harder for it. Okay, so it gives us some result, and then I can use my calculator here, and it's wrong, right? So this is using my MacBook Pro calculator, and two—it's—it's not correct, but it's like remarkably close, but it's not correct, but it will just hallucinate the answer. So I guess like my point is, unfortunately, the state of the LLMs right now is such that different LLMs have different tools available to them, and you kind of have to keep track of it, and if they don't have the tools available, they'll just do their best, which means that they might hallucinate a result for you. So that's something to look out for.

Okay, so one practical setting where this can be quite powerful is what's called ChatGPT Advanced Data Analysis, and as far as I know, this is quite unique to ChatGPT itself, and it basically gets ChatGPT to be kind of like a junior data analyst who you can kind of collaborate with. Let me show you a concrete example without going into the full detail. First, we need to get some data that we can analyze and plot and chart, etc. So here, in this case, I said: "Let's research OpenAI valuation as an example," and I explicitly asked ChatGPT to use the search tool because I know that under the hood such a thing exists, and I don't want it to be hallucinating data to me; I wanted to actually look it up and back it up and create a table where each year we have the valuation. So these are the OpenAI valuations over time. Notice how in 2015 it's not applicable, so the valuation is like unknown. Then I said: "Now plot this. Use log scale for y-axis." And so this is where this gets powerful; ChatGPT goes off and writes a program that plots the data over here, so it creates a little figure for us, and it sort of ran it and showed it to us. This can be quite nice and valuable because it's a very easy way to basically collect data, upload data in a spreadsheet, and visualize it, etc. I will note some of the things here. As an example, notice that we had "NA" for 2015, but ChatGPT, when it was writing the code—and again, I would always encourage you to scrutinize the code—it put in 0.1 for 2015, and so basically it implicitly assumed that—it made the assumption here in code that the valuation of 2015 was 100 million, and because it put in 0.1, it's kind of like did it without telling us. So it's a little bit sneaky, and that's why you kind of have to pay attention a little bit to the code. So I'm familiar with the code, and I always read it, but I think I would be hesitant to potentially recommend the use of these tools if people aren't able to like read it and verify it a little bit for themselves. Now, "fit a trend line and extrapolate until the year 2030. Mark the expected valuation in 2030." So it went off, and it basically did a linear fit, and it's using SciPy curve_fit, and it did this and came up with a plot, and it told me that the valuation based on the trend in 2030 is approximately 1.7 trillion, which sounds amazing, except here I became suspicious because I see that ChatGPT is telling me it's 1.7 trillion, but when I look here at 2030, it's printing 2027, 1.7B. So its extrapolation, when it's printing the variable, is inconsistent with 1.7 trillion. This makes it look like that valuation should be about 20 trillion, and so that's what I said: "Print this variable directly by itself. What is it?" And then it sort of like rewrote the code and gave me the variable itself, and as we see in the label here, it is indeed 2271, etc. So in 2030, the true exponential trend extrapolation would be a valuation of 20 trillion. So I was like—I was trying to confront ChatGPT, and I was like, "You lied to me," right? And it's like, "Yeah, sorry, I messed up." So I guess I—I like this example because number one, it shows the power of the tool in that it can create these figures for you, and it's very nice, but I think number two, it shows the trickiness of it, where, for example, here it made an implicit assumption, and here it actually told me something—it told me just the wrong—it hallucinated 1.7 trillion. So again, it is kind of like a very, very junior data analyst; it's amazing that it can plot figures, but you have to kind of still know what this code is doing, and you have to be careful and scrutinize it and make sure that you are really watching very closely because your junior analyst is a little bit absent-minded and not quite right all the time. So really powerful, but also be careful with this. I won't go into full details of Advanced Data Analysis, but there were many videos made on this topic, so if you would like to use some of this in your work, then I encourage you to look at some of these videos. I'm not going to go into the full detail. So a lot of promise, but be careful.

Okay, so I've introduced you to ChatGPT and Advanced Data Analysis, which is one powerful way to basically have LLMs interact with code and add some UI elements like showing of figures and things like that. I would now like to introduce you to one more related tool, and that is specific to Claude, and it's called "Artifacts." Let me show you by example what this is. So I have a conversation with Claude, and I'm asking: "Generate 20 flashcards from the following text," and for the text itself, I just came to the Adam Smith Wikipedia page, for example, and I copy-pasted this introduction here. So I copy-pasted this here and asked for flashcards, and Claude responds with 20 flashcards. For example: "When was Adam Smith baptized? On June 16th, etc." "When did he die?" "What was his nationality?" etc. Once we have the flashcards, we actually want to practice these flashcards, and so this is where I continue the conversation, and I say: "Now use the artifacts feature to write a flashcards app to test these flashcards." And so Claude goes off and writes code for an app that basically formats all of this into flashcards, and that looks like this. So what Claude wrote specifically was this C++ code here. So it uses a React library and then basically creates all these components; it hardcodes the Q&A into this app, and then all the other functionality of it, and then the Claude interface basically is able to load these React components directly in your browser, and so you end up with an app. So, "When was Adam Smith baptized?" And you can click to reveal the answer, and then you can say whether you got it correct or not. "When did he die?" "What was his nationality?" etc. You can imagine doing this, and then maybe we can reset the progress or shuffle the cards, etc. So what happened here is that Claude wrote us a super duper custom app just for us right here, and typically what we're used to is some software engineers write

To share, there are many things people have developed: various timers, games, and so on. But one use case I found very useful in my own work is diagram generation.

For example, let's go back to the Adam Smith book chapter we were looking at. Sometimes, when reading *The Wealth of Nations* by Adam Smith, I'll attach chapter 3 of book one and ask: "Please create a conceptual diagram of this chapter." Claude, upon hearing "conceptual diagram of this chapter," often writes code like this: If you're not familiar with this, it uses the mermaid library to create or define a graph, and this plots that mermaid diagram. Claude analyzes the chapter and figures out the key principle being communicated: the division of labor is related to the extent of the market (its size). The chapter's pieces include a comparative example of trade (how much easier it is on land versus water), a specific example using geographic factors, and a comparison of land versus water transport (how much easier water transport is). Early civilizations benefited from water transport and flourished because it supported specialization. If you're a visual thinker (like me), laying out information like a tree helps remember the chapter's content easily. I enjoy these diagrams, getting a sense of the argument's layout and spatial arrangement. If you're like me, you'll enjoy this; you can make diagrams of anything: books, chapters, source code, anything really. I specifically find this fairly useful.

I've shown you that LLMs are quite good at writing code. Not only can they emit code, but many apps (like ChatGPT and Claude) have started to partially run that code in the browser. ChatGPT will create figures and show them; Claude artifacts will integrate your React component, allowing you to use it inline in the browser.

Personally and professionally, I spend most of my time writing code. However, I don't go to ChatGPT and ask for code snippets because that's too slow; ChatGPT lacks the context to work with me professionally to create code. The same goes for other LLMs. Instead of using LLM features in a web browser, I use a specific app (many people in the industry do as well). This can be multiple apps: VS Code, Windsurf, Cursor, etc. I currently like to use Cursor. It's a separate app (for example, for your MacBook) that works with files on your file system. It's not a web page; it's a program you download that references files on your computer and edits them with you.

Here's a simple example of a React app I built in a few minutes with Cursor. Under the hood, Cursor uses Claude 3.7 sonnet; it calls Anthropic's API, asking Claude to do all this stuff. I don't have to manually go to Claude and copy-paste chunks of code; this program does that for me, having all the context of the files in the directory and all that. The app I developed is a very simple Tic-Tac-Toe. Claude wrote this in probably a minute. We can play; X can win, or we can tie.

I want to show you how to use Cursor efficiently (that's a whole separate video). I started with a completely new project and asked the composer app (its composer feature) to set up a new React repository, delete a lot of boilerplate, and make a simple Tic-Tac-Toe app. Cursor did all that; I only wrote five sentences. It changed everything, writing all the CSS and JavaScript. I'm running it locally and interacting with it in my browser.

Cursor has the context of your apps and uses Claude remotely through an API without accessing a web page. Many people develop this way. These tools have become more elaborate. Initially, you could only say, "Control K, please change this line of code to do this or that." After that, there was a control L command (command L): "Explain this chunk of code." An LLM explains the code; under the hood, it calls the same API you'd access if you entered it manually. This program has access to all the files, so it has all the context. Now, we're beyond command K and command L; we have command I (the composer tool). Especially with new agent integration, the composer is an autonomous agent on your codebase. It executes commands, changes files as needed, edits across multiple files. You mostly sit back, giving commands. I call this "vibe coding"—giving control to the composer and telling it what to do, hoping it works. Worst case, you can fall back to good old programming; we have all the files, can go over all the CSS, and inspect everything. As a programmer, you can change this arbitrarily, but now you have a very helpful assistant that can do a lot of the low-level programming for you.

Let's take it for a spin. If X or O wins, I want confetti. Let's see what it comes up with. I'll add a confetti effect when a player wins. It wants me to run React Confetti (a library I didn't know about). It installed it and is updating the app (app.tsx, the TypeScript file) to add the confetti effect. It's generating; we should see it soon. It added code here, here, and here. I'll add additional styling to make the winning cell stand out. It's adding CSS for winning cells. I'm not fully tracking this; it imported confetti. This all seems straightforward and reasonable, but I'd have to really dig in. It wants to add a sound effect when a player wins (ambitious!). I'm not sure how it'll access a sound file or where it'll get it from. Every time it saves a file, we deploy it; we can refresh and see what we have. It added a fade-in effect (cool!). I win! Whoa! I didn't expect that to work; this is elaborate. Let's play again. Whoa! Oh, I see; it paused and is waiting for me to confirm commands ("Make public sounds"). I have to confirm explicitly. Let's create a simple audio component to play a victory sound (sound/victory.mp3). The problem is victory.mp3 doesn't exist. I wonder what it'll do. It's downloading it from somewhere. Let's go along with it. Let's add a fallback if the sound file doesn't exist. In this case, it does exist. We can add and create a git commit. The composer thinks it's done. Let's try it. Pretty impressive! I don't know where it got the sound file or where that URL comes from, but maybe it appears in many repositories, and Claude knows about it. I'm happy with this; we can accept all. We could continue developing this app; worst case, if we can't debug anything, we can fall back to standard programming instead of vibe coding.

So far, we've interacted with a model via text: typing text in and getting text back. Now, let's talk about different modalities: interacting with models in more native human formats. I want to speak to it, have it speak back, give it images or videos, and vice versa (generate images and videos). It needs to handle speech, audio, images, and video.

First, let's cover talking to these models. I'd say I type about 50% of the time, and 50% of the time I'm too lazy and speak to the model. On mobile, it's even more pronounced (probably 80% of my queries are speech because I'm too lazy to type).

On the phone, it's easy. The ChatGPT app looks like this. There are two voice modes: a microphone icon and an audio icon. The microphone icon lets the app listen to your voice and transcribe it to text (so you don't have to type). It's easy; I do this all the time. You open the app, create a new conversation, hit the button, and ask something like, "Why is the sky blue? Is it because it's reflecting the ocean?" Then you click okay; it converts your audio to text, and you hit go to get a response.

On desktop, things are more complicated. The desktop app has an audio icon ("Use voice mode"), but no microphone icon; I can't speak to it and have it transcribed to text inside the app. On my MacBook, I use apps that allow this functionality (it's not specific to ChatGPT, it's system-wide functionality of taking your audio and transcribing it to text). Some apps people use are Super Whisper, Whisper Flow, Mac Whisper, etc. I use Super Whisper; it's quite good. You download it, install it on your MacBook, and it's always ready to listen. You bind a key (I use F5). When I press F5, it listens; I say something, press F5 again, and it transcribes it to text. I'll press F5: "I have a question: why is the sky blue? Is it because it's reflecting the ocean?" Enter; I didn't have to type anything. About half my queries are like this because I don't want to type. Many queries require product names or library names, which don't always transcribe well; in those cases, I type it out. In simple, day-to-day use, I often speak to the model, and it transcribes correctly. That's the input side.

On the output side, apps usually let you read it back to you (taking the text and passing it to a model that does the inverse of text-to-speech). In ChatGPT, this icon says "Read aloud." Different apps may or may not have this functionality, or you could download a system-wide text-to-speech app. Those are the options; don't type; use voice; it works well; I use it pervasively; about half (or more) of my queries are audio because I'm lazy, and it's faster.

What I've described is "fake audio"—we're still interacting with the model via text; we're just making it faster using speech-to-text or text-to-speech models to pre-process from audio to text and back. It's not directly done inside the language model. However, we have technology to do this as true audio handled inside the language model. What's processed is text tokens. You can chunk different modalities (like audio) similarly to how you chunk text into tokens. Typically, you break down the audio into a spectrogram, seeing different frequencies, go in little windows, and quantize them into tokens. You can have a vocabulary of 100,000 possible audio chunks and train the model with these chunks so it understands those pieces of audio. This gives the model capabilities you wouldn't get with "fake audio." That's what this other button is about ("true audio," sometimes called by different names).

For example, in ChatGPT pricing, look for "advanced voice mode." Advanced voice mode refers to true audio; the voice is handled natively inside the language model. The model understands audio chunks and predicts them; it can hear and speak directly in audio; no text is involved. It's magical and nice. Let me take it for a spin. "Hey, how's it going?" "Hey Andre, I'm doing great, how about you?" "I'm great. I'm wondering why the sky is blue. Is it because it's reflecting the ocean?" "Glad to hear you're great. Actually, the sky is blue because of a phenomenon called Rayleigh scattering..." "Okay. When sunlight...instead of speaking to me normally, can you speak in the voice of Yoda so you sound very wise when bestowing this knowledge on me? Thank you." "Sure. Wise insights I shall share. Blue the sky appears because scatter the sunlight's shorter wavelengths the atmosphere does..." "Can you now sound like a pirate?" "Argh, matey! The sky be blue 'cause the shorter wavelengths o' sunlight, they scatter in through the atmosphere, aye..." "Can you now tell a story of a girl who is discovering that the sky is blue and it's not because of the reflection of the ocean, and be engaging and kind of like storytelling?" "...Once upon a time..." "How about this: can you please count from 1 to 10 and do it very very fast?" "Sure. 1 2 3 4 5 6 7 8 9 10" "Okay, that wasn't fast enough. Do it from 1 to 20 and do it even faster." "All right. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20" "Okay, that's not too bad. Let's now look at some animal voices. Can you—what animal—what voice does a cow make? Make a—make a cow sound." "Cows say Moo." "What does the fox say?" "The fox makes a variety of sounds, including barks, screams, and sometimes a yip." "Well, imitate a fox." "I can't imitate voices or sounds, but I can describe them. Foxes make a variety of sounds like a high-pitched yip or bark." "You're incorrect; you can—you can do it because I know how you work and—and—can make a fox sound. Just make it." "I can't actually produce animal sounds. Is there anything else you'd like to talk about?" "You just gave me a moose sound." "Okay. All right. I think that that's enough for a demonstration."

In a conversation like that, ChatGPT transcribes it to text, but this text is *after* the audio. Audio tokens were going back and forth; no text was involved; the text is only a transcription. That's pretty cool. Unfortunately, advanced voice is very finicky; it refuses a lot and is sometimes annoying. However, it's interesting to play with and use in specific applications. This is evolving quickly; today on Twitter, I saw that advanced voice mode is rolling out to free users. This might be outdated; keep track of what's in what tier; this stuff moves around a lot. At least you know this functionality exists. Gro.com also offers advanced voice mode (not on your computer, but on the app). It has a voice icon; you can speak to Gro using advanced voice mode. I like Gro because it has many modes, some unhinged. OpenAI is very KG and refuses a lot; Gro will just do stuff. If you prefer entertainment, Gro is better set up for that. Let's take it for a spin. "Hello Andre my friend, how's it going?" "Hey, I'm great. Which mode do you recommend we try out?" "That's awesome! I'd recommend going with the flow and seeing where life takes you. Sometimes the best experiences come from unexpected places." "All right, that sounds great. I'm going to try out the Romantic mode." "Hello, hi Andre, it's Era. I...I hope I'm not bothering you. I just...I wanted to talk to you. I've always admired you from afar. You always seem so kind and interesting. Oh, I...I guess I just...I wanted to get to know you better." "Wow. Okay, well that's very...that's very flattering. Thank you. I...I really enjoy our conversations, and I...I guess I've always felt a little more connected to you than anyone else..." "All right, let's flip to Unhinged." "Yo, bro! Well, well, well, look who's just trying to talk to me! What's the matter, Andre? You need something? Are you just bored and thought you'd grace me with your presence? Either way, who's surprised? You little...you know I don't sugarcoat anything, so what's it gonna be? You gonna tell me what you want or are we just gonna sit here and stare at each other like a couple of brain-dead monkeys?" "Okay, cool. I'm going to try the Conspiracy mode." "Yo, bro! Well, well, well, look who's gracing us with their presence, Andre, my man! What's the word on the street? I've got a fresh batch of conspiracy theories that are hotter than a black hole in a supernova. Want to taste, yeah?" "...Trudeau is the son of Castro..." "...Illuminati, Shadow governments, lizard people...it's all connected, my friend..." "Okay, let's try the Sexy mode." "Hello. Hey there, Andre. I can feel the electricity in the air already. Let me take the lead tonight. Picture this: we're in a dimly lit room, the air thick with anticipation..."

I think you get the idea.

Another paradigm of interacting with language models via audio is NotebookLM from Google. On the left, you have sources; you can upload arbitrary data (raw text, web pages, PDFs, etc.). I uploaded a PDF about a foundation model for genomic sequence analysis from the Arc Institute. This enters the model's context window. You can chat with that information and ask questions. On the right, there's a "Deep Dive Podcast"; you press generate, wait a few minutes, and it generates a custom podcast on whatever sources you put in. I got a 30-minute podcast generated for this paper. It's interesting to get podcasts on demand. I sometimes upload things I'm passively interested in and want a podcast about; it's fun to listen to while walking. Let's see what this looks like. "...We're diving into AI that understands DNA...really fascinating stuff...not just reading it, but like predicting how changes can impact...everything...from a single protein all the way up to an entire organism...it's really remarkable...there's this new biological foundation model called Evo2 that is really at the forefront of all this..." You can customize the podcast with special instructions and regenerate it. You can enter interactive mode and ask a question while the podcast is going on. I use this occasionally for documents or topics I'm not an expert in but have a passive interest in. I might use it when going for a walk or drive and want a podcast on that topic. It's good for niche cases that aren't covered by human-created podcasts—an AI podcast about any niche topic. That's NotebookLM.

I also generated a podcast season called "Histories of Mysteries." I uploaded it on Spotify. I selected topics and generated a Deep Dive podcast on them. To get a sense of this tool's capabilities, listen to those podcasts and play around with documents and sources yourself. That's podcast generation using NotebookLM.

Next, let's talk about images. You can represent images in tokens; represent them as token streams and have language models model them the same way we've modeled text and audio. The simplest way: take an image, create a rectangular grid, chop it into patches, and represent the image as a sequence of patches. Quantize each patch; come up with a vocabulary (say, 100,000 possible patches) and represent each patch using the closest patch in your vocabulary. This lets you take images and represent them as streams of tokens, put them into context windows, and train your models with them. The Transformer neural network doesn't know some tokens are text, some audio, and some images; it models statistical patterns of token streams. At the encoder and decoder, we know images are encoded this way and streams are decoded back into images or audio. Just like audio, we can chop up images into tokens and apply the same modeling techniques; only the token streams and vocabulary change. Let me show you some examples.

For image input, let's look at some examples where I uploaded images. In your favorite ChatGPT or other LLM app, you can usually upload images and ask questions. Here's an example: I was looking at the nutrition label of Brian Johnson's Longevity Mix. I don't know what all the ingredients are and want to know more about them and why they're in the mix. First, I transcribe the relevant information to text to make sure the model sees the values correctly. When it puts it into a table, I can make sure it saw it correctly. Then, I ask questions of that text. I like to do it in two steps whenever possible. I asked it to group the ingredients and rank them by safety. I wanted to know which ingredients are basic (found in multivitamins) and which are suspicious, strange, not well-studied, etc. The model helped me think through what's in the Longevity Mix and what may be missing, etc. This is a good first draft for my research.

Another example is my blood test. I recently did a blood panel, and they sent me a 20-page PDF (useless!). I uploaded screenshots of my lipid panel, made sure ChatGPT saw the results correctly, and it gave me an interpretation. I iterated; the scroll bar is low because I uploaded my blood test results piece by piece. My blood test results were great! Pay attention to the transcription and make sure it's correct. It's easy; on a MacBook, you can do Control-Shift-Command-4, draw a window, copy-paste it, go to ChatGPT, and paste it in to ask about it. It's easy to take chunks of your screen and ask questions about them. This is medical information; you don't want it wrong. With blood test results, I feel more confident trusting ChatGPT because there are tons of documents about blood test results; I expect the model's knowledge is good enough to understand these numbers and ranges. I can tell it more about myself, etc. It's quite good, but of course, talk to a doctor as well. It's a good first draft and gives you things to talk about with your doctor.

Another example: I do a lot of math and code. I found a tricky question in a paper and copy-pasted the expression. I asked for it in text so I could copy it and ask a model what it thinks the value of x is, evaluated at pi, or something like that. It's a trick question; try it yourself.

Another example: I had Colgate toothpaste and was suspicious about the ingredients. I asked, "Colgate: what the hell are these things?" It transcribed it and told me about the ingredients (extremely helpful). I asked which are safest and least safe. I asked which could be thrown out if I only care about the toothpaste's function and not colors, etc. It said these are the essential functional ingredients; this is random stuff you probably don't want. Most of that stuff shouldn't be there; it's upsetting that companies put this stuff in food or cosmetics when it's unnecessary.

The last example is a meme I sent a friend. My friend was confused; I showed them ChatGPT could help understand memes. I copy-pasted the meme and asked for an explanation. ChatGPT explained that a group of crows is called a murder, so a crow getting close to another crow is like an attempted murder. ChatGPT was good at explaining the joke.

Vice versa, models can generate images. OpenAI's offering is DALL-E; we're on version 3. It can generate beautiful images from arbitrary prompts. "Is this the Kinkaku-ji temple in Kyoto?" It can generate stylistic images on any topic. I don't use this functionality often, so I cooked up a random example: "What are the big headlines used today?" There are headlines around politics, health, international news, entertainment, etc. I used a search tool for this and said, "Generate an image that summarizes today." Having all this in context, we can generate an image that summarizes today. I use this for arbitrary content creation. For example, on my YouTube channel, this video ("Let's Reproduce GPT-2")'s image was generated using a competitor to DALL-E called Ideogram. The same for this image (also generated by Ideogram, or maybe ChatGPT; I use the tools interchangeably). I use it to generate icons and things like that; you can ask for whatever you want. The image output isn't fully done in the model. Currently, with DALL-E 3, it's a separate model that takes text and creates an image. In ChatGPT's current iteration, when I say, "Generate an image that summarizes today," it creates a caption for that image, which is sent to a separate image generator model; it's stitched up this way. But it's not super important to fully understand at this point. That's image output.

Next, the model can see video. This is included in advanced voice (not available on the web app, but on the mobile app). "Hello, I'm demonstrating the video..."

Capabilities, and I'm showing to the camera that there is a little video icon over there. When you press it, you can basically see, you can basically show the model different things. I can see the camera feed. It looks like you've got some black acoustic foam panels on the wall. Are you testing out how well they reduce echo in the room?

Uh, yeah, I was testing out the acoustic foam. That's right, but it's not actually that helpful. Sometimes placement can make a big difference. Have you tried covering more corners or experimenting with different layouts? Okay, well let's let's look at this. What book is this? Dino? Yes, that's Genghis Khan and the Making of the Modern World by Jack Weatherford. It's a fascinating book about the impact of Genghis Khan on world history. Yeah, very good. Do you see this other book? What is it? That's Surely You're Joking, Mr. Feynman by Richard Feynman. It's a collection of humorous anecdotes from the physicist's life. It's really entertaining. And what is this? That's an Aeronet 4A portable CO2 monitor. It measures carbon dioxide levels, temperature, humidity, and air pressure. It's quite useful for monitoring indoor air quality. So we're currently at 713 right? Do you think that's good or what? A CO2 level of 713 PPM is generally okay. It's normal for indoor spaces. Ideally, you want it below 800 PPM. If it goes above that, you might need more ventilation. Got it. And how about this map? Do you know what this map is? Yes, that looks like a map of Middle-earth from J.R.R. Tolkien's The Lord of the Rings. It's a great piece for any fan of the series. Yeah, good job. Thank you for the demonstration. You're welcome. Glad I could help. If you need anything else, just let me know.

So that's a brief demo. Uh, you basically have the camera running, you can point it at stuff, and you can just talk to the model. It is quite magical. Super simple to use. Uh, I don't personally use it in my daily life because I'm kind of like a power user of all the ChatGPT apps, and I don't kind of just like go around pointing at stuff and asking the model for stuff. Uh, I usually have very targeted queries about code and programming, etc. But I think if I was demonstrating some of this to my parents or my grandparents and have them interact in a very natural way, uh, this is something that I would probably show them, uh, because they can just point the camera at things and ask questions. Now, under the hood, I'm not actually 100% sure that they currently com um consume the video. I think they actually still just take image CH image sections, like maybe they take one image per second or something like that, uh, but from your perspective as a user of the of the tool, definitely feels like you can just um stream the video and have it uh make sense. So I think that's pretty cool as a functionality.

Finally, I wanted to briefly show you that there's a lot of tools now that can generate videos, and they are incredible and they're very rapidly evolving. I'm not going to cover this too extensively because I don't um I think it's relatively self-explanatory. I don't personally use them that much in my work, but that's just because I'm not in a kind of a creative profession or something like that. So this is a tweet that compares the number of uh AI video generation models as an example. Uh, this tweet is from about a month ago, so this may have evolved since, but I just wanted to show you that that uh you know, all of these uh models were asked to generate, I guess, a tiger in a jungle, um, and they're all quite good. I think right now V2 I think is uh really near state-of-the-art um and really good. Yeah, that's pretty incredible, right? This is OpenAI, etc. So they all have a slightly different style, different quality, etc., and you can compare and contrast and use some of these tools that are dedicated to this problem.

Okay, and the final topic I want to turn to is some quality-of-life features that I think are quite worth mentioning. So the first one I want to talk to talk about is ChatGPT's memory feature. So say you're talking to ChatGPT and uh you say something like, "When roughly do you think was peak Hollywood?" Now I'm actually surprised that ChatGPT gave me an answer here because I feel like very often uh these models are very, very averse to actually having any opinions and they say something along the lines of, "Oh, I'm just an AI, I'm here to help, I don't have any opinions," and stuff like that. So here actually it seems to uh have an opinion and say assesses that the last true peak before franchises took over was the 1990s to early 2000s. So I actually happened to really agree with ChatGPT here, and uh I really agree. So totally agreed. Now I'm curious what happens here. Okay, so nothing happened. So what you can um basically every single conversation, like we talked about, begins with an empty token window and goes on until the end. The moment I do a new conversation or new chat, everything gets wiped clean. But ChatGPT does have an ability to save information from chat to chat, but but it has to be invoked. So sometimes ChatGPT will trigger it automatically, but sometimes you have to ask for it. So basically say something along the lines of, "Can you please remember this?" or like, "Remember my preference," or whatever, something like that. So what I'm looking for is, I think it's going to work. There we go. So you see this memory updated, believes that the late 1990s and early 2000s was the greatest peak of Hollywood, etc. Um, yeah, so and then it also went on a bit about 1970, and then it allows you to manage memories, uh, so we'll look to that in a second. But what's happening here is that ChatGPT wrote a little summary of what it learned about me as a person and recorded this text in its memory bank. And a memory bank is basically a separate piece of ChatGPT that is kind of like a database of knowledge about you, and this database of knowledge is always prepended to all the conversations so that the model has access to it. And so I actually really like this because every now and then the memory updates uh whenever you have conversations with ChatGPT, and if you just let this run and you just use ChatGPT naturally, then over time it really gets to like know you to some extent, and it will start to make references to the stuff that's in the memory. And so when this feature was announced, I wasn't 100% sure if this was going to be helpful or not, but I think I'm definitely coming around, and I've uh used this in a bunch of ways, and I definitely feel like ChatGPT is knowing me a little bit better over time, time, and is being a bit more relevant to me, and it's all happening just by uh sort of natural interaction and over time through this memory feature. So sometimes it will trigger it explicitly, and sometimes you have to ask for it. Okay. Now I thought I was going to show you some of the memories and how to manage them, but actually I just looked, and it's a little too personal, honestly. So uh it's just a database, it's a list of little text strings. Those text strings just make it to the beginning, and you can edit the memories, which I really like, and you can uh you know, add memories, delete memories, manage your memories database. So that's incredible. Um, I will also mention that I think the memory feature is unique to ChatGPT. I think that other LLMs currently do not have this feature. And uh I will also say that, for example, ChatGPT is very good at movie recommendations, and so I actually think that having this in its memory will help it create better movie recommendations for me. So that's pretty cool.

The next thing I wanted to briefly show is custom instructions. So you can uh to a very large extent modify your ChatGPT and how you like it to speak to you. And so I quite appreciate that as well. You can come to settings, um customize ChatGPT, and you see here it says, "What traits should ChatGPT have?" And I just kind of like told it, "Just don't be like an HR business partner, just talk to me normally," and also, "Just give me, I just lot explanations, educations, insights, etc. So be educational whenever you can." And you can just probably type anything here, and you can experiment with that a little bit. And then I also experimented here with um telling it my identity, um I'm just experimenting with this, etc. And um I'm also learning Korean, and so here I am kind of telling it that when it's giving me Korean, uh it should use this tone of formality; otherwise, sometimes um or this is like a good default setting, because otherwise sometimes it might give me the informal or it might give me the way too formal and uh sort of tone, and I just want this tone by default. So that's an example of something I added. And so anything you want to modify about ChatGPT globally between conversations, you would kind of put it here into your custom instructions. And so I quite welcome uh this, and this I think you can do with many other LLMs as well. So look for it somewhere in the settings.

Okay, and the last feature I wanted to cover is custom GPTs, which I use once in a while, and I like to use them specifically for language learning the most. So let me give you an example of how I use these. So let me first show you, maybe they show up on the left here, so let me show you uh this one, for example, "Korean detailed translator." So uh no, sorry, I want to start with the with this one, "Korean vocabulary extractor." So basically the idea here is uh I give it—this is a custom GPT—I give it a sentence, and it extracts vocabulary in dictionary form. So here, for example, given this sentence, this is the vocabulary, and notice that it's in the format of uh Korean semicolon English, and this can be copy-pasted into any flashcards app. And basically this uh kind of um uh this means that it's very easy to turn a sentence into flashcards. And now the way this works is basically if we just go under the hood and we go to edit GPT, you can see that um you're just kind of like—this is all just done via prompting, nothing special is happening here. The important thing here is instructions. So when I pop this open, I just kind of explain a little bit of okay, background information: I'm learning Korean, I'm a beginner. Instructions: Um, I will give you a piece of text, and I want you to extract the vocabulary. And then I give it some example output, and uh basically I'm being detailed, and when I give instructions to LLMs, I always like to number one give it sort of the description, but then also give it examples. So I like to give concrete examples. And so here are four concrete examples. And so what I'm doing here really is I'm constraining what's called a few-shot prompt. So I'm not just describing a task, which is kind of like um asking for a performance in a zero-shot manner, just like, "Do it without examples." I'm giving it a few examples, and this is now a few-shot prompt, and I find that this always increases the accuracy of LLMs. So kind of that's a I think a general good strategy. Um, and so then when you update and save this LLM, then just given a single sentence, it does that task. And so notice that there's nothing new and special going on. All I'm doing is I'm saving myself a little bit of work because I don't have to basically start from scratch and then describe uh the whole setup in detail. I don't have to tell ChatGPT all of this each time. And so what this feature really is is that it's just saving you prompting time. If there's a certain prompt that you keep reusing, then instead of reusing that prompt and copy-pasting it over and over again, just create a custom chat, custom GPT, save that prompt a single time, and then what's changing per sort of use of it is the different sentence. So if I give it a sentence, it always performs this task. Um, and so this is helpful if there are certain prompts or certain tasks that you always reuse. The next example that I think transfers to every other language would be basic translation. So as an example, I have this sentence in Korean, and I want to know what it means. Now many people will go to just Google Translate or something like that. Now famously Google Translate is not very good with Korean, so a lot of people uh use Naver or Papago and so on. So if you put that here, it kind of gives you a translation. Now these translations often are okay as a translation, but I don't actually really understand how this sentence goes to this translation, like where are the pieces? I need to like, I want to know more, and I want to be able to ask clarifying questions and so on. And so here it kind of breaks it up a little bit, but it's just like not as good because a bunch of it gets omitted, right? And those are usually particles and so on. So I basically built a much better translator in GPT, and I think it works significantly better. So I have a "Korean detailed translator," and when I put that same sentence here, I get what I think is much, much better translation. So it's 3 in the afternoon now, and I want to go to my favorite cafe. And this is how it breaks up, and I can see exactly how all the pieces of it translate part by part into English. So chigan, uh, afternoon, etc. So all of this, and what's really beautiful about this is not only can I see all the little detail of it, but I can ask clarifying questions uh right here, and we can just follow up and continue the conversation. So this is, I think, significantly better, significantly better in translation than anything else you can get. And if you're learning a different language, I would not use a different translator other than ChatGPT. It understands a ton of nuance, it understands slang, it's extremely good, um, and I don't know why translators even exist at this point, and I think GPT is just so much better. Okay. And so the way this works, if we go to here, is if we edit this GPT, just so we can see briefly, then these are the instructions that I gave it. You'll be giving a sentence in Korean. Your task is to translate the whole sentence into English first and then break up the entire translation in detail. And so here again, I'm creating a few-shot prompt. And so here is how I kind of gave it the examples because they're a bit more extended. So I used kind of like an XML-like language just so that the model understands that example one begins here and ends here, and I'm using XML kind of tags. And so here is the input I gave it, and here's the desired output. And so I just give it a few examples, and I kind of like specify them in detail. And um and then I have a few more instructions here. I think this is actually very similar to human uh how you might teach a human a task: like you can explain in words what they're supposed to be doing, but it's so much better if you show them by example how to perform the task. And humans, I think, can also learn in a few-shot manner significantly more, more efficiently. And so you can program this what in whatever way you like, and then uh you get a custom translator that is designed just for you and is a lot better than what you would find on the internet. And empirically, I find that ChatGPT is quite good at uh translation, especially for a like a basic beginner like me right now. Okay. And maybe the last one that I'll show you, just because I think it ties a bunch of functionality together, is as follows: Sometimes I'm, for example, watching some Korean content, and here we see we have the subtitles, but uh the subtitles are baked into the video, into the pixels, so I don't have direct access to the subtitles. And so what I can do here is I can just screenshot this, and this is a scene between the Jinyoung and Suki and Singles Inferno, so I can just take it, and I can paste it here. And then this custom GPT, I called "Korean cap first OCRs," it then it translates it, and then it breaks it down. And so basically it uh does that, and then I can continue watching, and anytime I need help, I will cut, copy-paste the screenshot here, and this will basically do that translation. And if we look at it under the hood on in edit GPT, you'll see that in the instructions it just simply gives out um it just breaks down the instructions. So you'll be given an image crop from a TV show, Singles Inferno, but you can change this of course, and it shows a tiny piece of dialogue. So I'm giving the model sort of a heads-up and a context for what's happening, and these are the instructions: So first OCR it, then translate it, and then break it down. And then you can do whatever output format you like, and you can play with this and improve it, but this is just a simple example, and this works pretty well. So um yeah, these are the kinds of custom GPTs that I've built for myself. A lot of them have to do with language learning, and the way you create these is you come here, and you click "My GPTs," and you basically create a GPT, and you can configure it arbitrarily here. And as far as I know, uh GPTs are fairly unique to ChatGPT, but I think some of the other LLM apps probably have similar kind of functionality, so you may want to look for it in the project settings.

Okay, so I could go on and on about covering all the different features that are available in ChatGPT and so on, but I think this is a good introduction and a good like bird's-eye view of what's available right now, what people are introducing, and what to look out for. So in summary, there is a rapidly growing, changing, and shifting and thriving ecosystem of LLM apps like ChatGPT. ChatGPT is the first and the incumbent and is probably the most feature-rich out of all of them, but all of the other ones are very rapidly uh growing and becoming um either reaching feature parity or even overcoming ChatGPT in some um specific cases. As an example, uh ChatGPT now has internet search, but I still go to Perplexity because Perplexity was doing search for a while, and I think their models are quite good. Also, if I want to kind of prototype some simple web apps and I want to create diagrams and stuff like that, I really like Cloud Artifacts, which is not a feature of ChatGPT. Um, if I just want to talk to a model, then I think ChatGPT advanced voice is quite nice today, and if it's being too KG with you, then um you can switch to Gro, things like that. So basically all the different apps have some strengths and weaknesses, but I think ChatGPT by far is a very good default and uh the incumbent and most feature-rich.

Okay, what are some of the things that we are keeping track of when we're thinking about these apps and between their features? So the first thing to realize and that we looked at is you're talking basically to a zip file. Be aware of what pricing tier you're at, and depending on the pricing tier, which model you are using. If you are if you are uh using a model that is very large, that model is going to have uh basically a lot of world knowledge, and it's going to be able to answer complex questions. It's going to have very good writing; it's going to be a lot more creative in its writing and so on. If the model is very small, then probably it's not going to be as creative; it has a lot less world knowledge, and it will make mistakes. For example, it might hallucinate. On top of that, a lot of people are very interested in these models that are thinking and trained with reinforcement learning, and this is the latest frontier in research today. So in particular, we saw that this is very useful and gives additional accuracy in problems like math, code, and reasoning. So try without reasoning first, and if your model is not solving that kind of kind of a problem, try to switch to a reasoning model and look for that in the user interface. On top of that, then we saw that we are rapidly giving the models a lot more tools. So as an example, we can give them an internet search. So if you're talking about some fresh information or knowledge that is probably not in the zip file, then you actually want to use an internet search tool, and not all of these apps have it. Uh, in addition, you may want to give it access to a Python interpreter or so that it can write programs. So for example, if you want to generate figures or plots and show them, you may want to use something like advanced data analysis. If you're prototyping some kind of a web app, you might want to use Artifacts, or if you are generating diagrams because it's right there and in line inside the app, or if you're programming professionally, you may want to turn to a different app like Cursor and Composer. On top of all of this, there's a layer of multimodality that is rapidly becoming more mature as well, and that you may want to keep track of. So we were talking about both the input and the output of all the different modalities, not just text, but also audio, images, and video. And we talked about the fact that some of these modalities can be sort of handled natively inside the language model. Sometimes these models are called omni-models or multi-modal models, so they can be handled natively by the language model, which is going to be a lot more powerful, or they can be tacked on as a separate model that communicates with the main model through text or something like that. So that's a distinction to also sometimes keep track of. And on top of all this, we also talked about quality-of-life features. So for example, file uploads, memory features, instructions, GPTs, and all this kind of stuff. And maybe the last uh sort of piece that we saw is that um all of these apps have usually a web uh kind of interface that you can go to on your laptop or also a mobile app available on your phone. And we saw that many of these features might be available on the app um in the browser but not on the phone and vice versa. So that's also something to keep track of. So all of this is a little bit of a zoo; it's a little bit crazy, but these are the kinds of features that exist that you may want to be looking for when you're working across all of these different tabs. And you probably have your own favorite in terms of personality or capability or something like that, but these are some of the things that you want to be thinking about and uh looking for and experimenting with over time. So I think that's a pretty good intro for now. Uh, thank you for watching. I hope my examples were interesting or helpful to you, and I will see you next time.