In the last video, we looked at writing tasks where you would specify a prompt to the last large language model, and have it generate a comparatively longer output than the input prompt. It turns out, is also useful for many reading tasks, and by that I mean tasks where you would input a prompt, and then have it generate usually a similar length or often shorter output than the input prompt. Let's take a look at some reading tasks starting with something that I use myself all the time, which is proof reading. Many times if I'm writing a piece of text, I will read through it carefully three or four times myself for spelling and grammatical errors, and even though I thought I prove read it carefully myself, a line model will still find errors in it that somehow I had miss. Here's an example of a prompts that you could try. Proofread the following text, and I find that if you tell it what you want the text before. Here's text intended for a website selling children's stuffed toys, and sometimes I ask you to check for spelling and grammatical errors as well as awkward sentences, and then have it rewrite it with corrections. This is a piece of text with some errors, and the output of the large language model fixes, snuggle was misspelled and it fixes this little piece of grammar over here. When I'm writing text myself, that I want to be quite confident it's free of spelling and grammatical errors, and sometimes also awkward sentences, I actually use this myself to proof read what I write. A second reading task that large language models are often used for, is to summarize long articles. One of my collaborators, Erik Brynjolfsson, who's a Stanford professor, once sent me an email linking to an article that he had written titled The Turing Trap. I knew it was a good article, but it was a very long article and I didn't have time to read the whole thing before I responded to his email. I actually use the following prompt and copy pasted his entire article into an web interface of a large language model, and had it quickly generate a summary for me. It turns out this paper that he had written talks about how human-like AI offers benefits, but there's a lot to be done by having AI augment humans rather than automate. But the point of Brynjolfsson's article on the Turing Trap was he was advocating that instead of having AI automate or replace human work, we should put more effort into having AI complement or to augment human work. With a large language model summarizing this long article, I was able to get back on this faster than if I had to read the entire article myself. By the way, this is a good article. Eventually I did read the entire article myself, and I really enjoyed it, but today, I do sometimes use large language models to summarize for me things that I don't have time to read in its entirety. This is a use case that you could go to one of the web interfaces of a large language model and use relatively quickly yourself. Now it turns out there's a software application version of this too that is taking off in businesses. Let me illustrate this with an example. Say you're a manager of a customer service call center where you have many customer service agents, like this person shown on the left with the microphone, having phone calls with customers, like this person shown on the right. If you have permission to record these phone calls between the agents and the customers, you can then run the phone calls through a speech recognition system to get a text transcript of the conversation. If you have many customer service agents having conversations, you end up with a lot of text transcripts. If you want to review what's going on in your call center, you probably end up with too much texts to read. Given a text transcript like this between the customer and an agent, what really happened in this call? One use of large language models would be to have it summarize this entire conversation and generate a short summary. Like MP401-27KX was reported as broken, and so on. If you were to take all of these different text transcripts and have a software application to generate the summaries, then you as a manager of this can take a quick look at all of these summaries and just spot if there are any issues, or any trends that you want to be aware of. A system like this would be implemented as a software application that uses a large language model, because it doesn't really make sense for you or anyone else to copy paste these conversations one at a time into the website of a large language model provider. In terms of customer service interactions, large language models are also used for customer email analysis. In an earlier video, you saw the example of taking a customer email, and deciding if it's a complaint, in this case, no, as well as what department to route this email. This will be another software application that uses a large language model. Let's take a deeper look at how one could build this application, focusing on the part of deciding what department to route this email. One thing you could do is write a prompt to tell the LLM to read the email and decide which department to route it to. You can specify the task and provide the email. But it turns out that with a prompt like this, you may find that the algorithm routes it to the complaints department in this case, which may or may not be a department that exists in your organization. This would be an example of where the LLM has been given insufficient context to know what are the names of the actual departments in your company that it should choose from. In contrast, if you were to update the prompt as follows, and say, read the email, choose the most appropriate department to rouse it to, and choose department only from the following list. In this case, given the set of choices you wanted to choose from, routes it to the apparel department correctly. The process of building an application using a large language model is again, not a lot uncommon. To write a prompt that doesn't quite work right the first time, and when you find it, route it to a nonexistent complaint department, then just update the prompt and that fixes the problem. One last application that I want to touch on is reputation monitoring, where you can use an LLM to build a dashboard to track your customer sentiments, positive or negative of your business or your products over time. For example, if you run a restaurant and occasionally your customers, write online reviews or send your emails describing their experience, you can then use a prompt like this, read the following review and classified as having the positive, negative sentiment, to have it decide automatically if each review was positive or negative. In this case, if the food was amazing or service a friendly, that would be classified as having a positive sentiment. Then by having software count the number of positive reviews per day, as well as the number of negative reviews per day, you can build a dashboard that tracks per day, all the time, how the sentiments are trending. Looks like the customer sentiment is pretty positive, but if ever it starts trending negative like this, with more negative reviews, then this dashboard can alert you to that maybe something's happening that we should pay attention to, and see if there's something we need to fix at the restaurant. In this video, we looked at a number of reading applications, including proof reading, summarization, email routing, restaurant review, sentiment analysis. If you can think of a task where you wish you had someone that could read a piece of text, and just say a few things or give a few quick indications of what was in that piece of text, that could be a good candidate for a reading task to get an LLM to do for you. Next, let's go onto the next video to take a look at chatting task.