Finally, an AI agent that actually works

Key Takeaways

this AI agent is able to access all my email inbox in response to people with my own writing styles and it can even review GitHub PR on my behalf if you are also a big fan of TV show Silicon Valley you might remember this thing so international access is you correct he's been messaging with it all day and he has no idea what the [ __ ] is going on because I just texted you and you were um we've been texting back and forth all day was that not you complicated question is your Shadow you is that a bot yes so I spent all day chatting with the black yes you made it AI gilfoyle yes and I was talking to it do you need the real me for this conversation people and manipulate them because you don't have the decency to want to interact with a co-worker so the concept of having a eyeshadowing you and talk to your colleague and friends is actually possible now ai agent has been the hottest topic this year it has developed so much past few months from baby AGI Auto gbt agent GPT and small AI what is Agent fundamentally agent is basically a combination of large language model plus memory plus planning skills plus tools so their large Lounge model can actually prioritize tasks that needs to be done use the right tools to execute those tasks and then decide what's the next best action to take with all the information collected however at the moment all those AI agents still feels like baby steps it has really high error rate in terms of the task to execute and also the tools you can use is quite limited most of them are just browsing their internet however this one AI platform that really open up the imagination of what the AI agent for personal assistant could look like so this new AI agent I want to talk about today is called hyperwrite it is Chrome plugin that Reagan more than 100K users it is initially an AI writing companion like Jasper but what they just introduced is a new feature called AI assistant it is basically a auto GPT that has access to your whole browser so you can ask it to book a flight which would go to website and book a ticket also have access to Dom nodes or LinkedIn and do things that just like you would do so the tool selection has been expanded massively and this tool currently is in Alpha 0.01 but it's already really impressive results once you test it out to use this ai's assistant you can go to Google Chrome plugin which I have a link below and install this plugin and once it's finished you can go to a website that you wanted to do task on and there in this few use case I want to share that really opened up my imaginations the first use case is emails and as you can see I'm pretty terrible at managing my inbox I can actually ask it to read and respond to all my unread emails so to use that once you are in Gmail you should see a little icon here and once you hover it it will give you a list of options so we will click on this personal assistant so what I wanted to do is respond to all my unread emails so our typing here respond to all my emails and click Start action once you click on that you will see a new window pop-up and on the right side it is showing it is working okay you can see the agent is thinking the current page is Gmail inbox and there are several on your email and actually open one unread email I have and start scroll down to this respond bar click on respond and typing a response and actually send out email on behalf of me and then it go back to the inbox page try to find the next unread email if you look at the syncing of Farm agents it is able to tell like one email is from Google photo it is automated email so it didn't require any response at the moment there's still some arrows like it will open the same email multiple times before it'll move on to the next one but as the accuracy improved I can imagine this became super useful and I wonder whether I can push this even further I will ask it to read all my unread emails and if it is personal email Write a response otherwise archive it and let me start this action so as you can see it is able to tell both the Amazon email and the other one are Promotional and just start archiving it automatically it is already doing a great job reviewed in archive five to six different emails for me it's not that useful for this email because 99 of them is promotional but if it is business email it can draft a response and flag the important message for you to review later and what's really cool about this is it is actually able to read all your Gmail data so it can start learning your writing styles so the next use case I want to share is lease generation in LinkedIn one leads generation strategy I know people do is they will try to find posts from potential customers and comment under their posts to warm up the connection so when they reach out the customer already have some impression or connection with this person and I think that's one use case that this personal assistant can potentially do what I would do is I will give it a task find a list of posts about generative Ai and then leave a comment to each post and let's get it running so you first they start searching generative AI in LinkedIn and then it switch to post tab so it can find all the posts and then open up the command box automatically and wrote a comment based on the content on behalf of me oh [ __ ] it actually posted the comment as well so I'm going to stop it before I spam a lot of people on LinkedIn but you can see how this is playing out and you can basically search for a lot of different posts and future high quality one and comment below and the other use case I want to share is getting this AI agent to review pull requests for me if you don't know what a pull request is especially if two people or a team working on the same code base they will need to review each other's work before it merge together and it's pretty tedious process that I will need to go to each pull request look at the file change making sure there's no arrows and if there is all good I will click on merge if it's not I will write a comment and give feedback to them and what I did here is I actually made a typo on purpose for one PR I'm submitting so I will remove the E from the message which created typo here and I just created pull request for this change I'm pretty curious to see whether AI agent can actually do this pull request for me okay so I will give you an instruction here help me review the pr approve if of good leave a comment if it needs improvements and start okay so it successfully go to the pull request page as it knows PR represents for pull requests and then it opens a PR that ready for review let's say whether it's able to go to file change or commits page to compare the change okay great so it does go to files change and let's see if it's able to spot the arrow oh cool so it actually automatically left a comments here it seems there's typo in the variable name and they click on the star review so now it successfully completes this task and the author will get notifications as here's a comment this is really impressive and this seemed to be a task that this AI agent is really good at and has pretty high success rate all right and the next use case I want to test out is lenses AI agent to write blog posts for me and automatically publish it because it has access to my webflow account and this is my personal website so I'll go to the blog post page to make sure there's the last room for Arrow and I will open this personal assistant giving instruction write and publish a blog post about problems of generative AI at least 200 words and fill in with both excerpt and body let's try alright so you can see that successfully create a new blog post page a viewing the title if you're in the description here and then start writing the actual blog post body so blog post here is still pretty short and I guess that's because of instruction I give it is 200 words if I give more it probably will input more stuff there as well once they've publish blog posts it go back to home page or webro and then looking for the publish button okay great so it's successfully published the blog posts and what's more impressive is it actually builds my website and try to click on that blog post page to kind of to validate if the task actually completed this is a step that I didn't expect the validation step is actually really cool all right so this is really impressive I don't think I would use this for actual quality blog post well this will be really good is writing tweets I'm pretty curious to see whether I can use this to write tweets every three hours about quotes from famous people so those are the use cases that really eye-opening for me but there are also a lot of stuff that this AI agent is not great at for example I asked it to find latest news about generative AI purchase and create a blog posts in the Google Doc and what I found is it can actually successfully do the research find the random in use but it's not great at using Google Docs so you can see it was successfully putting in the title of Google doc but it's unable to find where should I type the body of content and same case for Google sheet and a few other platform too however I'm still super excited about this because as it improves the quality of task execution the use case of personal assistant can start opening up and in the end I also want to share a mental model about agents that's weeks shared so in a launching agent webinar swix mentioned about this mental model currently when we talk about building agents A lot of people are trying to build level 5 fully autonomous self-driving cars but not enough people are actually focusing on level 2 or level 3 agents where the AI doing some specific tasks extremely well and let human to steer the direction or instruction next steps and I think this is really brilliant because when I think about all those agents products we tried in general it fell into two parts one is a planning of tasks quite often large analogy model is still not great in terms of deciding what's the best next step to do second is the quality of task execution because it's tried to get GPD to do a lot of stuff but the quality of task Fusion is not great so I'm personally very excited to see what are kind of level two and level three agents people start building for example I can say a word where the agent is specialized in certain tasks like try to find good information from internet exhaustively and human will decide what kind of next topic it should try to research further without spending hours and hours scrapping the internet and I think more specialized agents like this in build will eventually allow us to move towards a level five type of fully autonomous agents so this is today's video and I definitely highly recommend you try out this AI assistant and see what kind of interesting use case it's capable to do and I will post a few videos in terms of exploring what type of interesting and practical agents we can build so if you enjoy this video please subscribe and like and I see you next time