The AI Daily Brief: Artificial Intelligence News and Analysis - How to Get AutoGPT on Your Phone (And What People Are Actually Finding it Useful For)

Episode Date: May 9, 2023

A check-in on AutoGPT as well as the headline news: Updates from the US copyright office Pitchbook VC enthusiasm for AI Palantir stock price pops after AI announcement Google I/O Developer conference AI leaks IBM relaunches AI division as WatsonX in partnership with HuggingFace One researcher says 80% of jobs could be replaced by AI

On this episode of the AI breakdown, we are talking all about Auto GPD, a new tool for getting it on your iPhone, plus what people are finding it useful now, six weeks after it came onto the market and when some of the hype has died down. Before that, on the AI breakdown brief, we cover new updates from Google and IBM on AI strategy, as well as a bold prediction that 80% of jobs will be replaced by AI. Welcome back to the AI breakdown AI brief, the headline AI news you need in five minutes or less. We start today with updates from the U.S. Copyright Office. Obviously, AI has been a copyright challenge, and the Copyright Office has updated their guidance. Basically, they wanted to clarify that when humans create works with AI, it can still be copyrighted as long as there is sufficient human creation involved, such as bringing together lots and lots of AI created inputs into a larger hole. What can't be copyrighted right now are things that are entirely created by machines. Now, obviously, copyright is going to be one of the big issues as we figure out how to integrate AI with existing societal norms. Grimes continues to be on the front of this. I saw this quote from her yesterday where she said, I'm actually kind of stressed that people are starting to make competitively or maybe better quality grime-sounding songs than I do,

but it's also the most wonderfully poetic way to die and respawn in another career. If you want to hear one of those AI Grime songs, go check out the weekly recap from Saturday. It has a track from Angel Baby that uses Grimes AI at the end. Next on the AI breakdown, no surprise here, but VCs are excited about AI. This is from a pitch book survey, and they asked, in which areas of technology do you expect the most growth in adoption over the next 12 months? 36.1% said artificial intelligence, more than double the second place contender, which was climate tech at 18.1%.

Now, in addition, over two-thirds of investors that were surveyed 70.7% said that they thought the generative AI would spawn unicorns, no surprises there. I would only caution that there is going to be a temptation for people to write off much of AI because VCs are giving it such hype, don't fall prey to it. Just because VCs are overhyped and are going to invest in things that probably don't deserve their investments doesn't mean that the overall trend isn't correct. Part of why VCs are so keen to invest is that public markets are also recognizing the power of AI. Palantir last week showed off some of its new AI features and its stock has gone up more than 20% since then. They say that demand for AI tools is unprecedented.

Speaking of which, we've been keeping track of Google's plans. Google feels quite far behind, I think especially from where people would have expected Google to be. And this week is their big I.O. developer conference. Another set of leaked documents from inside Google gave us some hints about what's coming at I.O. Including the update of their large language model, Palm, which they're calling Palm 2 and includes more than 100 languages. That will be used to power barred, and they're hoping to reclaim some

momentum at this event. Speaking of reclaiming momentum, do you guys remember Watson? That IBM early AI product that was created to beat humans on Jeopardy? Well, that business line had basically gone nowhere. In fact, 15 months ago, IBM had sold its Watson Health Unit, but now they are bringing the Watson brand back. They've created something called Watson X, and effectively it is an AI training studio for enterprises. So when enterprises and big businesses need to train their own versions of models, Watts & X is supposed to help them do that more easily. CEO Arvin Krishna said we allow an enterprise to use their own code to adapt the model to how they want to run their playbooks and their code. Then they can deploy it for themselves without any danger of their code leaking. As quick as we might be

to write off a legacy competitor like IBM, I wouldn't do so in this case. First of all, I think it's very likely that enterprises with highly sensitive and proprietary data are going to want to train their own models. We already see OpenAI starting to talk about a chat GPT offering just for businesses that's private by default, but that next level of privacy and security is just running it on premise with your own uniquely trained models. So I think that there's going to be a market for that. Second, IBM is partnering with Hugging Face, one of the most dynamic players in the space, which is the default home for open source AI models, so there's something interesting there to watch. Finally, a buzzy little prediction that I'm sure we'll get lots of headlines.

AI expert Ben Gertzel says that AI could replace 80% of jobs in the next few years. He said this at a Web Summit conference, and he said that this was actually a good thing, that humans shouldn't be spending all their time on work that was irrelevant, when machines could just do it better. But even that optimistic view still has a really strong sticking point, which is what happens in the transition, and how society is able to deal with so much disruption all at once. Neither Ben nor basically anyone that I've seen has a really good answer to that, but that's the state of the conversation.

All right, guys, that's the AI Breakdown Brief. If you found this useful, please like this video or subscribe to the channel. And of course, you can also listen to this each day at the AI Breakdown podcast. I'll see you back here in a bit for the main AI breakdown. On today's AI breakdown, we are talking about how to get AutoGPT on your phone and what AutoGPT is actually useful for, about six weeks after it exploded onto the scene. About six weeks ago, people started talking about baby AGI and Auto GPT. These were both tools that took the idea of chat GPT,

but added some really interesting features that seemed like they would create a whole new set of functionality. Now, you could tell just how interesting they were to developers, because within days, basically, they were some of the most interacted with projects on GitHub. You had all the threaders on Twitter talking about, all the use cases that had popped up within the first week. You had people like me on YouTube who were doing the same thing. But then within a couple weeks, people started to maybe get a little bit more disillusion. They started to see what AutoGPT and Baby AGI couldn't really do.

So what we're going to do today on this AI breakdown is we're going to talk about, one, a new tool that puts an AutoGPT style experience on your iPhone. And two, we're going to do a bit of an update on what people are actually finding useful with AutoGPT, including one use case that I haven't seen really anyone else talking about. In order to get started, I think that there were three things about auto-GPT that people were really excited about relative to chat GPT. The first was that it was connected to the internet. It didn't have a cutoff on when its information came from,

which means it could be used for up-to-date research and other sort of queries that involve searching. Second was the idea that effectively it prompted itself. It didn't just respond to human input, but it actually, when given a task, created an additional subtask list to figure out how to get that thing. done. The third thing that people were excited about was the idea that it would actually execute the tasks, including spinning up other AI agents to do so. Now, I think what we've found so far, in short,

is that the first and the second have really come to fruition. Connection to the internet does give auto GPT an entirely new set of use cases that chat GPT currently isn't useful for. What's more, the fact that it prompts itself and creates its own task list has really interesting implications as well. Where I think for many, AutoGPT has fallen a little bit short, is around the execution. of tasks with other AI agents. You'll often see people complain that the tools that they're using are getting caught in loops, or thinking that something has been executed when it really hasn't. Now, I will also note that part of the problem here isn't just with AutoGPT, it's with the implementations of this type of technology that people are using. Specifically, the folks who are using no-code implementations where they're just off the shelf and there are a UI that someone is prepared,

obviously have a lot less control in terms of how they shape the experience of an Auto-GPT to actually do the things that they want it to do. broadly speaking, and we'll see this throughout the rest of the video, we've seen people move from very generalized use cases to more specialized use cases and they've been having more success as they've done so. But let's talk about I-baby AGI. Developer Nate Chan has been watching Auto-GPT since the very beginning and started tinkering almost immediately. And it didn't take long after a baby AGI and Auto-GPT were released for people to start putting web interfaces on top of it. A few that came out pretty quickly and that we've demoed on this channel include God Mode, Agent GPT, and IOMNI, which represented an even, newer trend of focusing in on a specific use case, in this case, online research. But back to Baby AGI, what Nate Chan wanted to do was get this type of experience on an iPhone. Sure, it was

great to have it in a web browser, but the way that so many of us interact with the internet is through our phones. One of the testers who had early access to Ibaby AGI is Lauren Marie. On May 3rd, she wrote, for the past few days, I've been testing the brand new Ibaby AGI app by Nathan Chan. I'm so impressed with what I've been able to do with it. She says the highlights include developed a simple interactive web page with a single prompt, some assembly required, conducted an SEO analysis and received SEO suggestions for my website, created a full travel itinerary for a trip to Italy with a budget of 3,000 USD, created a fitness plan based on personal preferences,

developed a list of tools to automate sales prospecting, made a newsletter outline with AI trends related to chat GPT4 and autonomous agents, wrote a Python script to send automated emails. This is using the 3.5 API, and if you look in the bottom right corner, that these didn't cost me more than two cents each to run. So let's look at one of her examples. This is the detailed travel itinerary for a vacation to Italy. On her phone, she's typed objective, create a detailed travel itinerary for my upcoming trip to Italy. I will be on a two and a half week excursion visiting Florence, Verona, and Lake Cuomo, and then she put in some more information

about what she wanted and her budget. I Baby AGI first created a task list, and then it started executing the tasks one by one. It looked for affordable restaurants. It looked for affordable hotels. it gave a set of options for each. Now, what about the tools for automated sales prospecting? The objective here was very clear. Give me a list of tools for automated sales prospecting. The task list that Ibaby AGI generated for itself included identify and extract relevant data from sales prospecting sources such as LinkedIn, company websites, and industry databases. Two, use natural language processing to analyze and categorize the extracted data based on criteria such as job title, company size, and industry. And three, use machine learning algorithms

to predict which prospects are most likely to convert into customers based on historical data and patterns. The tools that it came back with a few minutes later included lead IQ, zoom info, Clearbit, sales navigator, HubSpot sales, alongside a quick description of each. Now from there, it did a thing which you see these auto GPTs do quite often, which is kind of overcomplicate the task. So some of the additional tasks that created for itself include use machine learning algorithms to predict which prospects are most likely to convert it to customers based on historical data and patterns. Use a decision tree algorithm to identify the most effective outreach strategy for each potential lead based on their extracted data and score.

Implement a chatbot to automate initial outreach to potential leads and gather additional information to improve their score, et cetera, et cetera. At the end of it all, the most valuable information was simply this list of tools that Lauren was then presumably able to follow up with and figure out which one she was most interested in on her own. Now, overall, Lauren says, my favorite things about Ibaby AGI, quick and efficient. No looping. It never gets caught in a loop.

It may provide the same or similar info repeatedly, though. Finishes objectives cleanly, detailed outputs, intelligently. segments, tasks into completed and remaining, track spend, easy copy and share feature to make outputs more useful. The tool is super user friendly and a lot of fun to play with. Now, of course, when it comes to AI tools, you have to try them yourselves. And so I downloaded Ibaby AGI, and my first impression definitely resonated with some of what Lauren said, which that it was an extremely easy user interface. It made it feel very fast to begin this process. All you had to do was input your open AI key, and you were off to the races. The first thing I tried was a purpose

broad objective. I said, I want to create a newsletter that can get to 100,000 subscribers. I'm interested in lots of different topics. Ultimately, what Ibaby AGI spit out was kind of exactly the sort of generic list that you might expect. You could talk about dad stuff. You could talk about technology. You could talk about finance. No real specification. And obviously, that was more my fault than its fault. It shows, I think, the more abstract the objective is, the harder it is for these tools to do it. But then I narrowed things down. And one of my next inquiries was about whether it could help me come up with a top 10 list of guests to have on the AI breakdown. For this, I even gave it the criteria that I wanted them to be based on popularity and reach.

For this one, it did much better. It said, added a new task. Use a machine learning algorithm to rank the potential guests based on their popularity and reach, taking into account factors such as the size of their audience, engagement rates, and the number of media outlets they have been featured in. It then decided how it wanted to present the information to me, saying it will provide a report with the following information for each guest. Name, social media following, website traffic, media appearances,

the number of times they had been featured in relevant media outlets. It also said that it was going to develop a scoring system that takes into account the factors identified in step one, as well as any additional relevant data, such as academic credentials or industry awards. Now, the list it came back with was pretty good. It roughly matched what you would expect if you asked an AI to find you the 10 most popular people in the space to interview. Now, of course, what it didn't take into account for was the difficulty of getting them on a show. For example, it has people like Elon Musk, but it also took the additional step of clustering them based on whether they were academic. or whether they were industry leaders. And so overall, it did provide some interesting utility.

Now, I also asked it to plan a last-minute date night for me and my wife in New York, and I gave it some specifications to help it out. I said we wanted to be in the Lower East Side. We wanted a hotel under a certain price, and we wanted restaurants within walking distance of that hotel. It came back with a totally reasonable plan on the basis of that. And so again, I think you're seeing here that specificity matters. Now, this idea of specificity and specialization seems to be something that other people are experiencing as well. When I went back to agent GPT, which I haven't looked at for a couple weeks, I noticed that instead of just a blank box of do whatever you want, although it still has that, it also has a few different agents that it suggests. It has platformer GPT, write some code to

make a platformer game. Travel GPT, for example, plan a detailed trip to Hawaii. Research GPT create a comprehensive report of the Nike company. So these are clearly three different use cases for coding, for travel, and for research that Agent GPT seems to think are primary use cases for this technology. The emergence of tools like AOMNI, which specifically focuses on online research, also suggests that what people are finding is that the more that you can tune these models to go after specific use cases, the more likely to produce good results they are. I think this was almost inevitably going to be the pattern here, where we went from very

generalized use cases to actually honing in on the specific things that these tools could be most helpful for and designing new tooling around those use cases. What's more, I'm also seeing a lot more people start to figure out what the relationship between the auto and automation part of this should be versus the human part. Developer Nick Dobos, for example, had a long threat on April 28 that says, GPT and me, a baby AGI slash auto GPT style agent that creates full human and AI symbiosis. The only AI agent that can actually accomplish things today, because the task execution sub-agent is me, become a human and AI agent cyborg.

Nick then shows these detailed schematics, and basically what they come down to is that you use the AutoGPT style tool to create the task list to prioritize it, and then to say what needs to happen. Then instead of that third step where the AutoGPT just creates a new AI agent to do that task, the human becomes the execution agent. They return that result and then use that to enrich the AutoGPT's information so that it can reprioritize and give the human agent the next task. Now, in some ways, this is all very simple, but what it is is trying to actually figure out the process by which these sort of auto-GPT tools are really productivity enhancing without assuming that they're going to be able to do everything all on their own. Remember, these experiments are less than two months old, and so it doesn't surprise me that we're seeing both a move to specialization, as well as a move to people figuring out just how autonomous auto-GPT should really be. Now, the use case that I mentioned that I haven't really seen anyone talking about is really simple. It's brainstorming. One of the things that I've been underwhelmed with these tools around is their ability to go from an objective to a completed task.

It involves a ton of abstraction, a ton of steps, and the fact that you're seeing objective specialization, as well as people like Nick, experimenting with the human part of this, suggests that that's a difficulty for others as well. What I found auto GPs to be amazing for, however, is a partner in how to think about accomplishing different objectives. In other words, one of the most valuable parts of letting an auto-GPT, like I Baby AGI, go off to try to finish an objective, is to see how it actually approaches it. It's not just the tasks being completed that's valuable. It's which tasks it thinks are important in the first place, especially as an entrepreneur or a content creator or anyone who's working in a small environment where they have to be dynamic and do a lot of different things in context switch. Having a brainstorming partner that can help think from zero to one when it comes to important tasks is incredibly valid.

Sure, of course, we want these technologies that can do everything for us at the drop of a hat or the snap of fingers, but helping outsource our process brain and make sure that our process is most efficient is actually really valuable as well. Now, one of my biggest questions coming up with AutoGPTs is to what extent ChatGPT with browsing fundamentally changes the need for them. The two plugins that OpenAI has been working on for ChatGPT are, of course, code interpreter, which we talked about in a previous video, and chat GPT with browsing. As you would imagine, browsing gives chat GPT access to the internet, which is one of those major use cases of auto-GPT. Now, it still doesn't mean it's necessarily designed to auto-prompt itself

to figure out the tasks around an objective. But it's not impossible to me that what we'll really find is that a big part of what feels so amazing about auto-GPTs, at least right now, is the fact that they were connected and can go out and search the internet on our behalf. With GPT4 with browsing rolling out more fully, I think we'll have much more of a chance to understand. that in the weeks to come. All right, guys, that is it for the AI breakdown today. Go check out iBaby AGI. Nate Chan did a great job with it. If nothing else, you will have a ton of fun just

experimenting and learning and seeing what it does. If this video was valuable, please like and subscribe to the channel. Go follow the podcast on Apple or Spotify, wherever you listen. That's the AI breakdown. Come back tomorrow for another dose of the AI breakdown brief and the AI breakdown. Until then, peace.