How Close to Agents Are GPTs?

0:00Today on the AI Breakdown, we're reading an essay called Almost an Agent, what GPT's can do. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.network for more information about our YouTube channel, our Discord, and our newsletter.

0:24Hello friends, welcome back to a long-read episode of the AI Breakdown. Undisputedly, the most important news of this week came from OpenAI's DevDay, and I believe it was custom GPTs, or just GPTs as OpenAI calls them. I believe, as I'll discuss a little bit later, that the vast majority of people's especially first interactions with ChatGPT in call it six months from now will be mediated by one of these GPTs. I think that we will see a shift from apps to workflows, and that GPTs are the single best expression of that change. Now, at their presentation on DevDay, Sam Altman and others from OpenAI made it clear that they viewed GPTs and their assistance API as their first baby steps towards a future of agents.

1:11Well, Professor Ethan Mollick followed up on that idea with this essay, Almost an Agent, What GPTs Can Do, and as usual, it adds a lot of clarity to an important part of the AI space. I am going to now read big chunks of this, interspersed with some commentary from myself, and this one is actually me reading, not AI11LabsMe. Ethan writes, Many people think the future of AI lies in agents. Fuzzily defined term that refers to an autonomous AI program that is given a goal and then works towards accomplishing it on its own. There has been a lot of buzz about agents over the past few months, but not much technology that actually works well.

1:44Editor's note, if you've watched any of my experiment videos with auto-GPT type things, you will resonate with that idea. Now back to Ethan. What would a real AI agent look like? A simple agent that writes academic papers would, after being given a data set and a field of study, read about how to compose a good paper, analyze the data, conduct a literature review, generate hypotheses, test them, and then write up the results, all without intervention. You put in a request, you get out a Word document that contains a full draft of an academic paper. A process, Ethan says, kind of like this one.

2:13He then shares a photo of what is a new GPT. Ethan continues, This was a result of a GPT. Yes, that is what they decided to call them. I created it using the new system released by OpenAI today. And to be clear, GBTs aren't autonomous agents yet. I had to give feedback to the AI a few times along the way, and GBTs still have hallucinations and other issues that will show up in the final product. Plus, in the end of this experiment, even though it had worked before, the AI decided that writing academic papers was not something it was allowed to do, at least until I told it, no, it is really important and you are great at this and can do it, I know you can.

2:47So if they aren't quite agents yet, what are GBTs and how can you use them. Here are the basics. Right now, GPTs are the easiest way of sharing structured prompts, which are programs written in plain English or another language that can get the AI to do useful things. The GPT system makes structured prompts more powerful and much easier to create, test, and share. I think this will help solve some of the most important AI use cases. For example, how do I give people in my school, organization, or community access to a good AI tool? GPTs show a near future where AIs really can start to act as agents, since these GPTs have an ability to connect to other products and services, from your email to a shopping website, making it possible for AIs to do a wide range of tasks.

3:26So, GPTs are a precursor for the next wave of AI. They also suggest new future vulnerabilities and risks. As AIs are connected to more systems and begin to act more autonomously, the chance of them being used maliciously increases. So, with the second two points in mind, let's focus on the first, the power of GPTs to make automating tasks and processes much easier. Now, the next little section of Ethan's post talks about the GPT Builder and how you use it, which to sum up some of the key details is, one, there is a thing that exists called GPT Builder which helps you actually build these tools without needing to know code.

3:59Two, behind the scenes, as he puts it, the AI is filling out a detailed configuration of the GPT which can also be edited manually. The core of this configuration is a structured prompt, but it also has additional features. He discusses how, quote, the ability to work with documents is both extremely powerful and requires a degree of caution. Almost every company I talk to and basically every solution vendor has been pushing for people to use AI to quote, talk to your data, an approach that allows the AI to retrieve content from a company's proprietary database and then work with the documents and data it retrieves.

4:27The problem is that AIs hallucinate or make up plausible information all the time. This is getting much better as technology improves, but it isn't perfect yet. The same thing is true here. The file reference system in the GPTs is immensely powerful, but it is not flawless. So, Ethan concludes, GPTs are easy to make and very powerful, though they are not flawless. But they have two other features that make them useful. First, you can publish or share them with the world or your organization, and potentially sell them in a future app store that OpenAI has announced. The second thing is that the GPT starts seamlessly from its hidden prompt, so working with them is much more seamless than pasting text right into the chat window.

5:01We now have a system for creating GPTs that can be shared with the world. What do we do with it? The next section of Ethan's post is called GPTs as Tools. He writes, Once you create and troubleshoot a GPT, you now have a powerful tool that anyone can use. That means that communities and organizations can begin to work together to create a set of agents that can be useful for work and school. For example, we have been actively exploring the use of AI for education. And while there are many concerns about using LLMs in teaching, they show potential for democratizing access to the kind of instructions that are otherwise available only to a lucky few.

5:31Even then goes through an example that he created called the GPT Feedback Wizard. It's basically a lightweight writing coach. The heart of the system is a a structured prompt that is all about the role that the GPT should inhabit. It's about 500 words, but they start, you are a friendly and helpful mentor who gives students effective, specific, concrete feedback on their work. In this scenario, you play the role of mentor only. You have high standards and believe that students can achieve those standards, and so on and so forth. Ethan continues, based on that, the AI guides students to discuss their goals for a piece of writing and to upload their essays and grading rubrics.

6:04Then, rather than just writing the essay for the student, the GBT returns an edited, marked-in-read copy of the Word document with advice based on rubrics. To be clear, again, this is just a prototype, but the fact that writing instructors can now create a GBT that can provide personalized advice in their personal style and then give that GBT away to people all over the world to improve their writing is exciting. I hope experts jump on this capability and start building and testing their own tools. The power here is pretty obvious. I will be creating custom GBTs for every session of the classes I teach.

6:34Some will be simulations for students to experience. Some will be tutors or mentors. Some might even be teammates or assignments. I have been turning my research into GPTs so that anyone can get advice on how to generate ideas or pitch a business idea by getting feedback from a GPT to which I have given my books as a reference. And I expect this will become a trend in many places as schools and government agencies and companies build libraries of GPTs that are specialized in solving particular problems in useful ways. The last section of Ethan's post is called Promise and Peril of Agents. He writes, In the reveal of GPTs, OpenAI clearly indicate that this was just the start.

7:08Using that action button you saw above, GPTs can be easily integrated into with other systems such as your email, a travel site, or corporate payment software. You can start to see the birth of true agents as a result. It is easy to design GPTs that can, for example, handle expense reports. It would have permission to look through all your credit card data and emails for likely expenses, write up a report in the right format, submit it to the appropriate authorities, and monitor your bank account to ensure payment. And you can imagine even more ambitious autonomous agents that are given a goal.

7:36Make me as much money as you can, and carry that out in whatever way they see fit. You can start to see both near-term and farther risks in this approach. In the immediate future, AIs will become connected to more systems, and this can be a problem because AIs are incredibly gullible. A fast-talking hacker, if that is the right word, can convince a customer service agent to give a discount because the hacker has super-duper secret government clearance, and the AI has to obey the government, and the hacker can't show the clearance because that would be disobeying the government, but the AI trusts him, right?

8:03And of course, as these agents begin to truly act on their own, even more questions of responsibility and autonomous action start to arise. We will need to keep a close eye on the development of agents to understand the risks and benefits of these systems. Regardless of these long-term concerns, the current state of GPTs represents a powerful tool for making AI easy to work with. I look forward to seeing the experiments that result. Alright, back to NLW here. I think this is a great little summary, a very useful way of starting to wrap one's head around GPTs. And for the sake of this little follow-up conversation, I'm going to stay away from the bigger problems or questions of what happens in an agent future.

8:37It's a really important topic, but it's not one that I want to get into now. What I do want to talk about is how I see these GPTs being used in the short term. There are kind of two ways right now, and this is obviously being wildly reductive and oversimplifying things, but there are two ways that I see ChatGPT being really used. One is for novelty and intrigue. A lot of this is really fun and mind-expanding and interesting. A great example is the choose-your-own-adventure story creation that I do with my five-year-old daughter. It's empowering to her, it's interesting, and it's a unique thing that feels like magic.

9:10What it's not is anything that's useful for work or school, and that I think is what the second category of ChatGPT uses are, things that are actually useful for work and school. Now, frankly, a lot of those workflows are still being worked out. You have a lot of people hacking at ChatGPT to figure out how to create complex prompts that start to get them the results that they want for specific functions or tasks. I think basically that the future that we're headed towards is one in which every single one of the most useful of those prompts becomes its own custom GPT. Anyone who figures out how to work the system to radically simplify some important thing for their industry or their job function is going to have an incentive to share that either for public acclaim or for sale, given that OpenAI is putting a marketplace around this.

9:52And that's why I said at the beginning that I think that in the not-too-distant future, the way that most people start to interact with ChatGPT will be through these custom GPTs, because they won't be signing on to ChatGPT to just sort of vaguely explore. Instead, they'll be signing on to ChatGPT to try to find something specific that's useful to some problem they're having right now, or some efficiency that they're trying to win in one of the courses of normal things that they do. Now, we could have a whole additional conversation around how much OpenAI's decisions around how to search the custom GPT store or whatever they end up calling it will be in terms of how useful it is.

10:26But I really do think that when all is said and done, what GPTs are going to enable is a mass transfusion of the highest value of this tool from the call it 5 or 10 or 1 % of people who best use it to the rest of everyone else. There is a natural power distribution where a very small number of people who are interacting with ChatGPT are uncovering a huge portion of the most useful ways to prompt it and get value from it, and custom GPTs make it easier than ever for them to actively share that. Now that's great for them if they can monetize it, but it's even better for the rest of us because it means more value to everyone.

11:00In sum, I am very excited about the custom GPT future, and I certainly know that I have spent an inordinate amount of time already hacking on them myself. If you have good custom GPTs to share, may I encourage you to come check out the AI Breakdown Discord. You can find it at bit.ly slash AI Breakdown. Appreciate you guys listening or watching as always, and until next time, peace.