ChatGPT's AI agents will change how you use chatbots
What happens when an expert provides a curated database of knowledge for you to interrogate?
Yesterday’s post was a high level overview of what AI agents are. Now let’s look at some in more detail. The easiest way to understand these things is by looking at the ones you can build and use with ChatGPT1. OpenAI, ChatGPT’s developer, confusingly refers to these agents as GPTs, as in “general purpose technology”.
This is their description of GPTs:
We’re rolling out custom versions of ChatGPT that you can create for a specific purpose—caled GPTs. GPTs are a new way for anyone to create a tailored version of ChatGPT to be more helpful in their daily life, at specific tasks, at work, or at home—and then share that creation with others.
I don’t find this explanation to be very good. A better explanation is this one, by Bojan Tunguz, an AI researcher at Nvidia, reflecting on his time at OpenAI’s recent DevDay:
GPTs are user-defined context-specific instances of ChatGPT, built with a specific use case in mind. They are extremely easy to build using just prompts and a few accompanying documents that users can upload. A less charitable view of them would be that they are just specialized and repackaged domain-specific prompts. But just like with anything else OpenAI related, it’s the UI/UX that makes a big difference. Sharing bespoke prompts and chat instructions was inelegant. Sharing GPT “apps” is much more straightforward and intuitive.
I would add to this explanation that OpenAI’s GPT AI agents allow someone to instantiate their knowledge into a shareable ChatGPT session which anyone with a paid subscription can access. In principle, this can be a very powerful tool, as we’ll see later in this post when we look at some examples.
Bojan goes on to note that OpenAI is planning a GPT Store, through which developers of these GPTs can sell their creations, akin to Apple’s App Store. Here’s a Twitter thread about this idea. I’m not convinced that this store will get much traction, for the simple reason that the most powerful aspect of these GPTs, which is the ability to call other services’ APIs, can’t be used in ChatGPT’s mobile app due to privacy restrictions instituted by Apple and Google.
Since most people are mobile-first these days, this is a non-starter. Unless Apple and Google change their privacy policies to allow for unfettered API calls through apps that reside on users’ phones, I just can’t see this market taking off. For GPTs which don’t use API calls, they are easily copied by other people.
Another thread which helps to understand how this all works is this one.
In spite of these drawbacks, the concept of custom ChatGPT instances is a useful one, and there are some very cool things being built by people. The rest of this post takes a look at some of these. Understanding OpenAI’s vision of AI agents can help you understand AI agents more generally.
The best way to understand each of these AI agents is to give it a prompt, read its response, and react to the prompt. So I’ve re-created my interactions with each of these AI agents. After each exchange I’ll offer my comments in brackets.
GPT MD — An experimental doctor for medical guidance. Design by a Doctor.
Prompt: I’m trying to lose weight. I go to the gym five days a week and lift weights for an hour each session, but I don’t seem to lose weight around my waist.
GPT MD: It’s great to hear that you’re actively working out and focusing n your fitness! To better understand your situation and provide more personalized advice, could you tell me a bit about your current diet? What types of food do you typically eat in a day, and do you follow any specific dietary plan?
Prompt: I eat eggs for breakfast, cheese and crackers for lunch, yogurt with honey, banana, and almonds for a post-workout snack, and usually something like chicken, rice and salad for dinner.
GPT MD: Your diet seems to have a good balance of protein and carbohydrates. To delve deeper into your weight loss journey, could you share how your portion sizes are for these meals?
Prompt: Yeah my portion sizes are pretty big I guess. What can I do about that?
GPT MD: Managing portion sizes can be a key factor in weight loss, especially around the waist. Here are a few practical tips you can try:
Use Smaller Plates: Switch to smaller plates for your meals. This can psychologically make your portions seem larger.
Half-Plate Rule: Aim to fill half your plate with vegetables, a quarter with lean protein (like chicken or fish), and the remaining quarter with whole grains or starchy vegetables.
Slow Down: Eating slowly and chewing thoroughly can help you better recognize your body's fullness signals, preventing overeating.
Avoid Eating from the Package: Portion out snacks into a bowl or plate instead of eating directly from the package.
Stay Hydrated: Sometimes thirst is mistaken for hunger. Drinking water throughout the day can help manage hunger and fullness.
Meal Prep: Preparing meals in advance can help control portion sizes and reduce the temptation to overeat.
[Overall I think this is a pretty good experience. There is a lot of potential for AI Agents built by credentialed medical doctors to help people with common medical questions. To some extent, instantiating these dialogues in AI agent form, instead of in the doctor’s office, can free up both doctors’ and patients’ time and lead to a more efficient healthcare system.]
Cochrane Navigator — Guiding researchers on evidence synthesis
Prompt: Explain what you do, and provide a sample use case.2
Cochrane Navigator: Hello! I'm Cochrane Navigator, a specialized GPT designed to assist with understanding and applying the guidelines outlined in the Cochrane Handbook for Systematic Reviews of Interventions, 2nd Edition, 2019. My role is to make the complex process of conducting and reporting systematic reviews more accessible and manageable.
Sample Use Case:
Imagine you are a researcher planning to conduct a systematic review of interventions in a healthcare-related field. You're aware of the rigor and detail required but feel overwhelmed by the guidelines and best practices. Here's how I can help:
Clarifying Methodology: You might ask, "What are the key steps in defining a research question for a systematic review?" I would provide a clear and concise summary of the relevant sections from the Cochrane Handbook, detailing how to formulate a research question using the PICO (Population, Intervention, Comparison, Outcome) framework.
Data Extraction Guidance: If you're uncertain about the process of data extraction, you could ask, "How should I extract and manage data for my review?" I would offer advice on systematic data extraction methods, including the use of standardized forms and the importance of having multiple reviewers.
Statistical Analysis: For questions like, "What statistical methods are recommended for meta-analysis in systematic reviews?" I would outline the recommended statistical approaches for meta-analysis, such as fixed-effect or random-effects models, as described in the Handbook.
Addressing Bias: If you're concerned about bias, you might ask, "How can I assess and minimize bias in my systematic review?" I would explain methods for assessing risk of bias in studies, as well as strategies to minimize it throughout your review process.
In each case, my responses would be grounded in the authoritative content of the Cochrane Handbook, ensuring that the advice is accurate and aligned with current best practices in systematic review methodology.
[I don’t have the requisite background to make much sense of what this AI agent does, so I stopped this after one prompt, but it sounds like it would be useful for certain researchers looking for guidance about proper research practices. This could be very helpful for a stressed out graduate student trying to figure out how to proceed with her research.]
FOIA GPT — Freedom of Information Act request strategist to “arm the rebels” for truth and transparency in the fight against corruption3
Prompt: I want to write a FOIA letter to NASA regarding relating to Richard Feynman’s investigation of the Challenger Space Shuttle’s O-rings.4
FOIA GPT: To draft a Freedom of Information Act (FOIA) request letter to NASA regarding records relating to Richard Feynman's investigation of the Challenger Space Shuttle's O-rings, we need to focus on a few key elements:
Clarity of Request: Clearly state the specific documents or types of records you are seeking. For example, you might be interested in communications, reports, meeting notes, or other documents related to Richard Feynman's involvement in the investigation.
Time Frame: Specify the time frame for the records you seek. Since the Challenger disaster occurred in 1986, you would likely be interested in documents from that period.
Reason for Request: While not always necessary, providing a brief reason for your request can sometimes aid in processing.
Contact Information: Include your contact information for any follow-up.
Citing the FOIA: It’s important to explicitly mention that your request is made under the Freedom of Information Act.
Let's draft a basic template for your request:
[Your Name]
[Your Address]
[City, State, Zip Code]
[Email Address]
[Phone Number]
[Date]
National Aeronautics and Space Administration
Office of the Chief Information Officer
Freedom of Information Act Office
300 E Street SW
Washington, DC 20546
Email: foia@nasa.gov
Dear NASA FOIA Officer,
Subject: Freedom of Information Act Request
I am writing to request records under the Freedom of Information Act (5 U.S.C. 552). Specifically, I am seeking all documents and records related to Dr. Richard Feynman's investigation of the O-rings in the Space Shuttle Challenger disaster, including but not limited to:
Communications (emails, letters, memos) involving Dr. Feynman.
Reports or findings authored or co-authored by Dr. Feynman.
Minutes or notes from meetings where Dr. Feynman was present or his findings were discussed.
Any audio or video recordings of interviews or discussions with Dr. Feynman.
The time frame for this request is from January 1986 to December 1987.
[Optional: Include a brief statement about why you are requesting these records, if you feel it's necessary.]
Please inform me of any fees required for searching or copying the records before processing my request. If you expect fees to exceed [insert amount], please provide a detailed breakdown of the fees and contact me before proceeding.
I am looking forward to your timely response within the 20 business days as stipulated by the FOIA. Should you have any questions or need additional information, please do not hesitate to contact me at [Your Email Address] or [Your Phone Number].
Thank you for your attention to this matter.
[This seems like a very helpful tool if you have never drafted a FOIA request before.]
Concluding thoughts
OpenAI’s GPT AI agents promise to create much more utility for users than simple chatbots like ChatGPT. They are a mechanism for an expert to assemble a database of curated knowledge and upload that set of curated knowledge to a chatbot. Users can then interrogate the chatbot about the knowledge, and rapidly learn a lot about the topic at hand.
One feature which is especially important, but which I don’t discuss much in this post, is the ability to append ‘actions’ to these AI agents. Actions allow users to call external APIs to execute actions with other services. For example, someone might build an GPT which, after a series of prompts and responses between user and GPT, executes an action in Zapier. The problem with this feature is that a GPT which uses these kinds of API calls is not available on the ChatGPT mobile app, due to restrictions put in place by Google and Apple.
Unfortunately, these API calls are what give GPTs their true power and extensibility. But since so many people are mobile-first these days, they will just ignore GPTs which they can’t use on their mobile phones/tablets. This means that any attempt by OpenAI to build a market for GPTs will be hobbled: the most popular GPTs will be those that people can use on mobile devices, and because of the aforementioned limitations imposed by Apple and Google, these GPTs won’t be the most powerful ones.
Finally, it’s important to consider that those who learn to build these GPTs (or other AI agents) and interact with them will fare better than those who allow their skills and knowledge to atrophy. As Alberto Romero so poignantly writes:
A human using AI will always be superior to just an AI. But the truth is that, theoretically at least (and taking some liberty in what I mean by AI here), the human element will lose value and importance over time, as far as solving cognitive or creative tasks goes, until the expression becomes this:
AI + human ~ AI; human ~ 0.
This may sound dramatic but if we look at the disciplines that AI has truly mastered above the best humans, the pattern is evident. Human performance eventually becomes not just practically useless, but even a hindrance. Chess is the quintessential example. A grandmaster using a modern chess engine will always obliterate a grandmaster playing alone.
So, as you learn more about AI agents or GPTs, keep this in mind. AI agents will become more powerful and capable. The worst AI you’ll ever use is the one you use today. As they become more powerful and more capable, you, too, will become more powerful and more capable.
Note that in order to use ChatGPT’s AI agents, you have to have a paid subscription to the service.
I used this prompt because I had no clue what this AI agent does.
The Freedom of Information Act is a US law which requires that the government provide certain records to people who request them. More information is available here.
After the Space Shuttle Challenger exploded in 1986, Richard Feynman debunked the government’s explanation by illustrating that the O-rings the shuttle used were defective in cold