The Autonomous Agents Are Coming and It’s Time to Prepare

Photo by Michael Herren on Unsplash

Over the last six months or so, we have witnessed large language model’s public and commercial breakthrough. Impressive and useful as AI models like ChatGPT and GPT-4 undeniably are, what we see is only the finely polished end result.  Beneath the user-friendly surface, we don't see the $100 Million investments that have gone into training GPT-4, the heavy carbon footprints it leaves behind, or the underpaid workers in developing countries who have worked long hours to filter out toxic content for ChatGPT.

Furthermore, large language models have been trained on generations of user-generated content, collective, unpaid efforts that we can now pay to access. In this light, the incredible capabilities of modern large language models may be slightly deceiving, and a natural consequence of the knowledge- and power gap between tech monopolies and the general public. We are being served a delicious, expensive steak in gourmet style, unaware of the unpaid labor that has gone into slaughtering the cow and preparing the meat.

It's now time for lawmakers to step in and regulate the development and use of AI. This job requires lawmakers to be extraordinarily foreseeing. By the time new regulation such as EU's AI Act finally comes into force, the landscape is probably very different from today.

Allegedly, ChatGPT has already broken the EU Commision’s plan to regulate AI. And now, as ChatGPT and other large language models are on the EU Commission’s radar, many experts including OpenAI CEO, Sam Altman, have said the age of giant AI models is already over. 

BigTech companies are racing against each other to capture people’s attention with increasingly seductive measures, and lawmakers are falling several labs behind. The only way that the turtle can beat the rabbit in this particular race would be if regulators can expect the coming developments years and decades ahead.

Judging from recent developments, the next big thing in AI could very well be autonomous agents.

Autonomous Agents

I see autonomous agents as the next generation of advanced chatbots. While ChatGPT is useful for carrying out simple, text-based tasks, the next generation of advanced chatbots has much more agency. In other words, they will be able to act on their own accord without much or any human influence. 

As of today, the technology for widespread use of autonomous agents is not ready. However, that doesn’t mean that we shouldn’t be. Here’s a brief overview of recent developments.

ChatGPT Plugins

One month ago, OpenAI announced that they had implemented initial support for plugins in ChatGPT.

The new web-browser plugin makes it possible for ChatGPT to access the internet and retrieve up-to-date information outside of its training data. A feature that users have been asking for ever since ChatGPT’s arrival.  Another new plugin by OpenAI is an experimental code interpreter that can use Python in a sandbox environment, handle uploads and downloads, and make life easier for programmers as well as programming-interested people without skills in the field.

With additional plugins to third-party services such as Klarna, Expedia, OpenTable Shopify, Zapier, and Slack  ChatGPT can help you to pick out the best deals and discounts, give you travel tips, write shopping lists and buy groceries, make dinner reservations, compose business e-mails and send them in Gmail or via chat messages in Slack.

HuggingGPT

Earlier this month, Hugging Face published a paper: HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face.

The authors show how ChatGPT can act as a controller to manage existing AI models with the use of language. By connecting ChatGPT and other AI models, you can prompt ChatGPT not only to generate text, image, video, and audio but also to identify and describe objects across different formats of content.

For example, you could upload a document.jpg to ChatGPT, ask it to retrieve some information you are looking for, and make it read the text out loud for you. Or you can prompt ChatGPT to count the number of certain objects in an image, then describe the image in detail, and make a video based on the description.

In this way, ChatGPT can be used to solve complicated AI tasks by leveraging other AI models.

AutoGPT

Last week, #AutoGPT was the top-trending hashtag on Twitter. AutoGPTs have become an extremely popular area of study and experimentation for open-source developers. 

In essence, open source models like AutoGPT and BabyAGI or Microsoft’s Jarvis aim to give “the brain” of large language models like GPT 3.5 or GPT-4 arms and hands to carry out tasks based on a defined goal or objective.  For instance, you can ask an AutoGPT to find the best and cheapest restaurant nearby and book a table. Or you could ask it to destroy humanity, establish global dominance, and attain immortality.  With access to a large language model, the internet, and equipped with long- and short-term memory, the AutoGPT will hereafter prompt itself with tasks to attain the provided goal.

While AutoGPTs requires a bit of coding skill to set up and use, you can try out the web-based version, AgentGPT, to get a taste of the experience. 

ChatGPT Meets the Sims

In a recent experiment conducted by Stanford University and Google Research, so-called, "generative agents" simulate human behavior and act on their own in a simulated world reminiscent of The Sims. According to the paper:

By connecting our architecture to the ChatGPT large language model, we manifest a small society of twenty five agents in a game environment. End users can observe and interact with these agents. If an end user or developer wanted the town to host an in-game Valentine’s Day party, for example, traditional game environments would require scripting tens of characters’ behavior manually. We demonstrate that, with generative agents, it is sufficient to simply tell one agent that she wants to throw a party (..) They spread the word about the party and then show up, with one agent even asking another agent on a date to the party, all from this single user-generated seed suggestion.

You can follow the generative agents in the simulated Sims-like environment here.

The Future

In my mind, the simplest way to think of autonomous agents is as personal, digital assistants. They are like ever-present butlers who can take care of our needs in the digital world. And help out with tedious tasks that no one enjoys doing. As AI grows in agency, it’s not hard to imagine that we will eventually see autonomous companies which thrive without employees or management.

Autonomous agents will magnify the depth and scale of the serious issues we face with AI and social media already today . Issues such as privacy, smartphone addiction, AIs black box problem, the alignment problem, etc. will have whole new meanings.

If you are interested to follow my writing on these issues and haven’t done so already you can subscribe to The Gap below.