TheTechOasis
Posts
The AutoGPT Revolution, The Era of GPT-4 Agents Is Here

The AutoGPT Revolution, The Era of GPT-4 Agents Is Here

GPT-4 is Now An Autonomous Agent, But What Does that Mean?

Ignacio de Gregorio Noblejas
April 16, 2023

🏝 TheTechOasis 🏝

🤖 This Week’s AI Insight 🤖

AutoGPT is taking the AI world by storm. Here’s why

In the last Lex Fridman’s podcast, with Max Tegmark, a famous MIT researcher and scientist, they mentioned the potential implications of developing an autonomous GPT-4, one that could reevaluate and debug its tasks in a recursive loop with no human participation.

The implications, they mentioned, could go far beyond simple productivity increases, to the point of being a literal risk for humankind.

Well, that exists now in the form of AutoGPT, and everybody is going nuts for it.

But why?

One prompt and that’s it

With the standard GPT-4 model, all interactions are one prompt at a time.

You request/ask it something
It answers
You answer back in case you need more context

And so on.

The results and potential applications, in this case, are already endless, but there’s a “problem”.

It’s always a conversation-based activity.

Again, this isn’t necessarily bad, but what if I just want to tell it to do something and let it do its thing?

Humanoid in the style of Agostino Arrivabene. Source: Author with Diffusion Model

This, obviously, isn’t possible with ChatGPT just as it is, we need something more.

Also, there’s another problem.

Its memory is also very limited, constrained to the time boundaries of that specific interaction, making it unfeasible to create complex tasks with ChatGPT.

Additionally, we have the issue of taking action; ChatGPT can explain to you how to do something, but it can’t help you to actually do it (besides code).

But these problems are a thing of the past with AutoGPT, a model that opens a staggering new number of use cases and potential for LLMs like GPT-4.

Memory, Decision, and Action

AutoGPT is a new, usable library that leverages OpenAI’s GPT-4, Pinecone vector databases, and the LangChain framework to exceed GPT-4’s capabilities to become a fully autonomous, long-term memory agent.

Particularly, it uses the different tools as follows:

GPT-4: The central figure, used for language understanding, reasoning, and task creation/prioritization
LangChain: To allow GPT-4 to execute actions by integrating it with other tools
Pinecone: A vector database to store the data and allow GPT-4 to lengthen its context window. It’s its memory, to be clear.

In short, AutoGPT can, given a single pre-structured prompt with a goal and a set of tasks, execute them, measure the results, and create and prioritize new tasks depending on the results of the previous ones.

All by itself.

It just needs a pre-structured prompt like the one below:

Name your AI: “Newsletter Creator”

Describe your AI’s role: “An upcoming writer wanting to spread the word of AI through a weekly newsletter and make a living of it.“

Enter five goals for your AI:

“Find an interesting, trendy idea that I will enjoy writing about and my readers will enjoy reading to”

“Define the brand, logo, and vision for this project”

“Create an appropiate landing page with its unique domain name previously purchased for the ocasion”

“Code the landing page for people to subscribe to my newsletter and choose the email service provider that best suites the topic in question”

“Define a monetizing strategy that, above all, ensures the newsletter stays accessible and free-based”

And… that’s all the interaction you need with the model, with AutoGPT proceeding then to perform the tasks in that order, according to a similar schema to the one below:

As you can see above, the model progressively executes tasks, evaluates them, defines new ones if necessary, and reprioritize task execution, all while using long-term memory to not forget what the end goal is, which is the role defined in the initial prompt.

Freakish.

But the real question is… does this take us closer to Artificial General Intelligence?

A new paradigm, but not for AGI

Lots of people, especially hypers or doomsayers, say AutoGPT is an obvious next step for AGI.

And I agree… partially.

Yes, AGI, if ever achieved, will be autonomous, but autonomy doesn’t implicitly mean sentience.

In other words, while all sentient beings are more or less autonomous, not all autonomous systems are sentient, and we can see this every day with automation technologies like robotics.

But I do feel that AutoGPT does represent a new frontier… for prompting.

For this, instead of me explaining, what better than one of the godfathers of modern AI, Andrej Karpathy, to do it for me:

Next frontier of prompt engineering imo: "AutoGPTs" . 1 GPT call is just like 1 instruction on a computer. They can be strung together into programs. Use prompt to define I/O device and tool specs, define the cognitive loop, page data in and out of context window, .run().
— Andrej Karpathy (@karpathy)
6:44 PM • Apr 2, 2023

In other words, what Andrej is saying is that autonomous systems like AutoGPT will change how we interact with machines, to the point of reducing it to mere, first-principle instructions.

Imagine a world where all you need to do to a machine is explain to it what you want, one time.

Seems like science fiction, but we’re getting closer to it… by the day.

If you care to read the original AutoGPT paper, here’s the link:

AutoGPT: Task-driven autonomous agent

Yohei Nakajima