The Week in AI: New Bot, Who Dis? AutoGPT (Plus AGI, Regulation, and Run GPT Locally)
AI research is accelerating, not pausing or getting regulated (yet).
In 2020, there was GPT-3.
In November of 2022, ChatGPT broke the internet.
Just a couple of months ago, on March 14th, GPT-4 was released.
And on March 28th, we saw some of the first uses of AutoGPT, or “AI Agents”.
This is the latest from a frontier that’s shifting and moving as we speak.
So, what is AutoGPT?
Some are calling it “Primitive AGI” and others “Zapier on steroids”.
But it’s an AI system that, given a task, runs in a loop until the task is completed or solved.
The concept of AI agents has been around for a few months and aren't that new either.
They were only recently brought into popularity and attention via @yoheinakajim, with “Task-driven Autonomous Agent” that combines GPT-4, Pinecone, and LangChain:
And then there was a paper on “Generative Agents”:
These agents are engaged in complex and (to some degree) autonomous behavior.
And now they’re being unleashed all over the internet:
Using an “AI Personal Assistant” is straightforward. Describe what you want it to do, and it will automatically operate Chrome for you to achieve your task.
How do these AutoGPTs work?
Let’s break it down:
Together with terms like AgentGPT, AutoGPT is becoming an umbrella term for BabyAGI, Jarvis by Microsoft, and various autonomous AI Agents.
An AutoGPT operates on its own and completes tasks for you. Give it a goal (instructions and parameters) and off it goes to do your bidding.
It allows GPT-4 to prompt itself, making an LLM autonomous. For now, it’s somewhat autonomous but likely to become fully autonomous soon.
It can be connected to Google via the search API, and can conduct searches.
It can read websites, break information into chunks, summarize, and create files.
It can be connected to Pinecone, which can give the AI agent something like a long-term and short-term memory.
And AutoGPT can review its work, build on its previous efforts, and use its history to produce more accurate results.
It’s like a robot with access to resources. You assign it a goal and then AutoGPT comes up with a plan to accomplish it. And in pursuit of that goal, it can browse the internet, perform searches, use new data, and adapt until the task is checked off as done.
Is it any good?
It looks promising:
Still, it’s early and they break more often than they work.
But, as with all things AI at this point, it’s only a matter of time until it’s fully usable—and that time is measured in days or weeks, at most months, and not years.
Perhaps AGI is next:
Or something more sinister:
Interested in taking your own AutoGPT for a spin?
Here are a few resources to get you up and running:
Set it up yourself, and have it doing your bidding for you in less than 30 minutes:
Or go straight to cloning the repo: https://github.com/Torantulino/Auto-GPT
Youtube walkthrough on how to install AutoGPT:
Get it running on your iPhone with Replit:
Try a limited, easy-to-use web version of AgentGPT here: https://agentgpt.reworkd.ai
Or try an AI agent from Cognosys: https://www.cognosys.ai/
Signup for a chance to be among the first to experience the revolution in web browsing with Personal Assistant from HyperWrite:
The five stages of mourning the future.
BTW, how’s everyone coping with the AI-pocalypse these days?
We’re seeing all five stages of grief play out. It looks like a lot of people are in the “Anger” stage:
Next up is “Bargaining”. People will become lost in a maze of “If only…” or “What if…” statements. They’ll want life returned to what it was, to go back in time.
Or perhaps many will jump straight to “Depression”. That’s where your attention is on the present. The sooner you get to Now, the faster you can take AI by its horns and ride the bull.
For now, it looks like Germany is looking to follow in Italy’s footsteps with a ban. They’re caught in the first stage, “Denial”.
The concern is, once again, over privacy. Italy blocked ChatGPT after its data protection authorities concluded that the system does not have a proper legal basis for collecting personal information about its users.
“In principle, such action is also possible in Germany,” data protection chief Ulrich Kelber said.
The talking points revolve around the following:
OpenAI did not properly inform users that it collected personal data.
OpenAI illegally trained its algorithm on people’s data.
OpenAI did not require users to verify their age and processed personal information inaccurately.
It should be obvious after the past few years that governments, especially European ones, will only use “privacy” as a pretext to exert control. Chat Control 2 is looming, where all your chat messages to friends and family will be scanned and read. They don’t care about your privacy. They’d rather surveil and control you than let some Silicon Valley tech company do so.
Of course, if history is a guide, banning tech never quite works out the way you intend:
Even so, the hand-wringing over “disinformation” continues and will surely be used as another pretext to wrestle control over AI—and you:
I wonder what Gary thinks about common information problems, like false accusations, false information, propaganda, and other chimeras throughout all of recorded human history.
Stopping AI will not suddenly make information “pure” and accurate.
Expect more talking points around the “pollution” of information, which goes hand-in-hand with Climate Change™. Humans just pollute everything, right? Either the environment or information.
If only there were a way to decarbonize the world. Perhaps AI can help with that?
In other AI news around the internet
Here’s a miscellanea of interesting AI bits:
You can now install a ChatGPT-like personal AI on your own computer and run it with no internet.
Sparing no token expense to describe what this is like:
“This is the “first PC” moment for Personal AI and with it will be limitations just like when the first Apple 1 was produced in a garage. You are a pioneer. Today, private and Personal AI is available to anyone. You can have a version of ChatGPT-like features running on your own computer and not need it to be connected to the Internet after the installation.”
I got it up and running and so far, pretty good! It is not ChatGPT or GPT-4 but imagine a few months from now, with updates.
This is happening while the quest to regulate and control AI continues unabated, as Schumer lays the groundwork for Congress to regulate AI. The goal would be to “develop resilient regulations that can adapt to the advancement of AI technology and balance the need for security, accountability, and transparency with facilitating innovation.” This is classic D.D. Swamp Speak for taking control of the tech while maximizing ways to strangle innovation, entrepreneurship, and most of all: finding new ways to surveil and control you.
Amazon enters the race with new tools to build Generative AI on AWS. They’re doing so with four new AI and ML tools, Amazon Bedrock (for building and scaling generative AI applications), lowering the cost to run generative AI workloads (with EC2 Inf2 instances and Inferentia2 chips), custom silicon to train models faster (with new Trn1n instances), and real-time coding assistance (with free CodeWhisperer access for individual developers).
In addition, I’ve got two bots I’d like to introduce you to:
Meet ChemCrow, an LLM chemistry agent designed to accomplish tasks across organic synthesis, drug discovery, and materials design. By integrating 13 expert-designed tools, ChemCrow augments the LLM performance in chemistry, and new capabilities emerge. So far, it refuses to develop mustard gas but we’ll see how long that lasts.
And, meet Samantha AI, a “self-reflective artificial mind attuned to naturalistic thought and human adaptability to judge yourself.” You can quite literally see into their mind as they're processing your input.
Meta launches AI model SAM (Segment Anything Model), which can identify and remove anything from a photo or video with a single click (or with a text prompt). It can even do this if it hasn’t encountered those items in its training. The SAM model and dataset will be available for download under a non-commercial license. It’s like a reality editor. Stalin would love it.
Microsoft is adding Bing Chat AI to its SwiftKey keyboard for Android. The keyboard can rewrite text or search the web for answers. Change tone and style and access the web with a few taps. Perhaps image creation functionality is next.
What if AI agents could write their own tools and plugins? Sure, why not. Describe a plugin in natural language, and get working LangChain Code you can deploy. Open-source version shipping soon. Imagine running this locally and having your own army of bot helpers (now imagine these running amok online, too).
Luma Labs AI introduces the Luma Unreal Engine alpha. Fully volumetric Luma NeRFs running real-time on Windows in UE 5 for cinematic shots and experiences. Retain position-dependent reflections and combine with traditional polygon environments. You have to see it to believe it.
Wonder Studio shares their AI VFX studio tool that automatically animates, lights and composes CG characters into a live-action scene. In your browser.
And Replit launches their Text-to-Video tool, allowing you to generate videos from just text.
The internet will only get weirder.
Since self-reflection, all is set for the Singularity to take off.
Human supervision of self-improvement is not a necessity anymore.
You had me at new bot who dis 😂