Sparks of AGI: AutoGPT

AI that runs without human intervention

April 14, 2023

Good Morning!

Today’s article is brought to you by: 50% human and 50% AI 

Market Update

📊 Stocks jumped Thursday after the PPI (Producers Price Index) - a measure of prices paid to producers - fell 0.5% from February to March, reflecting easing price pressures and moderating inflation conditions.

📈 Interest-rate futures suggest investors are betting a 25bps rate hike in May will be the last raise for the Federal Reserve.

🏦 JP Morgan, Wells Fargo, and Citi report Q123 Earnings tomorrow - investors are hoping to get more color into the behind-the-scenes of the Banking Crisis - specifically on consumer trends & what lending availability might look like going forward.

Banger Tweet of the Day

Inflation down good, recession bad.

Top Stories

🤖 Amazon Joins Microsoft, Google, in AI Race Spurred by ChatGPT (WSJ)

  • AWS is taking a “picks & shovels” approach to AI products by focusing on developer tools like access to LLMs and custom-made chips to run AI software more efficiently.

  • AWS will sell access to multiple AI models while also rolling out its own - Titan - built more for handling internal business data.

  • They also plan to release CodeWhisperer - a programming assistant that will compete directly with Microsoft’s Copilot.

🤝 Walmart sells Bonobos to Express in $75 million deal (CNBC)

  • Quite the flop - Walmart originally purchased Bonobos for $310 million in 2017 while trying to grow its online presence. Tough loss, but we’re talkin’ pennies to Walmart.

🐤 Twitter Changes its Name to X Corp, Moves to Nevada, and Rolls out plan to buy Stocks (Barron’s)

  • Elon officially changed the legal name of Twitter to X corp - the name of his first major company & longtime vision of a “Super-App” that does everything from messaging to payments to social to investing.

  • His plans are coming along somewhat - Twitter just announced a partnership with eToro to allow users to trade stocks & crypto within the app.

  • The company also relocated its corporate structure from Delaware to Nevada - which reportedly grants more discretion and protection to company management.

Sparks of AGI: AutoGPT

AutoGPT is the Next Big Thing taking place in AI right now. So much so, the former head of Tesla AI called it “the next frontier of prompt engineering.”

So what is it?

In my words: Autonomous GPT, short-named AutoGPT, is a program that interacts back & forth with GPT-4 (think ChatGPT) fully on its own without user inputs.

The model combines:

✍️ GPT-4 to generate text

💻️  Internet access to search Google searches & gather info

🧠 long-term memory for recall

💾 file storage for outputs

…all to work on its own without a human intermediary.

It’s still early days - the script often trips out or gets confused and rendered worthless, but when I played around with it I had moments where I felt like I was interacting with a human-level computer.

Bring in AutoGPT

To help explain, I brought in AutoGPT himself. herself? itself.

I asked AutoGPT to write an article describing itself (more details at the bottom).

Here’s what AutoGPT had to say (formatted and slightly edited for brevity):

What is AutoGPT?

AutoGPT works by leveraging the power of GPT-4, a state-of-the-art language model created by OpenAI, by creating AI agents that utilize the language model to perform various tasks on the internet.

Dall-E’s Idea of “AI Agents”

How it Works

  • When the user inputs a task for AutoGPT to perform, the application breaks it down into a series of smaller tasks that the AI agent can execute autonomously.

  • The AI agent then generates a series of "thoughts," "reasoning," and "criticisms" to explain to the user what it is doing and why. This makes it easy for users to understand how the AI is arriving at its conclusions and allows them to intervene if necessary.

Interesting Use Cases

AutoGPT has potential in a range of fun and interesting use cases. Here are a few examples:

  • Songwriting: AutoGPT can be trained on the lyrics of different genres and even mimic a certain artist's style. This can help aspiring songwriters and music producers come up with new and unique lyrics.

  •  Screenplay writing: AutoGPT can help screenwriters create complex stories or even write short dialogue scenes by generating character interactions.

  • Content creation: AutoGPT can create high-quality news articles, blog posts, and marketing content faster and more efficiently than humans. This can help businesses save time and money while increasing their output.

  • Academic & Market research: AutoGPT can assist researchers in summarizing large volumes of research data into easy-to-read summaries. This can save researchers time and effort while improving their output.

My Interjection: It’s not there yet by any means, but it’s not hard to imagine a near future where this agent can run its own business. It will source workers, generate content, write software, and sell products all on its own. This is the killer use case, but we’re not really there yet. AutoGPT is still only as smart as the underlying tech. ok, back to AutoGPT.

Overall, AutoGPT's ability to simplify complex tasks can save individuals and businesses time, effort, and money, while improving overall output and results.

End Script.

My Thoughts & Experiences with AutoGPT

My thoughts: 

Not too shabby! To be clear, my use-case to have it write an article isn’t that different from something GPT-4 can do on its own. There are much cooler use-cases (like my interjection above).

The big improvement comes from its ability to string together multiple prompts, allowing the script to review and iterate upon itself, improving the output and accomplishing tasks with no human supervision.

Newsletter-GPT / How I Got That Article

I downloaded AutoGPT here and ran it on my local device. Once I ran it I gave it the following parameters:

  • Name: “Newsletter-GPT”

  • Role: an AI designed to write a short essay on AutoGPT

  • Goals: Craft an article on AutoGPT, comment on what AutoGPT is, why it’s a big deal, how it works, and include interesting use cases.

AutoGPT then went through the following steps:

Task 1: Ran a google search to better understand itself.

Task 2: AutoGPT then wrote a rough draft of the article. It was pretty good, but didn’t fulfill the 5 goals listed above.

Task 3: Save the article to a .txt file on my desktop.

Task 4: AutoGPT then prompted itself to proofread the essay for clarity and overall flow.

Task 5: AutoGPT then began breaking up the 5 goals into individual tasks and writing articles on each one, updating & saving the local file along the way.

End Result: I had to turn it off after a few runs as it got caught in a recursive loop of writing & proofreading the same few things over and over. However, the end result was pretty good and watching it work was a thing of beauty.

It was actually more sensational for me than using ChatGPT for the first time. Watching the program string together multiple prompts based on its own instruction and deploy agents to fulfill tasks felt like I was working with AGI_v1.0.

Today’s Challenge: Share this article with a few friends!

Much Love,

Reply

or to participate.