Igniting the Spark of Discovery
Exploring the Potential of AGI and AI Agents in a New Era of AI
Prompt engineering is simply storytelling and knowing how to ask the right questions.
It all started in November 2022
For a few weeks I had been seeing people were using some new AI to generate pictures. I got into the Midjourney rabbit-hole pretty quickly thereafter.
As someone who saw the potential of the Internet in 1992, I saw something that smelled interesting. After a few hours of research I found out that not only were these images purely generated by AI, but - more importantly - there was an open source version that I could download and run on my own machine - completely offline.
I purchased a new laptop within hours of discovering the offline capability with this tech. I ended up with an Nvidia 8GB GPU that I still use today. I was running Automatic1111 / Stable diffusion that evening. It seemed like pure magic to me as I am a computer scientist but not a machine learning expert. The offline aspect was important to me as I knew this technology would soon be able to be in everyone’s hands - “these things were going to be running on our phones one day” was my immediate thought.
Days later ChatGPT was released, and that changed everything again. I found it fascinating, but it still seemed like a toy… (a Stochastic parrot not capable of reasoning) not capable of autonomous agency due to its lack of decision making reliability.
The Sparks, March 2023
But then, there was something that shook me to my core. Something that sparked my interest and passion in this generative field. It began my chase towards intelligent systems that can solve complex tasks.
I was having my morning coffee, checking in with AI news, when I came across a research paper review by an academically leaning YouTuber. It was the Sparks of AGI research paper (https://arxiv.org/abs/2303.12712) I pulled the paper up and started reading it immediately.
I had to re-read the paper just to make sure I wasn’t seeing things. GPT-4 was passing Senior Developer interview questionnaires for working at Amazon.
You could see it learn to reason by reviewing some of the examples.
The Theory of Mind example was very telling for me as the kind of advancement and the types of questions you can give the models. It could role-play, which is an important part of agentic tech currently.
Below is Sébastien Bubeck one of the authors of the paper re-telling and explaining the paper, but more importantly what it was like to witness history. What the reader here needs to know is that most people in the industry were surprised and shocked at the capabilities of GPT-4 when this paper was released. As I read this paper, it was apparent to me that there was something extremely exciting - and a bit scary - in the capabilities that were being demonstrated in that paper.
Research Paper-palooza 2023
The rest of 2023 was a torrent of research papers on the capabilities of large language models, but most particularly it focused on GPT-4, for good reason: at least a dozen reasons outlined in that paper. The paper showed that there was more to discover, and of course, internally as soon as they discovered this at Microsoft, they went all-out starting separate research efforts into agent technology that turned into the open source project AutoGen which came out as just a Proof of Concept based on a research paper.
Less than a year later, the AutoGen team is still conducting research in this area as a means of solving more complex tasks. One of the members provided a formal update.
The research in this space is the primary topic of this newsletter.
These papers are important as they shape why prompt engineering, story telling, and asking questions more succinctly has become a focus for so many. Stay tuned!