Let me begin with a confession: it’s becoming increasingly challenging to keep up with the rapid flow of experimentation, innovation, and research results emerging around large language models (LLMs). As I delve into the notes accompanying this video, I’ll mention a few examples that have caught my attention.
I’ve talked about GPTs and new services that enable LLMs to generate their own tasks. From a video, I published on April 6:
As LLMs, chatbots, and instances of these models connect to other systems on the internet, a new complex system arises. This has led to the excitement around AutoGPT projects. These connect GPT systems to other LLMs and APIs to coordinate tasks and execute them in a more generalised way among the connected systems. This dramatically expands the capabilities of these systems and creates a “super system” that can perform a broader range of tasks.
We’ve been experimenting with these at Exponential View, inputting high-level objectives and observing how they recursively work to achieve their goals. It’s strikingly similar to the process a human would go through when breaking down complex tasks. As AutoGPTs mature, we’re seeing impressive results, even though people are still figuring out how to use them effectively.
There are other developments worth our attention. I recently read a paper discussing how scientists used LLMs to help decode the language of sperm whales. This fascinating development could contribute to a fundamental shift in our understanding of communication among advanced mammals — although we are nowhere near to inventing a warp drive, we may achieve cross-species understanding similar to that portrayed in Star Trek IV: The Voyage Home sooner than 2286!
LLMs also have their limitations, such as their restricted context windows. Some have found ways to work around this by having GPT-4 create a compressed language that encodes previous conversations, allowing for extended interactions. GPTs may also help us rethink how we write software: moving from rigid, circuit-based applications to understanding-based applications of almost unlimited input and dynamic properties.
It’s absolutely remarkable to witness the growing capabilities as we chain these systems together or embed them in large collections of software.
The rapid advancements in LLM technology make me think about historical paradigm shifts, like the Copernican Revolution or the Gutenberg Moment, which forced us to change our perspectives and systems. With LLMs challenging copyright, privacy, and other societal norms, it begs the question: are we the priests or the scientists of our current worldview?
