How a big shift in training LLMs led to a capability explosion

In April 2023, a few weeks after the launch of GPT-4, the Internet went wild for two new software projects with the audacious names BabyAGI and AutoGPT.

“Over the past week, developers around the world have begun building ‘autonomous agents’ that work with large language models (LLMs) such as OpenAI’s GPT-4 to solve complex problems,” Mark Sullivan wrote for Fast Company. “Autonomous agents can already perform tasks as varied as conducting web research, writing code, and creating to-do lists.”

BabyAGI and AutoGPT repeatedly prompted GPT-4 in an effort to elicit agent-like behavior. The first prompt would give GPT-4 a goal (like “create a 7-day meal plan for me”) and ask it to come up with a to-do list (it might generate items like “Research healthy meal plans,” “plan meals for the week,” and “write the recipes for each dinner in diet.txt”).

Then these frameworks would have GPT-4 tackle one step at a time. Their creators hoped that invoking GPT-4 in a loop like this would enable it to tackle projects that required many steps.

But after an initial wave of hype, it became clear that GPT-4 wasn’t up to the task. Most of the time, GPT-4 could come up with a reasonable list of tasks. And sometimes it was able to complete a few individual tasks. But the model struggled to stay focused.

Sometimes GPT-4 would make a small early mistake, fail to correct it, and then get more and more confused as it went along. One early review complained that BabyAGI “couldn’t seem to follow through on its list of tasks and kept changing task number one instead of moving on to task number two.”

By the end of 2023, most people had abandoned AutoGPT and BabyAGI. It seemed that LLMs were not yet capable of reliable multi-step reasoning.