Maximizing Efficiency with Human-AI Collaboration | Leveraging Each Other's Strengths

There has been a tremendous about of exciting new product innovation in the past year, giving everyone a glimpse of the world to come when we all have potent models at our side. It's no longer difficult to imagine the role models like GPT-3 will play in the world shortly, but it still needs to be determined precisely how our relationship with models will evolve.

As we have grown PromptLoop - it has become clear that specific patterns are emerging with how humans can use language models to propel their work without placing them in the wrong role.

Where do the models win?#

Let's start with what Large Language Models (LLMs) like GPT-3 are and the models that power Promptloop is good at in 2022. These models are trained on billions or lines of text to generate a coherent view of language(s) and how words typically go together. They are fast and accurate at stitching together text or pulling out essential words or themes, making them perfect for PromptLoop users.

Where do they fall short?#

In late 2022, the model's most vital skills also feed into its weaknesses. Language models trained on open internet data also cause some of their shortcomings. It's scarce to find a comment or post on the internet saying, "I don't know ."Predictably, the models rarely return uncertain answers and prefer to repeat confident but incorrect statements or responses. "Out of the box" models often present challenges with error correction and misleading information. A lot of what we work on with PromptLoop is correcting for this tendency behind the scenes. Logical explanations are similarly tricky to produce.

Backfilling and the Human and AI loop#

Lack of logical prowess does not limit the usefulness of these models, especially when we can use the best logical engines around - well-tuned spreadsheet models.

The teams getting the most out of PromptLoop are chaining our formula with error and labeling functions that provide the best of both worlds. They use language models when human reading and writing would otherwise be needed and check and process the results with the formulas that already power their spreadsheets.