Microsoft study shows AI copilot development can be overwhelming

[ad_1]


summary
Summary

A study by researchers at Microsoft Research has identified problems developers face when building AI copilots.

As more companies deploy AI copilots powered by large language models (LLMs) to assist users with tasks in applications such as Word, Excel, programming, and image and video creation, software developers are entering uncharted territory in integrating these AI technologies.

Microsoft researchers interviewed 26 professional software developers responsible for copilot development at a variety of companies. Their key finding is that development processes and tools have not kept pace with the challenges and scope of AI application development.

The development process for a copilot follows a rough sequence of exploration, implementation, evaluation, and productization. But because of the unpredictable nature of AI, the process is “messy and iterative,” the researchers say.

Ad

Ad

The “messy and iterative” development process for AI copilots. | Image: Henley et al.

Developers must identify relevant use cases, assess feasibility with different technologies, and ultimately deliver a product to real users – each of which presents its own set of challenges. The study categorized the pain points into six areas:

1. Prompt engineering is time-consuming and requires extensive trial and error to balance context and token count. It is “more art than science” and models are “very fragile” with “a million ways you can effect it.”

2. Orchestrating multiple data sources and prompts to understand user intent and control workflows is complex and error-prone.

3. Testing is crucial but tedious because of LLM unpredictability. Developers run tests repeatedly looking for matches or creating expensive benchmarks.

4. There are no best practices for working with LLMs. Developers rely on Twitter and papers in a rapidly evolving field that demands constant rethinking.

Recommendation

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top