For greater than a decade, Stack Overflow has been the go-to discussion board for builders searching for solutions to coding questions. At its peak within the mid-2010s, the positioning noticed greater than 200,000 new questions every month. These days are gone. For the reason that arrival of AI assistants resembling ChatGPT, Stack Overflow’s utilization has fallen off a cliff. The truth is, from March 2023 to March 2024, the quantity of recent questions dropped from ~87,000 to ~58,800, a 32% decline in only one yr. By the tip of 2024, month-to-month questions have been down 40% yr over yr, hitting ranges not seen since 2009. Builders, who beforehand relied on group validation for correct solutions, now more and more belief an algorithm’s replies, that are instant and assured (but not at all times right).
This shift creates an ironic drawback: These massive language fashions (LLMs) owe their spectacular information bases to human-generated content material from platforms like Stack Overflow, creating their very own cannibal’s dilemma of kinds. I’ve written at size about this drawback, however the true query isn’t about how you can recapture what we’ve misplaced, however how you can transfer into the longer term. As these communities fade, what is going to LLMs use for future coaching, and the way will builders make sure the accuracy and reliability of the solutions these instruments present?
Paradoxically, the AI assistants displacing Stack Overflow are themselves merchandise of Stack Overflow’s success. LLMs resembling ChatGPT turned “good” by coaching on huge troves of human-generated content material—a lot of it scraped from websites like Stack Overflow. Each well-explained reply, each debated finest follow, and each upvoted code snippet on these boards turned grist for the AI mill. In a really actual sense, the information that flows from group Q&A platforms is the bedrock on which future iterations of LLMs are constructed. Early fashions have been educated on information units that included hundreds of thousands of Stack Overflow posts capturing the nuances of programming issues and their options.
