Descriptions:
Gergely Orosz, author of The Pragmatic Engineer newsletter, takes the stage at the AI Engineer conference to discuss a quietly spreading behavior at some of the world’s largest tech companies: ‘token maxing’—the practice of deliberately inflating AI usage metrics to look productive on internal leaderboards, regardless of whether the AI usage is actually generating value.
Orosz describes specific incidents at Meta, Microsoft, and Salesforce. At Meta, engineers were reportedly asking AI agents to summarize documentation they could read themselves, or running autonomous agents on throwaway tasks, just to avoid appearing in the bottom quartile of a company-wide token usage leaderboard—a leaderboard that was eventually shut down after press coverage, though the behavior reportedly continued. Salesforce introduced a $175/month minimum AI spend target. At Microsoft, similar leaderboard dynamics drove comparable gaming. Orosz draws a direct parallel to the ‘lines of code’ productivity metric era and warns the same incentive distortions are playing out at a faster pace and larger scale.
Beyond token maxing, Orosz offers a broader view of how the software engineer role is evolving: testing, DevOps, and increasingly product management are collapsing into a single generalist function, a trend AI is sharply accelerating. He argues that getting good at AI-assisted development is a slow, practice-dependent skill—not something that follows from understanding the underlying architecture—and that the engineers extracting the most value share a common trait: low ego and a willingness to abandon prior intuitions about how software gets built.
📺 Source: AI Engineer · Published April 21, 2026
🏷️ Format: Interview
