Amazon's AI Memo Sends Mixed Signals to Engineers
Amazon made public this week a document on the role of the developer in the age of AI. Terk Terkowitz, a VP and Distinguished Engineer, wrote it. It first went to Amazon's software engineers and their managers in September.
Corey Quinn, who wrote about the memo for The Register, calls it a good document. He also reads it as an attempt to make the future of software engineering feel less bleak, at a time when many engineers fear for their careers.
What the memo says
The memo stresses customer outcomes and engineering judgement. It says that work that gains nothing from current AI tools should be documented, and it puts it this way: "If your work does not benefit from current AI tooling, that is a finding, not a failing."
It also offers a way to measure adoption. Look at the 90th percentile, the teams burning the most tokens, and check whether their delivery rose by the same amount.
The memo calls input metrics "deeply flawed". The next paragraph says to use them anyway, carefully. It then warns that people who do not adopt AI are falling behind.
Tokenmaxxing and a shut-down leaderboard
Quinn points to KiroRank, an employee-built leaderboard that ranked people by token usage. Amazon shut it down in May. People had created junk tasks to climb the ranks. The practice is now called "tokenmaxxing", and the memo names it as being at odds with "frugality".
Quinn says Meta did something similar with "Claudenomics" and 85,000 employees in the same week. Amazon also reportedly announced another 250 job cuts this week. Quinn says he cannot claim the cuts and the memo are related, but that every engineer who read the memo assumes they are.
Quinn's concern for managers
Quinn argues that the three positions in the memo, taken together, let a manager choose. A manager can cite "a finding, not a failing" when a strong engineer skips an AI tool. The same manager can cite "falling behind" when it is time to put someone on a performance plan.
He asks whether promotion and performance decisions will reward long-term outcomes or only what is cheap to measure. In his words, if you give someone a metric, they will turn it into a target.
Quinn wrote the column after two days at GitHub Universe. He says agents there were said to start roughly 40% of pull requests. Jay Parikh argued in his keynote that useful output means completed outcomes, such as features shipped and bugs fixed, rather than more code.
Quinn notes that GitHub's business model charges on consumption, not output. AWS likewise bills Bedrock and Kiro customers by the token. Inside Amazon, tokenmaxxing is a problem to be stamped out. Outside, in his words, it is the revenue plan.
Source: The Register