Amazon sends chills through its workforce with mixed-message memo about AI use
B · Useful, limited A column by Corey Quinn in The Register on an Amazon engineering memo about AI use; it quotes the memo and adds the KiroRank episode and a sharp reading of mixed incentives, but it is opinion built on excerpts, with some facts marked as reported.
Summary based on the full text·Original: about 5 minutes to read
In short
A column by Corey Quinn (The Register, 31 October 2026) on a memo by Amazon VP and Distinguished Engineer Terk Terkowitz about the developer's role with AI, first sent internally in September and since published by Amazon. Quinn likes much of it: it puts customer outcomes and engineering judgement first, accepts that some work does not benefit from AI, and suggests checking whether the heaviest token users actually deliver more. His objection is that the memo also says input metrics are flawed but usable and warns that non-adopters fall behind, which lets managers justify either reward or punishment. He ends on a conflict: Amazon fights wasteful token use internally while selling tokens to customers.
Key numbers
- 90th percentileAI-adoption level at which the memo suggests checking whether the heaviest token users actually delivered more
- 250job cuts Amazon reportedly announced the same week (unconfirmed link to the memo)
- roughly 40%share of pull requests GitHub said at its Universe conference are now started by agents
As reported by the source.
English·Web page·opens theregister.com
What it reports
- The memo says: "If your work does not benefit from current AI tooling, that is a finding, not a failing", and asks engineers to document such cases and raise them with managers.
- It calls input metrics "deeply flawed", then allows careful use of them, then warns that non-adopters are falling behind. Quinn reads these three positions as cover for inconsistent performance management.
- It recommends looking at the 90th percentile of AI adoption, the teams burning the most tokens, and checking whether delivery rose to match.
- In May, Amazon shut down KiroRank, a staff-built leaderboard ranking people by token use, after people created junk tasks to climb it ("tokenmaxxing"); the memo calls this at odds with frugality. Quinn says Meta ended a similar ranking that week.
- Quinn notes Amazon reportedly announced another 250 job cuts that week and says engineers will link the two, though he cannot say they are related.
- At GitHub Universe, agents were said to start roughly 40% of pull requests, while the keynote argued that useful output means completed outcomes; Quinn points out that consumption billing does not reward outcomes.
How they got it
Opinion column based on the published memo, press reports and the author's attendance at GitHub Universe.
Limits
- Commentary; the memo is quoted only in short excerpts and the reading of motives is the author's.
- Job-cut and Meta details are relayed, not sourced in the column.
- The author is a known critic of AWS pricing.
Why it matters for companies
Many firms now track AI use through token counts, seat activity or leaderboards. The KiroRank episode shows the predictable result: once usage is a target, people generate usage. The memo's better idea is to compare the heaviest users' delivery with everyone else's, and to treat a documented case of AI not helping as useful information. Quinn's test is the practical one for leaders: check what happens to the engineer who reports that AI made a task worse. If that person's review suffers, the stated policy does not matter.
Summary based on the full text. Read the original for the full argument and data.Read the original on theregister.com