https://keepingaihonest.com/lane/token-conservation/ Ethical AI news: the rolling record of what AI costs people and the planet, and who's fighting back. en-us Sat, 31 Oct 2026 21:45:08 GMT 30 https://keepingaihonest.com/item/2026-10-31-ai-optimized-code-used-more-energy-than-human-code-in-42-of-44-setups/ https://keepingaihonest.com/item/2026-10-31-ai-optimized-code-used-more-energy-than-human-code-in-42-of-44-setups/ Sat, 31 Oct 2026 21:45:08 GMT A study accepted at the IEEE/ACM International Conference on Software Engineering (ICSE 2027) and posted to arXiv measured hardware-level energy as 11 open-source language models, each under four decoding settings, repeatedly rewrote code for speed on all 1,000 problems of the EffiBench benchmark. In 42 of the 44 model-and-setting configurations, the resulting code used more energy per problem than the human-written reference solution, with overheads from 4% to 886%; the two exceptions would need on the order of a million to ten million runs to pay back the energy spent generating them. Capping the number of output tokens cut inference energy by up to 69% against the default setting, and mixture-of-experts models used substantially less inference energy than dense models of similar size, the authors report. Why it matters: Faster is not greener. Here the machine spent energy writing code that then spends more energy every time it runs. Teams using AI to optimize code should measure joules, not just milliseconds, and the cheapest fix is the boring one: cap the tokens. Via arXiv - AI energy/carbon/water (https://arxiv.org/abs/2610.32891v1). https://keepingaihonest.com/item/2026-10-31-amazon-memo-tells-engineers-token-usage-is-a-flawed-measure-of-ai-work/ https://keepingaihonest.com/item/2026-10-31-amazon-memo-tells-engineers-token-usage-is-a-flawed-measure-of-ai-work/ Sat, 31 Oct 2026 21:45:08 GMT Amazon published a memo to its software engineers on Oct. 27 calling token usage, AI tool adoption and counts of code changes "deeply flawed" measures of AI work, warning that using them as targets could raise costs ("tokenmaxxing is at odds with frugality") and adding: "If your work does not benefit from current AI tooling, that is a finding, not a failing." The memo, posted by vice president Terk Terkowitz, still says those metrics "correlate well with the outputs that matter," calls for using them "with care and judgement," and tells managers to study the teams at the 10th and 90th percentiles of AI adoption. Per CNBC, the memo circulated internally in September, Amazon shut down an internal leaderboard in May after finding employees were tokenmaxxing (using AI as much as possible with little regard for output) to climb it, and a spokesperson said Amazon has not capped employees' token use. Why it matters: Amazon names the waste and keeps the metric that drives it. As long as token counts "correlate" with performance, the 90th percentile reads like the goal. The line worth copying is the permission to stop: "a finding, not a failing." Every employer tracking AI usage should put that in writing. Via Amazon (https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai). https://keepingaihonest.com/item/2026-10-31-bank-of-america-and-capital-one-route-ai-work-to-smaller-models-to/ https://keepingaihonest.com/item/2026-10-31-bank-of-america-and-capital-one-route-ai-work-to-smaller-models-to/ Sat, 31 Oct 2026 21:45:08 GMT Bank of America sends AI queries through Orchestra, a routing layer that uses a small classifier model to pick the most cost-effective model for each query, and chief technology and information officer Hari Gopalkrishnan called routing to lower-cost models an important cost tactic, CIO Dive reported from the Evident AI Symposium in New York on Oct. 22. Capital One distilled its fine-tuned open-weight models into smaller ones for Chat Concierge, its car-buying assistant, which chief enterprise AI officer Prem Natarajan said cut latency fivefold and raised customer engagement for dealerships by 50%. Gopalkrishnan also said Bank of America plans to double its AI budget next year. Why it matters: Routing is conservation without the sermon: send the easy question to the small model. But the savings here are being spent, not banked. Bank of America is doubling its AI budget anyway. Efficiency that funds more volume is a discount, not a cut. Via CIO Dive (https://www.ciodive.com/news/bank-of-america-capital-one-manage-ai-costs/873112/). https://keepingaihonest.com/item/2026-10-31-tokwatt-energy-database-one-ai-coding-agent-task-uses-440-times-a/ https://keepingaihonest.com/item/2026-10-31-tokwatt-energy-database-one-ai-coding-agent-task-uses-440-times-a/ Sat, 31 Oct 2026 21:45:08 GMT TokWatt, an open database posted to arXiv by Abderaouf Bahi of Chadli Bendjedid University in Algeria, collects 134 published estimates of AI inference energy from 13 sources (2023–2026) and rescales them to a common boundary covering the server plus facility overhead such as cooling. Its reference values are 0.34 watt-hours for a short chat answer (range 0.16–0.99 Wh), 4.7 Wh for a reasoning answer, 124 Wh for a short video clip and 150 Wh for one task handed to a coding agent, about 440 times a short chat answer. Five independent estimates for default production chat systems such as ChatGPT and Gemini agree within a factor of 1.8; the author notes that 12 of the 19 reference values rest on a single source and that the values leave out idle capacity and the energy used to make the hardware. Why it matters: The quick chat answer was never the big cost. The agent is. One delegated task here costs what hundreds of short questions do. Builders should count the loop, not the prompt, and users deserve that number before they hand off the job. Via arXiv - cs.CY (AI) (https://arxiv.org/abs/2610.39846v1).