Published elsewhere
Each link goes to the original, which stays the canonical version. The research itself lives in Papers.
Articles
Long-form writeups published elsewhere.
Posts
Shorter arguments and threads.
- LinkedInAEQ Was Never a Cost MetricAnswering PointFive's study of 2,908 paid Claude Code sessions, which found that stripping context to save tokens pushed bills up 7 to 46 percent. AEQ holds the answer constant and asks what the architecture needed to produce it: 345 tokens against 1,615 on the same question, same model.Read on LinkedIn
- XSwept My AI Stack After the LiteLLM BreachA supply-chain compromise that reached 2,488 organizations and 153GB. The stack came through clean, and not by luck: no gateway was ever adopted, groq and google-genai run direct. Abstracting 100+ providers was the whole value proposition, and exactly what made it worth compromising.Read on X
- XQuantization Did Not Degrade Capability, and Size Did Not Order ItA 4-bit quantized 9.7B model on a 16 GB consumer desktop passed 3 of 5 workload classes at zero marginal compute, matching a frontier API model on retrieval, synthesis and quantitative derivation. The 12B passed fewer. Parameter count did not predict adequacy, and the measurement cost about $0.02 a cell.Read on X
- LinkedInAgent Efficiency Quotient: A Better Benchmark Than Cost-Per-TokenAnswering Chen Goldberg in CIO Dive one layer up, in the architecture. The bloated build spent 5.51x the tokens on OpenAI and 2.04x on Anthropic for the identical answer, and agreed with itself three times in five at temperature 0 where the lean build agreed five in five.Read on LinkedIn
- LinkedInCheaper Models Won't Save Bad ArchitectureThree architectures, same model, same query, a 32% to 54% spread that comes to fractions of a penny. The argument moves off cost and onto attention: latency, reliability and capability do not get cheaper when tokens do.Read on LinkedIn
- XHow Efficient Is Your AI Agent? The Metric Nobody's Tracking YetWritten while Nadella, Musk and Blundin were all calling the end of SaaS and celebrating token deflation. Three agent architectures, same model, same query, and a cost difference that came out negligible.Read on X
- LinkedInLast Week I Asked: Is SaaS Dying, or Just Dumb SaaS?Follow-up drawing on the replies to the first post, narrowing the claim to something measurable.Read on LinkedIn
- LinkedInIs SaaS Dying, or Is It Just Dumb SaaS?Opening argument on where per-seat workflow software is actually exposed, and where it is not.Read on LinkedIn
Something missing here? Send it to michael@bucketbranch.com.