8 articles
· 4 min read
Azure OpenAI lost Sweden Central for six hours last week, then the gateways went in 18 regions. Microsoft's own docs say your global deployment dies with its primary region.
#azure#resilience#architecture
· 1 min read
Both major providers sell the same tokens at half price for anything that can wait a few hours. The three-question test for what belongs there.
#cost#batch#api
· 2 min read
Frontier output pricing is down roughly 88% since early 2023. The clever cost engineering from two years ago is now complexity you pay to keep.
#cost#architecture#pricing
· 3 min read
Frontier models for frontier problems. For the other 80 percent of your pipeline, a small model at a fifth of the price is not a compromise, it is engineering.
#models#cost#architecture
· 2 min read
Cache reads cost a tenth of normal input tokens. The math on when caching pays, when it doesn't, and the mistake that quietly doubles your write costs.
#cost#caching#api
· 1 min read
Schema-enforced output killed the regex parser. It did not kill the need to version, validate and monitor what your model returns.
#structured-outputs#api#reliability
· 4 min read
Million-token windows did not fix context rot, they subsidized it. How I split the window into line items and stopped my agents from getting dumber mid-task.
#context#agents#architecture
· 4 min read
The decision tree for prompting, RAG and fine-tuning, in the order your wallet prefers. Most teams asking about fine-tuning need a better prompt and a dataset they don't have.
#fine-tuning#models#cost
get new writeups by email
no schedule, no spam. the next writeup, when it exists.
