All pieces
9 pieces
-
Why doesn't HNSW have to check every vector? ›
The simplest exact search measures every stored vector, so the bill grows with the index. Step through the eight moves that let HNSW skip most of the measuring.
Aug 10, 2026
-
Why AI benchmarks that lasted four years now die in two years ›
The flagship 2020 AI exam lasted four years before models passed 90%. The 2021 and 2023 tests fell in 17 and 24 months. Why benchmark scores now expire.
Jul 30, 2026
-
Why the same AI model gets three different scores ›
Claude Opus 4.5 scored 45.9, 50.2, and 55.4 on the same benchmark, and only the harness changed. Why leaderboards partly rank the software wrapper around the model, and what to pay for first.
Jul 23, 2026
-
Why the flat $20 AI subscription broke in 2026 ›
The top 10% of users on flat AI plans burn 70% of the tokens. How the $20 subscription broke in spring 2026, and the hybrid price that replaced it.
Jul 16, 2026
-
Why LLM context windows lose accuracy long before they fill ›
Models pass the needle test at 99 percent, yet 10 of 12 lose half their accuracy by 32K tokens. Where the usable context window ends and how to measure yours.
Jul 10, 2026
-
Why doubling an agent's steps quadruples its bill ›
An AI agent re-reads its whole history on every step, so cost grows with the square of the steps. Why doubling an agent's steps roughly quadruples its token bill.
Jul 1, 2026
-
Why thinking longer makes some AI answers worse ›
A reasoning model scored 36 out of 100 on a task it got 42 right with no thinking at all. Why test-time compute has a peak, and how to set the budget.
Jun 25, 2026
-
Why AI models top the benchmark and fail the job ›
An agent that clears 95% of steps finishes a 20-step job only 36% of the time. Why benchmark scores collapse on longer real work, and where to add checks.
Jun 18, 2026
-
Why AI products leak half their revenue in a year ›
AI-native software keeps about 48% of revenue after a year versus 82% for traditional SaaS, and the gap tracks price. Retention benchmarks by price tier.
Jun 11, 2026