Agent Costs Hinge on KV Cache; Failed Subtitle Edit Sparks AGI Skepticism

Chris Alexiuk notes that agent running costs are dominated by KV cache hits and misses rather than compute, with DeepSeek compressing KV cache to about 890 bytes per token; meanwhile, a user's agent still misplaced subtitles despite explicit instructions, prompting complaints that AGI is not here yet.

2026-09-12 ~ 2026-09-12 · 2 related posts