DeepSeek just made reading your prompts cheaper than thinking
DeepSeek's new open model charges less compute to read a prompt than to answer it. The 8B/16B split signals where model pricing is heading.
AI
The Least Intelligent Part of My AI Stack Is the Most Reliable
My most trusted pipeline step has zero machine learning in it. A Rust binary turns JSON into PNG at about 50ms per slide, and it replaced a headless browser that kept breaking my batch jobs.
llama.cpp's 42x prompt lookup speedup lives in a fork
A fork of llama.cpp made prompt lookup drafting up to 42x faster. The cache-side numbers matter more, and none of it is merged upstream yet.
Why the Self-Hosted Clone Matters More Than the Launch
Jev launched September 15. Eleven days later its open source clone had 27,645 GitHub stars. The clone is the story.
Tokens got two orders of magnitude cheaper. What still meters you?
Cost per task fell about 100x in a year. The interesting question now is what a billion cheap tokens buy, and which parts of your stack still charge per query.
Weekly Research: Automation Is Up, Vibe Coding Is Fading
ai workflow automation pinned at 100 on Google Trends this week, up from a quarterly average of 22 last year. Vibe coding stayed flat. What the numbers say.
Give your Hermes agent a memory that stays
Your agent forgets everything between sessions. Wire Uteke into Hermes and it starts keeping decisions, preferences, and project facts.