oMLX is more than an Apple Silicon inference server with continuous batching. Its 256-token cache blocks, SSD tier, model eviction, and fairness scheduler expose a familiar game-engine problem: kee...

Joe Wilson’s SQLite for Everything makes a compelling case for the database as a local function call and a file. For games, that file is the hard part: save synchronization, engine packaging, untru...

MoneyPrinterTurbo does not make its default videos by synthesizing pixels from a prompt. I read the v1.3.5 release, source, configuration, and API path to see a more useful truth: it is an open-sou...

I read Meituan’s LongCat-Video paper, weights, and inference code to understand its real idea: make text-to-video, image-to-video, and long continuation one condition-prefix problem, then make the ...

FreeToken is an edge-native MoE serving engine that runs DeepSeek-V4-Flash on a gaming desktop and GLM-5.2 on a single workstation GPU. I read the paper and the repo to find what it actually does d...

A Source Audit separates modern MCP metadata, legacy initialization and optional dual-era behavior before applying the roadmap’s stateless-server advice.

A LinkedIn post listed 12 concepts for mastering AI agents. I read every linked article, verified the claims against primary sources, and built the one layer none of the twelve names as a concept -...

Vectorize wired Hindsight memory into Omnigent, a meta-harness that wraps Claude Code, Codex, Cursor, OpenCode and Pi. I read the source to find out where the bridge actually lives — and found one ...

A deep dive into Q00/ouroboros - the spec-first Agent OS where a Socratic interview gates code behind an ambiguity score, the grading command never reaches the worker, and an evolutionary loop runs...