Attention — Sep 20, 2026
What caught my attention on Sep 20 (6 items).
JEV Is Just a Text Classifier…
Prompt Engineering (Article:00:47) - Published 2026-09-18 18:00 - Discovered 2026-09-20 00:47
The video argues that JEV, despite its hype, is not a brand-new model category but a fast, instruction-following classifier built on ideas that date back to BERT-era single-pass encoders and 2019 zero-shot natural-language-inference pipelines. The presenter frames frontier LLMs as slow, deliberate reasoners, while many software tasks need quick classification. JEV accepts raw text and typed questi…
JEV: RAG Reranking for 1/10th the Cost
Prompt Engineering (Article:09:34) - Published 2026-09-19 18:00 - Discovered 2026-09-20 09:34
The video explains how JEV, a model from TypeSafe AI, can be used as a low-cost, instruction-following reranker inside retrieval-augmented generation pipelines. The presenter argues that RAG systems often contain several decision points—such as reranking, filtering, and citation selection—where a model must choose which retrieved chunks are most useful. Traditional cross-encoder rerankers are fast…
Technopolitics
Podcast – Cory Doctorow’s craphound.com (Podcast:16:24) - Published 2026-09-20 04:56 - Discovered 2026-09-20 16:24
In this episode, Cory Doctorow opens with a detailed tour schedule, mentioning upcoming talks in Berkeley, Cambridge, Brighton, Notre Dame, Calgary, Winnipeg, Vancouver, Ottawa, Kilkenny, and Montreal, as well as plans for his graphic novel Unauthorized Bread. He also shares personal updates, including buying a Framework 13 Pro laptop, praising its repairability and touchscreen while complaining a…
Spec-driven vibe coding in action
NoPlaceLikeLocalhost (Article:18:23) - Published 2026-09-17 18:00 - Discovered 2026-09-20 18:23
In this episode, the creator demonstrates his “spec-driven vibe coding” workflow, a hybrid approach that lets an LLM write all the code while requiring every implementation to be based on a carefully refined specification. Starting a new project called Dust to Dominion, an ambitious Asteroids-inspired mining strategy game, he begins with only documentation: a project overview, application conf…
Running AI group chat in very low VRAM with LuxTTS
NoPlaceLikeLocalhost (Article:18:25) - Published 2026-09-15 18:00 - Discovered 2026-09-20 18:25
In this video, the creator tests whether the TalkWithMe AI group chat application can run on a very low-VRAM system, specifically 4GB. TalkWithMe combines speech-to-text, text-to-speech with cloned voices, and an LLM for group conversations. The host first demonstrates voice cloning by creating a Daenerys Targaryen persona, showing that reference audio strongly affects the clone’s tone: an angry c…
LLM Battle 8
NoPlaceLikeLocalhost (Article:18:26) - Published 2026-09-13 18:00 - Discovered 2026-09-20 18:26
In LLM Battle 8, the host tests three small local language models—Spark-X2.5-4B, Ornith-1.5-9B, and IBM Granite-4.2-8B—on a practical coding task aimed at automating game-asset work. The task has two parts: first, write a Python script that rotates a spaceship sprite through 360 degrees and combines the frames into a single sprite sheet; second, build a viewer that displays the sprite sheet in a r…