AI
5 items · 5 articles in this topic
Low-key AI day: local inference, compression, benchmark skepticism and agent infrastructure. The claims are mostly unverified project pages.
- 1
The llama-modchip project claims to run a 284B DeepSeek model on 64GB RAM without a GPU. The summary gives no speed or quality figures.
Hacker News (front page) ↗10 min read
- 2
MaxxBench researchers question whether a claimed DeepSWE benchmark score for an AI model is valid.
Hacker News (front page) ↗9 min read
- 3
A Show HN project claims to compress a 31B video and audio diffusion model to 14B parameters with INT8 and six sampling steps.
Hacker News (front page) ↗12 min read
- 4
A LoreKit blog post argues AI agent memory needs a standard open format, comparable to what OpenTelemetry did for telemetry.
Hacker News (front page) ↗7 min read
- 5
Fireworks argues that automatically routing queries across multiple AI models through managed inference saves money. The claim comes from a vendor that sells such inference.
The Register ↗5 min read