Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU
Llama-modchip aims to run a 284-billion-parameter DeepSeek model using 64GB of RAM without a GPU. The provided information does not explain its technical approach. (github.com)
Ensemble is both an AI coding agent and an executable teaching book about building one. Its central claim is that useful specifications should be earned through development, rather than written in exhaustive detail before coding or reconstructed after the fact. The developer and AI first discuss a design, then draft only the next chapter’s summary and exercises. They pair-program the implementation, recording uncertainties, corrections, decisions, and tradeoffs. Once the code passes exercise graders, those notes become a chapter that explains how and why to build the system. A fresh agent then attempts the exercises from the chapter, revealing gaps that can be corrected. The resulting specification aims to transfer the judgment needed to reproduce the implementation, not merely list features.
The approach is embodied in successive editions: Edition 2 used the book to build Edition 3, whose application is written in C# 14 on .NET 10 and developed within Ensemble. The repository preserves chapter solutions, graders, implementation evidence, and earlier editions; it also distinguishes current application code from Go-based grading and publishing tools. Chapters document their own verification limits, and the README gives explicit build, test, and grading commands, including the requirement for a .NET 10 SDK. The project’s broader lesson is that developers gain expertise through guided implementation, so documentation should teach the construction process and reasoning rather than serve as a blueprint assumed to be complete in advance.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Llama-modchip aims to run a 284-billion-parameter DeepSeek model using 64GB of RAM without a GPU. The provided information does not explain its technical approach. (github.com)
A GNU maintainer’s controversial take prompted debate about AI-generated code in open-source projects. Commenters discussed project leaders who use AI tools while distancing themselves from code ownership. (social.treehouse.systems)
ThinkStream says AI can generate work, but people still need to understand it. The provided page content contains only an anti-bot prompt. (thinkstream.app)
Maple uses manifold-activated pruning to compress mixture-of-experts language models below 2 bits per weight. Its developers say it runs Gemma 4 26B at 45 tokens per second on a 6GB GPU. (github.com)
A longtime Claude Code user is switching to Pi after growing frustrated with incompatible instruction-file formats, limited VSCode session support, and bugs in the GUI extension. They say Claude Code has prioritized adding features over fixing reliability and usability problems. (caleb.software)
Anthropic says AI has already helped develop models with capability gains the company did not initially recognize, and it may soon reach its definition of recursive self-improvement. The company warns that faster AI-driven development could create safety risks and says its Responsible Scaling Policy includes plans to monitor and mitigate them. (techcrunch.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.