A project demonstrates running large language models on devices with minimal RAM by using custom kernels, memory optimization, and flash storage streaming. Results show 16B models running on 1GB of RAM.
From Hacker News
Google launches Gemini 4 Hydrogen, a faster, more efficient model for everyday coding and agentic workflows, with improvements in speed, cost, and task completion across multiple domains.
From Google DeepMind
This guide provides scripts and containers to fine-tune TRELLIS.2, a 3D generation model, using LoRA adapters. It includes instructions for running on Hugging Face Jobs or locally with Docker.
From Hugging Face
Researchers show that small local language models can generate Linux kernel semantic patches with guidance, outperforming direct prompting and achieving high precision.
From arXiv
ADAPT is a new tool that uses an autonomous coding agent to generate focused regression tests for pull requests, achieving higher success rates than existing methods.
From arXiv
A self-taught software engineer created a terrain generator using diffusion models, published a SIGGRAPH paper, and gained significant attention online. The process involved years of studying math and ML, followed by experimentation with code and model training.
From Lobsters
JioStar’s chief architect discusses lessons learned from deploying AI at scale, focusing on evaluation, guardrails, and production challenges.
From OpenAI
Rasoi is a tool that uses visual diffs to help developers understand how coding agents change a codebase's structure. It was tested with 12 developers and showed promise in identifying unexpected restructurings and ensuring changes stay within scope.
From arXiv
A new method called SatisFix helps coding agents better follow task requirements by identifying and correcting implementation choices that don't fully meet specifications.
From arXiv
A study evaluates whether a self-learning migration agent for Oracle-to-PostgreSQL improves over time, finding mixed results with significant memory leakage and limited policy adaptation.
From arXiv
A new testing approach, Spec-Driven PBT, uses code specifications to derive test properties, finding more unique bugs than traditional Code-driven PBT while writing fewer tests.
From arXiv
A new tool called isloss detects silent information loss when translating conversations between LLM APIs, using a shared intermediate representation and explicit rules to identify avoidable data loss.
From arXiv
KernelAudit introduces an independent-audit framework to evaluate the quality of validators used in generating GPU kernels, identifying false-reject and false-accept issues in existing validators.
From arXiv
EvoX is a new agent runtime framework that improves the success rate of LLM agents by treating the harness as an operational policy, rather than just a tool wrapper. It outperforms existing harnesses across multiple models and benchmarks.
From arXiv
GHA-SmellFix is a new tool that uses agentic LLMs to detect and refactor GitHub Actions workflow smells. It compares two designs—joint and decomposed—and finds that decomposition helps control when patches are generated but doesn't universally improve quality.
From arXiv
JoyFrontBench is a new benchmark for evaluating front-end coding agents, using goal-oriented acceptance contracts to distinguish between partial implementation and full application success.
From arXiv
LOHRE is a new framework that improves UDT recovery by combining LLM-based hypothesis generation with execution validation and cross-function evidence.
From arXiv
A false smoke alarm triggered a chain reaction that could have destroyed Jane Street's NYC datacenter. The alarm shut down cooling systems, leading to a dangerous rise in temperature before a systems engineer manually pulled power cords to prevent disaster.
From Hacker News