20d ago
Eliminate GPU Waste by Cutting the Retry Tax
Proposes queuing/pacing infrastructure to reduce GPU waste from retry amplification in agent traffic.
20d ago
MiniMax M3 Medium hits 73.17% F1 on DeepSearchQA, near GPT-5 High
Evaluation of MiniMax M3 Medium reasoning model on DeepSearchQA benchmark using an agent harness with search tools, achieving results near GPT-5 High.
20d ago
People vs. the AI Overlords
Opinion essay on how LLM-driven coding workflows and automated code review agents are straining open source maintainers.
20d ago
NanoGPT Speedrun Frontier
Benchmark comparing 18 frontier models on nanoGPT optimizer speedrun via autonomous agent runs, with curated agent trajectories.
20d ago
Low-Rank Adaptation
Wikipedia-style explainer on LoRA, a parameter-efficient fine-tuning technique for large language models.
20d ago
Can Agents Use a Computer Yet? We've Got the Data
Discussion of the state of computer-use AI agents in production, focusing on back-office automation use cases.
20d ago
How LLM Serves a Request
Technical primer explaining how LLM inference servers process requests, covering KV cache, prefill, and decode mechanics.
20d ago
DeepSeek's Price Increase Goes Beyond GPU Costs
Local-first chatbot CLI tool with persistent memory across sessions using Ollama models.
20d ago
Where are people finding GPU capacity?
Analysis of DeepSeek's API price increase and the multiple business factors behind it beyond GPU costs.
20d ago
Show HN: Locum – Grok Bot delegates tasks to Claude/Codex CLI on your machine
A tool letting Grok Bot delegate coding tasks to locally-run Claude Code and Codex CLIs via an MCP tunnel connector.
20d ago
Fast and Hard Code
Essay on how LLMs and coding agents are shifting programming language choice toward performance-oriented 'hard languages' like Rust and Zig.
20d ago
Claude Code status line for worktree, branch, context and quota
A dependency-free bash status line for Claude Code showing worktree, branch, context and quota usage.
20d ago
Show HN: ePub reader that lets your coding agent read along
A CLI companion that lets a coding agent read along in an EPUB reader with web search integration.
20d ago
If you're not using AI to attack your own systems, your adversaries will
Article on AI agents as both offensive security tools and new attack surfaces for organizations.
20d ago
Can an AI agent serve at a Foreign Office?
A feasibility pilot where Claude Opus acted as a fictional foreign ministry desk officer, evaluated against a diplomatic AI evaluation framework.
20d ago
Why your local LLM feels dumber than it is
Technical post explaining how hardware/software implementation differences cause local LLM inference to underperform benchmarks.
20d ago
When AI art has no author: Generated images can't trace to training data (study)
MIT CSAIL research shows 'attribution decay': large-scale generative models make individual training examples untraceable to outputs.
20d ago
Show HN: TechSkills – Open-source skill modules for AI coding agents
TechSkills is an open-source library of markdown-based skill modules giving AI coding agents domain-specific engineering knowledge.
20d ago
An iOS app that runs AI agents and a complete voice pipeline on the device
Show HN: tool for structural code search across public GitHub repos using ast-grep patterns.
20d ago
Show HN: Git workflow as an AI-agent skill
iOS app running on-device AI agents and a full local voice pipeline, with optional web research fetching.
20d ago
Fork of antirez h3.c highly optimized for CUDA and DGX Spark (~15.5x speedup)
A CUDA/DGX Spark-optimized fork of h3.c achieving ~15.5x speedup, alongside native MiniMax-H3 video generation inference on Apple Silicon.
20d ago
Show HN: I turned Apple-style scroll-video websites into a reusable AI skill
Show HN: GitX, a portable Git workflow skill for AI coding agents to manage commits, branches, and PRs.
20d ago
Hotcrp.com – AI agents and bot accounts
HotCRP now supports connecting AI agents via MCP to read submissions, set tags, and write conference reviews, with confidentiality caveats.
20d ago
Show HN: Stash, turn your Instagram saves into notes Claude finds on its own
Stash transcribes saved Instagram posts into markdown notes that Claude can proactively surface without prompting.
20d ago
Grasp: Simple protocol for code collaboration that uses interoperable servers
Model Genome project fingerprints LLMs via architecture, tokenizer, and weights to detect if trained from scratch or derived.
20d ago
Model Genome: Fingerprinting Whether an LLM Was Trained from Scratch or Derived
Show HN: private local AI photo gallery app for Windows with on-device search and AI renaming.
20d ago
Multi-Agent Harness for Visual Design
Foreman is a self-hostable AI-agent software factory that automates dev tasks into reviewed draft pull requests.
20d ago
Show HN: Free Vision TUI Library for Go
Myli is a pre-alpha Python multi-agent harness for proposing JSON-patch changes to visual design documents.
20d ago
TT-AMX, a zero-copy Tensor-Train inference engine for Apple Silicon
Bare-metal C++ engine running Tensor-Train compressed linear layers on Apple Silicon's AMX coprocessor for efficient on-device inference.
20d ago
Show HN: Figranium – Build browser tasks visually, execute via API (Dockerized)
Open-source, self-hosted browser automation tool that turns visual workflows into API endpoints, an Apify alternative.
20d ago
Claude Watermarks AI-Generated Text
Detailed explainer video/lecture on how Claude's new text watermarking mechanism works technically.
20d ago
Show HN: Knowl – agent memory that retires stale facts when they change
Open-source agent memory tool that resolves conflicting facts by superseding stale ones, benchmarked on MemoryAgentBench.
20d ago
StateM: Stateful control for long-horizon agents
Open-source CLI state machine tool for making long-horizon AI agent workflows reliable via inspectable graphs of states and checks.
20d ago
Show HN: I make my AI agents file paperwork before they're allowed to code
Tool that gates AI coding agents behind requirements/design documentation and review workflows before code changes are accepted.
20d ago
Show HN: Running a full AI coding agent inside Cloudflare Durable Object
Community project running DeepSeek's coding agent harness inside Cloudflare Durable Objects/Workers for edge deployment.
20d ago
Show HN: Pisesh, a zero-dependency TUI for finding and resuming Pi sessions
Zero-dependency terminal UI tool for finding, tagging, and resuming Pi coding-agent sessions.
20d ago
Ante: Self-contained coding agent that lives in your terminal and self-organizes
Self-contained Rust-based coding agent CLI with local GGUF inference, claiming top Terminal-Bench 2.1 performance and low resource use.
20d ago
Improving matrix multiplication exponent w. modern optimization and AlphaEvolve
arXiv paper on improving matrix multiplication exponent bounds using AlphaEvolve and modern optimization methods.
20d ago
Show HN: VinvAI runs and observes your services to find bugs and perf issues
Local tool that connects runtime traces to source code, feeds evidence to coding agents, and verifies fixes via acceptance tests using Thompson sampling.
20d ago
hlquery, An open-source C++20 search engine
Open-source C++20 search engine with full-text, hybrid ranking, vector similarity, and AI-assisted search modules via REST API.
20d ago
How to Fingerprint AI Models When Prompts Lie
Technique to identify underlying LLM serving infrastructure via tokenizer artifacts and API response quirks rather than prompts.
20d ago
Show HN: Plainsay – 12 MB Mac dictation with local Whisper and Parakeet
Native Mac dictation app (12MB) running local Whisper and Parakeet speech models offline.
20d ago
Five things make agent-built UI look generic. Each takes 30 seconds to check
Analysis of five common design tells that make AI agent-built UIs look generic, with measured examples.
20d ago
Security assessment found open model near frontiers
Benchmark of 10 AI models' cybersecurity vulnerability rediscovery capability using 32 fresh CVEs across 96 runs.
20d ago
Continue coding agent is dead. Alternatives?
Discussion asking for alternatives after Continue coding agent was acquired by Cursor and repo archived.
20d ago
CyberStrike – open-source AI harness for offensive security (AGPL)
Open-source AGPL AI harness turning any LLM into an autonomous penetration testing agent with 13+ agents and security skills.
20d ago
The hold that could not release itself
Technical postmortem on a production bug where a stale safety hold blocked a contract migration in a course-sync pipeline.
20d ago
Embedded AI
Book chapter on embedded AI covering hardware, data, and deployment for resource-constrained devices with hands-on projects.
20d ago
10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over
Essay arguing that synthetic data, judges, and simulation are progressively replacing human-generated components in AI training pipelines.
20d ago
I turned Unix talk from 1983 into the interface for my AI
Experimental bridge connecting the 1983 Unix talk command to an LLM for character-streamed conversation.