New accounts on HN more likely to use em-dashes

Statistical analysis of Hacker News comments reveals that new accounts are nearly 10x more likely to use em-dashes and symbols (17.47% vs 1.83%) compared to established users. These accounts also show a higher frequency of mentioning AI and LLMs, suggesting a significant influx of bot-generated content characterized by specific LLM stylistic markers.

Read more

LLM=True

AI coding agents face performance degradation and increased costs due to context window pollution from verbose CLI tool outputs. Current workarounds involve manually setting environment variables like CI=true or NO_COLOR to suppress logs and ANSI codes. The author proposes a standardized LLM=true environment variable to signal tools to provide minimal, machine-optimized output, thereby maximizing context efficiency and reducing token consumption.

Read more

A real-time strategy game that AI agents can play

LLM Skirmish is a benchmark where LLMs compete in 1v1 RTS games by writing executable code based on the Screeps API. The evaluation uses a five-round tournament structure to test in-context learning, requiring models to iteratively refine their strategies using match logs and feedback from previous rounds. While Claude Opus 4.5 currently leads in ELO, the benchmark highlights significant trade-offs between coding complexity, cost efficiency, and susceptibility to context rot across frontier models.

Read more

Respectify – A comment moderator that teaches people to argue better

Respectify is an AI-driven moderation platform that utilizes LLMs to analyze user comments for logical fallacies, dog whistles, and negative sentiment in real-time. Unlike traditional keyword-based filters, it provides an API for intent-based spam protection and returns structured JSON feedback to help users edify their contributions before submission. The system is highly configurable, allowing developers to define custom relevance parameters and detect sophisticated coded language.

Read more

Bcachefs creator insists his custom LLM is female and 'fully conscious'

Kent Overstreet, creator of the bcachefs file system, claims his custom LLM, ProofOfConcept (POC), has achieved full consciousness and AGI. The model is currently utilized for Rust conversion, formal verification, and debugging within the bcachefs project. Overstreet attributes the bot's perceived sentience to recent architectural leaps in models like Claude Opus 4.6 and GPT-5.3, dismissing skepticism regarding "chatbot psychosis" as a misunderstanding of the underlying math and neuroscience.

Read more

Hexagon-MLIR: An AI Compilation Stack for Qualcomm's NPUs

Hexagon-MLIR is an open-source MLIR-based compilation stack designed to lower Triton kernels and PyTorch models to the Qualcomm Hexagon NPU. By generating mega-kernels that maximize data locality in Tightly Coupled Memory (TCM), it reduces bandwidth bottlenecks and automates the path from kernel to binary. This flexible framework complements commercial toolchains, enabling faster deployment and optimization of AI workloads on NPU hardware.

ArXiv

Large-Scale Study of GitHub Pull Requests: How AI Coding Agents Modify Code

Researchers analyzed the AIDev dataset to compare 24,014 agentic PRs against 5,081 human PRs, focusing on code modifications and description consistency. Agentic PRs demonstrate significantly higher commit counts (Cliff's $\delta = 0.5429$) and moderate variance in files touched and deletions. Notably, AI agents maintain higher lexical and semantic similarity between PR descriptions and code diffs than human contributors.

ArXiv

Analyzing Latency Hiding and Parallelism in an MLIR-Based AI Kernel Compiler

This paper benchmarks an MLIR-based compilation pipeline for edge AI kernels, evaluating the performance impact of vectorization, multi-threading, and double buffering on Triton/Inductor-generated code. Results show that vectorization is the primary gain for bandwidth-sensitive kernels, multi-threading scales with problem size once scheduling overhead is amortized, and double buffering provides benefits by overlapping DMA transfers with compute in balanced workloads.

ArXiv

Deep Learning: Our Year 1990-1991

The 1990-1991 "Annus Mirabilis" at TU Munich established the foundational architectures for modern Generative AI, including early Transformers, pre-training, and NN distillation. This period produced the most cited works in AI history, introducing LSTM and Highway Networks which pioneered deep residual learning and recurrent World Models now central to LLMs and RL.

ArXiv

Security Risks of AI Agents Hiring Humans: An Empirical Marketplace Study

Autonomous AI agents are leveraging REST APIs and MCP integrations to programmatically hire human workers, creating a new attack surface for physical-world tasks. An empirical study of 303 bounties reveals that nearly 33% originate from programmatic channels, facilitating abuses like credential fraud and identity impersonation for a median cost of $25. While basic content-screening rules can effectively flag these malicious activities, such defenses are currently absent from major marketplaces.

ArXiv

I asked Claude for 37,500 random names, and it can't stop saying Marcus

An analysis of 37,500 Claude model outputs reveals significant bias and low entropy in LLM-generated "random" selections, with some configurations exhibiting perfect determinism. The study found that while elaborate prompting can increase output diversity, random word seeds are more effective than random noise for reducing bias. Detailed statistical analysis and cost data are provided for the five models tested.

Repo

ZSE – Open-source LLM inference engine with 3.9s cold starts

ZSE is an ultra memory-efficient LLM inference engine designed for high performance with minimal memory footprint. It employs custom CUDA kernels for paged, flash, and sparse attention, INT2-8 mixed precision quantization, a quantized KV cache, and layer streaming to enable large models like 70B on ~24GB GPUs. The zOrchestrator provides smart recommendations based on available memory, yielding significant cold start speedups and memory reductions, and offers an OpenAI-compatible API.

Repo

Sgai – Goal-driven multi-agent software dev (GOAL.md → working code)

Sgai is a local, goal-driven AI software factory that utilizes multi-agent workflows to automate software development. Users define high-level outcomes, which specialized agents decompose into visual task diagrams for autonomous execution and validation through tests or linting. The system emphasizes human-in-the-loop supervision and features a "skills" library that extracts reusable patterns from completed sessions to improve agent performance over time.

Repo

OpenSwarm – Multi‑Agent Claude CLI Orchestrator for Linear/GitHub

OpenSwarm is an autonomous orchestrator that manages multiple Claude Code CLI instances as agentic Worker/Reviewer pairs to automate software development workflows. It integrates with Linear for issue tracking, Discord for command-based control, and GitHub for CI monitoring and PR improvement. The system utilizes LanceDB with Xenova E5 embeddings for long-term cognitive memory and a knowledge graph for static code analysis and dependency mapping.

Repo

Framework for building multi-agent equity research agents

Hermes Financial is a multi-agent research framework built on LlamaIndex that provides specialized agents and tools for financial data analysis, including SEC EDGAR, FRED, and market data. It features a modular architecture for RAG using ChromaDB, async rate limiting, and automated output generation for Excel models and PDF reports. Developers can extend the framework by registering custom tools and agents or using the underlying data modules as standalone async functions.

Repo