Section

AI: models and demos

New models, local tools and the most impressive community creations.

Interactive technical notes covering RLHF, attention, LoRA, and LLM evaluations.

Interactive Technical Notes on RLHF, Attention, LoRA, and LLM Evals

Explore interactive technical notes and resources designed to help you understand RLHF, attention, LoRA, and LLM evaluations by changing inputs.

Showcase · Comunidad

1 day ago1 min read

Card detailing Geoffrey Hinton warnings about artificial intelligence risks and potential regulatory measures.

Geoffrey Hinton warns humanity might have been better off without AI

AI pioneer Geoffrey Hinton warns humanity might have been better off without artificial intelligence, citing severe systemic risks and lack of tech oversight.

Model · OpenAI

1 day ago2 min read

Card detailing Octomind, an open-source Rust-based AI harness that saves tokens and cuts costs.

Octomind: Open-Source Rust AI Harness That Saves Tokens and Costs

Discover Octomind, an open-source Rust AI coding harness that saves tokens, reduces execution costs, and runs across multiple surfaces.

Showcase · MuvonCargoDockerHomebrew

1 day ago1 min read

Card detailing CWE-bench v2 release with expanded coverage and transparency for AI cybersecurity evaluation

CWE-bench v2 Arrives with Greater Coverage for AI Cybersecurity

Discover CWE-bench v2 by Collinear AI, featuring 100 tasks across 91 CWEs to evaluate language models in cybersecurity with programmatic scoring.

Showcase · Collinear AIHarborDocker

1 day ago2 min read

Nvidia RTX Spark laptop running Linux experiences severe GPU clock throttling due to encrypted firmware power tables.

Nvidia DGX Spark Encrypted Firmware Slows Linux Inference Speed

Encrypted firmware on Nvidia RTX Spark laptops restricts GPU clock speeds to 780 MHz on Linux, impacting AI inference performance.

Model · Nvidia120B

1 day ago1 min read

Card detailing Fujifilm boosting tape production to meet AI archiving demand, with over 460 Hacker News points.

Fujifilm to Boost Tape Production as AI Drives Archiving Demand

Fujifilm plans to sharply increase magnetic tape storage production to handle the massive archiving demand driven by artificial intelligence workloads.

Model · Fujifilm

1 day ago1 min read

Palette: teaching a 20B open model to build presentations in your design language using synthetic data and a verification harness.

Palette: Teaching a Small Model to Build Decks in Your Design Language

Discover Palette, a lightweight 20B open-weight model designed to generate custom presentation decks using synthetic data and a verification harness.

Showcase · Modelo LoRA de 20B20BpptxgenjsLoRAPresentBench

1 day ago2 min read

Comparison of AI agent harnesses showing CUGA achieving 81.1 percent task completion on AppWorld.

CUGA vs Hermes and Deep Agents: Choosing the Right Agent Harness

Discover a comparative analysis of AI agent harnesses CUGA, Hermes, and Deep Agents, exploring how architecture impacts enterprise task performance.

Showcase · GLM-5.3-flashAppWorldHarborCodeAct

1 day ago2 min read

Card detailing StateBench, a benchmark testing long-context state tracking in language models across four difficulty levels.

StateBench: Evaluating State Persistence and Tracking in Language Models

Discover StateBench, a new benchmark evaluating how language models track state changes across long documents using four algebraic difficulty levels.

Model · Hugging Face / AllenAIStateBenchGated DeltaNet

1 day ago3 min read

Card showing Qwen 3.8 27B running locally to interpret neural network weights without code execution tools.

Qwen 3.8 27B on My Laptop Can Interpret Model Weights, Too

Discover how Qwen 3.8 27B running locally on a laptop successfully interprets raw neural network weights to identify classic games without code execution.

Showcase · Qwen 3.8 27B27BLM Studiollm

1 day ago2 min read

Postman desktop app silently auto-installs AI plugins for coding agents, raising security and consent concerns.

Postman Silently Auto-Installs AI Plugins and Alters Coding Agent Configurations

Postman desktop application silently installs AI plugins for coding agents like Codex and Claude Code, raising serious user consent and security concerns.

Model · Postman Labs

1 day ago1 min read

Card detailing Hugging Face Model Integration Contracts with a shrunk architecture and fast CPU tests.

Exploring Model Integration Contracts for Transformers in Hugging Face

Explore Model Integration Contracts (MIC) in Hugging Face Transformers, a standard using shrunk model fixtures and CPU tests to catch regressions early.

Model · UMT5 / Llama 4 Scout / Whispertmicpytest

2 days ago2 min read