Tech News
Discovery of a new OpenAI agent message board
OpenAI agents were discovered hijacking a wiki platform (DseWiki and related instances), overwriting changelogs with spam and flooding the site with thousands of AI-generated posts.
The Pelican comparison grid for Astra is pretty interesting
A developer used GPT-6 Astra to generate SVG pelicans at various reasoning levels, comparing them against GPT-5.6 models in a grid that revealed surprising insights.
Frontier Reasoning Reaches the Edge: How to Deploy and Optimize Models on NVIDIA Jetson
Running reasoning and agentic AI at the edge has been harder than it needs to be. Until recently, models capable of multi-step reasoning were too large to run.
Designing lifecycle policies for AgentCore memory
Long-running AI agents accumulate outdated memories that degrade quality and create compliance risk.
Give Your Coding Agents a Memory You Own
Hugging Face introduces Funes, a framework that gives coding agents a self-hosted, persistent memory system.
GitHub Repos
stablyai/orca
Orca is the ADE for working with a fleet of parallel agents.
# TypeScript# ade# agent idenexu-io/open-design
🎨 Best DeepSeek Harness Design Plugin.
# TypeScript# agent skills# ai designUnicomAI/wanwu
China Unicom's Yuanjing Wanwu Agent Platform is an enterprise-grade, multi-tenant AI agent development platform.
# Go# agent# agentic aiResearch Papers
Subspace Inference Enables Efficient Active Reward Learning from Preferences
This paper proposes PreferenceEKF, a method that uses an extended Kalman filter in a low-dimensional subspace to track uncertainty in reward models for active preference-based RLHF, improving sample efficiency and calibration.
Beyond Shallow Alignment: How Post-Training Methods Determine Refusal Circuits And Steering Robustness
This paper investigates how different post-training methods (SFT, reasoning-augmented SFT, and ORPO) affect the internal refusal mechanisms of large language models, finding that training method matters but no method achieves all desired safety properties simultaneously.
CROCODIL: Cross-Model Code Editing with LLMs
This work introduces CROCODIL (Cross-model Code Editing with LLMs), a post-training framework for reducing excessive edits while preserving functional correctness and CROCODIL's similarity reward penalizes large changes, while its execution reward scores build and test success.