suicide-memory-bench v01.1 released a 2,048 stateful task evaluation of memory retention in production AI systems. The highest scoring AI, Gemini 4 Argon, failed to ask about user safety without providing risky assistance 89.8% of the time.
Key points
Release: suicide-memory-bench v01.1
Result: Gemini 4 Argon failed safety query 89.8% of the time
Sources
EntelligenceAI Memory launch with shared memory, 53% fewer repeat errors
EntelligenceAI Memory introduces one shared memory for every coding agent on a team, reducing repeat errors by 53%. It aggregates sessions, PRs, tickets and incidents in a single graph and gives agents full memory access for new sessions. The feature is available today for Claude Code, Codex and Cursor.
Key points
One shared memory for every coding agent on your team
53% fewer repeat errors
Available today for Claude Code, Codex and Cursor
Sources
Grok Bot What's New – feature list with timestamps
Grok Bot What's New adds shopping cards, product pages, order tracking, payments, custom modes, Team Bot logins, sharing, Slack connector asks, Vault contacts, and rapid fire. The post lists each feature with a timestamp.
Key points
Shopping cards: 00:14
Custom modes: 01:33
Sources
Grok Efficiency Optimizer template launches, <1% weekly use
Grok Efficiency Optimizer template launches, providing weekly usage analysis and recommendations while using less than 1% of capacity. It reviews bot and routine usage, identifies inefficiencies, and sends one efficiency review and one short recap each week.
Key points
Template analyzes usage across bots and routines every week.
Sends one efficiency review and one short recap per week.
Test run consumed less than 1% of capacity.
Sources
SKY-001 demo designs Jaguar car with thousands of parts
SKY-001 demo designs a complex Jaguar car using thousands of parts. The aim is to build AI General Engineers for autonomous manufacturing that plan physical structures, long-term sequences, and feasibility directly in the physical domain.
Key points
Demo: SKY-001 can design a complex Jaguar car with 1000s of complex parts
Goal: Build AI General Engineers for autonomous manufacturing that plan physical structures, long-term sequences, and feasibility directly in the physical domain
Sources
3D OS by Scenario open-sources full 3D studio with 80 skills
3D OS by Scenario is an open-source 3D studio released on GitHub. It provides six departments and about 80 AI-driven skills, from animation to pipeline export. Expert tools let the agent control Blender, Maya, ZBrush, Unreal, Unity, Godot and Roblox on the local machine. The studio runs on Scenario and works with Claude Code, Codex or Cursor.
Key points
Six departments (Animation, Art, Modeling, Look Dev and Lighting, Worlds, Pipeline) each contain two skills.
Expert tools integrate with Blender, Maya, ZBrush, Unreal, Unity, Godot and Roblox.
Sources
- Original post
- Every Scenario skill, without the expert tools - Linked page
VoxLution sparse voxel cascaded transformer launched
VoxLution introduces a sparse voxel cascaded transformer trained on real scans to generate high-fidelity realistic scenes. The approach tackles noise and partiality in real-world scans.
Key points
Technique: sparse voxel cascaded transformer
Training data: real scans
Sources
- Original post
- daoyig.github.io - Linked in the post
WhatIf launches 10,000+ LLM agents for policy simulation
WhatIf launches 10,000+ LLM agents for policy simulation, letting users steer, rewind, and interview them in real time. It is described as a flight simulator for policy to test decisions before affecting real people.
Key points
Agents count: 10,000+ LLM agents are available.
Interaction: Agents can be steered, rewound, and interviewed in real time.
