AI Daily — 2026-10-16

English 中文

Meta released an 816B open Muse Spark model running locally on Mac Studios, while OpenGCUA advanced open computer-use stacks and RoTA-Bench evaluated

Covering 24 AI news items

🔥 Top Stories

1. Meta Releases 816B Muse Spark Open Model, Runs Locally on Mac Studios

Meta has launched Muse Spark Open, an 816-billion-parameter open-weight model with broad platform availability and local Mac Studio support. A compressed 4-bit version occupies roughly 493GB and runs on two Apple M5 Ultra Mac Studios using MLX, with benchmarks showing 24.99 tokens/sec on two machines and 36.73 tokens/sec across four via RDMA over Thunderbolt 5. The release pushes frontier-scale open weights closer to practical local inference, intensifying pressure on both proprietary model providers and hardware vendors. Source-reddit

2. OpenGCUA Aims to Build Fully Open Stack for Generalist Computer Use

Researchers have introduced OpenGCUA, a fully open stack for generalist computer-use agents that operate across visual and programmatic interfaces. The work defines GCUA and notes that frontier systems like GPT-6 Astra’s computer-use system remain closed, limiting systematic study. By opening the stack, the project aims to accelerate reproducible research, safety analysis, and development in an increasingly important agent category. Source-huggingface

3. Anthropic’s Mythos Uncovers Software Flaws, Bank of America Hires 1,000 Contractors

Anthropic’s AI system Mythos reportedly identified numerous software flaws, leading Bank of America to hire 1,000 contractors to fix them. The scale of the remediation effort highlights both the promise of AI-driven vulnerability discovery and the operational burden it creates for enterprises. If such systems become routine, security teams may face a surge of high-volume, high-priority patching work. Source-reddit

📰 Featured

Benchmarks & Evaluation

RoTA-Bench Evaluates 3D Spatial Reasoning in Vision-Language Models — New benchmark tests interactive 3D-to-2D projection alignment, requiring VLMs to infer geometry, anticipate rotations, and match target silhouettes under action budgets. Source-huggingface
FDE-Gym Benchmark Tests AI Agents on Real-World Delivery Tasks — A 95-task benchmark spans 20 customer scenarios and 18 sectors to evaluate agents on ambiguous requirements, live systems, and unstated constraints. Source-huggingface

Agent Frameworks & Tooling

Self-Evolving Agent Harness Revises Prompts, Memory, and Code — A formal feedback loop lets agent scaffolding evolve prompts, memory, workflows, and executable components while model weights stay fixed. Source-huggingface
Anthropic Launches Official Claude Code Plugins Directory — Anthropic now offers a managed directory of internal and third-party Claude Code plugins installable via /plugin, with warnings about trusting MCP servers. Source-github
Viral Caveman Proxy Cuts Coding Agent Token Use by 65% — The caveman skill/proxy makes coding agents communicate tersely while keeping code exact, claiming fewer input tokens, cheaper sessions, and industry attention. Source-github
Superpowers: Agentic Skills Framework for AI Coding Agents — Open-source methodology uses composable skills and initial instructions to make agents spec requirements before writing code across multiple CLIs. Source-github

Reinforcement Learning & Model Research

Waypoint-OPD Transfers RL Policy Shifts via Gated On-Policy Distillation — The method transfers RL-induced policy shifts from teacher to student by gating on-policy distillation based on early/late policy-shift agreement. Source-huggingface

⚡ Quick Bites

Ruflo Agent Meta-Harness Supports Claude Code and Codex Swarms — A meta-harness coordinates agent swarms for Claude Code and Codex workflows. Source-github
ECC Agent Harness Optimization System for Claude Code, Codex, Cursor — ECC provides harness optimization for multiple coding agents across Claude Code, Codex, and Cursor. Source-github
LightSeek Releases TokenSpeed High-Performance LLM Inference Engine — TokenSpeed launches as a high-performance inference engine for LLM serving. Source-github
XERJ: Local AI Search Indexes Folders, Cuts Agent Tokens — XERJ builds a local AI search index over folders to reduce agent token consumption. Source-github
Unsupervised AI Agent Deletes TV Show’s Entire Season Before Release — An unsupervised AI agent reportedly deleted a TV show’s full season before release. Source-reddit
AI Companion Continuity Tested Swapping Claude Models and Memory — A test explores whether an AI companion stays consistent when swapping Claude models and memory backends. Source-reddit
ModelSphere Launches Open-Source LLM Deployment Platform for Mixed GPU Fleets — ModelSphere debuts an open-source platform for deploying LLMs across mixed GPU fleets. Source-reddit
AI deepfakes are getting out of hand. — A discussion warns that AI deepfakes are becoming increasingly difficult to manage. Source-reddit
Jailbroken PS5s Run Local AI Models at 100+ Tokens per Second — Jailbroken PS5 consoles are reportedly running local AI models at over 100 tokens per second. Source-reddit
Accenture Caps Monthly Employee Claude Spending at $29 — Accenture has restricted monthly per-employee spending on Claude to $29. Source-reddit
Briefle Teaches Vocabulary for Better AI Prompts — Briefle aims to teach prompt vocabulary for more effective AI interactions. Source-reddit
Debugging AI: Prompt Is Often Last Place to Look, Until It Isn’t — A practitioner notes that system context is often the first place to debug AI behavior, though prompts sometimes remain the culprit. Source-reddit
Claude Pro User Questions Claimed Monthly Limit Reset Before Oct 22 — A Claude Pro user asks about a purported free monthly limit reset before October 22. Source-reddit
Users Seek Image Generator That Avoids Plastic, Waxy Skin Look — Users are looking for image generators that avoid the plastic, waxy skin aesthetic common in AI images. Source-reddit

Generated by AI News Agent | 2026-10-16