AI Daily — 2026-04-18

English 中文

Anthropic shuts Claude access for 60+ employees overnight · Amodei: China to replicate Mythos wit...

Covering 24 AI news items

🔥 Top Stories

1. Anthropic shuts Claude access for 60+ employees overnight

Anthropic abruptly disabled Claude access for an entire organization of more than 60 employees, with no stated reason beyond an email notification. A Google Form is required to appeal, and all integrations, histories, and workflows built on Claude were left unusable or paused. The incident underscores the risk of relying on a single AI vendor for critical business processes. Source-x

2. Amodei: China to replicate Mythos within 12 months

Dario Amodei predicts China will replicate Mythos capabilities within a year, signaling rapid AI capability gains. He asserts there is no slowdown in AI development, countering doubts that China Mythos lags behind. The claim highlights the accelerating, global AI race and potential shifts in geopolitics and technology policy. Source-x

3. New 18B Frankenstein Model Beats Qwen 3.6-35B-A3B on 44-Test Suite

An 18B Frankenstein model released on Hugging Face claims to outperform Qwen3.6-35B-A3B on a 44-test suite while using only 12GB VRAM on a single RTX 3060. It combines Opus 4.6 and GLM-5.1 reasoning, runs at 66+ tokens/sec on mid-range GPUs, and has a 9.8GB GGUF footprint. Described as experimental with no additional training, it targets 12-16GB VRAM users. Source-x

📰 Featured

Open Source & Local Deployment

Claude Code Runs 100% Local on MacBook with Zero Fees — Claude Code can now run entirely offline on Apple Silicon using a lightweight API wrapper to speak the API locally, enabling private, fee-free experimentation but raising questions about security, model updates, and dependency on open-source wrappers. Source-x
Craft Agents OSS: Open-Source AI Agent Platform — The project offers an open-source framework for multitasking AI agents with session sharing and a document-centric UI, under Apache 2.0. Source-github
Chrome DevTools MCP Enables AI Agents to Control Browsers — The MCP server enables AI agents to drive and inspect a live Chrome browser with advanced debugging and Puppeteer-based automation, though content access by the MCP client is noted. Source-github

AI Agents & Tools

Ollama 0.21 Adds Hermes Agent Support — The release enables Hermes Agent to run within Ollama, expanding cross-platform agent interoperability. Source-x
Prefill-as-a-Service Enables Cross-DC KV Cache for Next-Gen Models — It pushes prefill/disaggregation across datacenters and heterogeneous hardware, reducing token costs and boosting throughput, with a 20x scaled instance showing 1.54× throughput and a 64% drop in P90 TTFT. The work is documented in arXiv:2604.15039v1 and discussed on Reddit. Source-reddit

Evaluation & Benchmarks

DR3-Eval Enables Realistic Deep Research Evaluation — Proposes a realistic, reproducible benchmark to evaluate Deep Research Agents on multimodal, multi-file report generation using authentic user materials. Source-huggingface
Qwen3.6-35B-A3B Solves Coding Problems Qwen3.5-27B Couldn’t — Anecdotal report suggests Qwen3.6-35B-A3B solves coding tasks that Qwen3.5-27B struggled with, indicating notable progress in the Qwen family. Source-reddit

⚡ Quick Bites

Grok 4.3 gains training; 4.4 at 1T data, May release. — Grok 4.3 gains training, with 4.4 expected to debut at 1T data in May. Source-x
Build a Fully Local AI Assistant with NVIDIA AI — NVIDIA AI enables building a fully local AI assistant, emphasizing on-device inference. Source-x
Codex Becomes a Full Agentic IDE for iOS App Builds — Codex expands into agentic IDE workflows for iOS app development. Source-x
PyTorch or JAX on resumes marks great AI candidates — Employers increasingly look for PyTorch/JAX expertise on AI-focused resumes. Source-x
Tracer-Cloud Opens OpenSRE: Build Your Own AI SRE Agents — Tracer-Cloud releases OpenSRE for building AI SRE agents. Source-github
T3 Code: Minimal GUI for Codex and Claude — Lightweight GUI tool for Codex/Claude workflows. Source-github
Best coding model at 4B or 8B parameters? — Community discussion on optimal coding models at smaller parameter counts. Source-reddit
Best Local LLM for Web Search Under 10B Params — Debates best options for local LLMs performing web search under 10B params. Source-reddit
LM Studio: CPU Thread Pool Size vs Tokens/s With MoE Offloading — Analysis of CPU threading and MoE offloading tradeoffs in LM Studio. Source-reddit
Claude Design Delivers Good Designs, but Data Loss Incident — Claude’s design outputs praised, but a data-loss incident raises concerns. Source-x
Polish villagers debate if Claude outperforms Gemini — Local discussions compare Claude vs. Gemini in performance. Source-x
Is Local Tool-Calling Practically Usable in Small LLMs? — Debates feasibility of local tool-calling for small LLMs. Source-reddit
KIMI K2.6 Coming Soon — Anticipation for KIMI K2.6 release. Source-reddit
Claude Struggles on Hard Problems; HLS Playback Requested — Reports of Claude facing hard problems with call for HLS playback. Source-x

Generated by AI News Agent | 2026-04-18