AI Daily — 2026-03-01

English 中文

WSJ: U.S. Used Claude AI for Iran Air Strikes · GUI-Owl 1.5 Releases Native GUI Agent Family

Covering 26 AI news items

🔥 Top Stories

1. WSJ: U.S. Used Claude AI for Iran Air Strikes

The Wall Street Journal reports that CENTCOM has been using Anthropic’s Claude AI for intelligence assessments, target identification, and simulating battle scenarios in its operation against Iran. The usage spans commands worldwide, illustrating how embedded AI tools are in military planning, with CENTCOM declining to comment on specific systems. Source-twitter

2. GUI-Owl 1.5 Releases Native GUI Agent Family

X-PLUG’s Mobile-Agent announces GUI-Owl 1.5, a native GUI agent family spanning 2B/4B/8B/32B/235B with Instruct & Thinking, built on Qwen3-VL. It enables desktop, mobile, and browser automation, claiming state-of-the-art results on 20+ GUI benchmarks and strong end-to-end performance. Weights are on HuggingFace, with a technical report and README; demos include Mobile-Agent-v3.5 via Modelscope and Bailian, offering a limited free API. Source-github

📰 Featured

LLM

Perplexity Enables AI-Powered Portfolio Management in Seconds — Perplexity can generate a minimalist portfolio management tool capable of handling 100+ clients with dashboards and client views from a single prompt. What once took months and six-figure budgets can now be created in seconds, effectively building AI-native financial infrastructure in real time. The development signals a broader disruption of traditional wealth-tech app models. Source-twitter
Qwen3.5-35B-A3B Runs Real-Time on M4 Laptop — Qwen3.5-35B-A3B, a 35B-parameter model, reportedly runs locally on an M4 chip at about 49.5 tokens per second, enabling real-time inference on a laptop. The demonstration highlights rapidly advancing local AI capabilities on consumer hardware and hints at broader edge deployment possibilities. Source-twitter
Perplexity Computer Enables 1-Person Startup to Nearly Reach $1B — Patrick Moorhead lauds Perplexity Computer as the first platform he’s used that nearly end-to-end enables a one-person startup to approach a $1B valuation by matching the right model with the right tool. He praises its ability to align models and tools for the job, calling it impressive. The post is from X (formerly Twitter). Source-twitter
PSA: Check KV cache quantization before blaming 30k-context models — Long-context local agents (around 30k tokens) can hallucinate or mishandle tool calls. The author argues aggressive KV cache quantization (Q4/Q8 in llama.cpp or ExLlamaV3) is often the real cause, despite little impact on short-context perplexity. The post cites OpenClaw testing with malformed JSON outputs around 30k tokens and urges practitioners to check KV cache quantization before tweaking prompts or precision. Source-reddit

AI Agents

Agent-browser adds Electron control for desktop apps — An Electron skill for agent-browser lets agents control desktop apps built with Electron, including Discord, Figma, Notion, Spotify, and VS Code. It can also be used to debug Electron apps and to extend coding agents via npx skills add vercel-labs/agent-browser —skill electron (or —skill slack for Slack). Source-twitter

AI Policy

Criticism of SCR designation on Anthropic by DoW — An AI-policy advocate argues that enforcing the SCR designation on Anthropic would be damaging to the AI industry, the country, and Anthropic itself. They emphasize de-escalation, prioritize safe superintelligence and broad benefit sharing over corporate competition, and urge the DoW to reverse the decision despite potential criticism. Source-twitter

Open Source

Nous Research Releases Open-Source Hermes Agent for Personal AI — Hermes Agent is a fully open-source AI agent you install on your own machine, connecting your messaging accounts and becoming a persistent personal assistant that learns, builds skills, and runs tasks on a schedule. It supports multiple models via Nous Portal, OpenRouter, or your own VLLM/SGLang endpoint, with a terminal UI (TUI) rather than a web interface. Developed by Nous Research, the project underpins data generation and RL training for tool-calling models. Source-github
PaddlePaddle: Open-Source Deep Learning Platform from China — PaddlePaddle (飞桨) is one of the first independent R&D deep learning platforms and has been open-sourced since 2016. It provides high-performance single-machine and distributed training, cross-platform deployment, and a rich ecosystem of core frameworks, libraries, tools, and services, serving 23.33 million developers, 760,000 companies, and 1,100,000 models. Source-github

AI Tools

Data.gouv MCP Server Enables AI Chatbots to Access Datasets — The official MCP server from data.gouv.fr lets AI chatbots such as Claude, ChatGPT, and Gemini search, explore, and analyze datasets from the French Open Data platform directly through conversation. A public instance at https://mcp.data.gouv.fr/mcp provides a no-access-restriction endpoint for seamless, natural-language querying of datasets. Source-github

Hardware

Reverse-engineered Apple Neural Engine trains 110M MicroGPT — An author claims to reverse-engineer ANE private APIs using Claude to bypass Core ML, building a bespoke training pipeline that trains a 110M MicroGPT. They argue the Apple Neural Engine offers extreme power efficiency and suggest LoRA training for 3B/7B models, with larger-scale training possible on clusters. Source-reddit
AMD GPU firmware update boosts Vulkan with ROCm and llama.cpp — An AMD GPU firmware update, combined with a new llama.cpp Vulkan build, reportedly delivers a substantial Vulkan performance boost on Qwen3.5-35B-A3B-Q8_0. Tests were conducted on Debian GNU/Linux with ROCm 7.12 nightly and a Vulkan-enabled llama.cpp build, showing marked improvements over prior Vulkan results. The post also references energy efficiency comparisons and earlier weaker Vulkan performance on older Qwen3.5 configurations. Source-reddit

⚡ Quick Bites

AGI Emergence Requires Engineering, Not Pure Theory — Francois Chollet argues that emergent properties from scaling algorithms are effectively unpredictable, making AGI more of an engineering challenge than a theoretical one. He emphasizes that progress in AGI comes through building and discovery rather than purely formal analysis. Source-twitter
AI in Warfare: Flawed High-Stakes Decision-Making Model? — A tweet‑level critique argues that leaders could delegate critical warfare decisions to a basic AI, likening it to Microsoft Clippy. It advocates for truth-seeking AI and cites Katie Miller’s Feb 28 post and the grok account, highlighting safety concerns in high-stakes military decision‑making. Source-twitter
Honor Builds Phone With Integrated AI Robot Companion — Honor is building a smartphone that doubles as an AI robot, with a pop-up camera acting as the AI’s eyes. The concept envisions a continuously active AI companion that can function as an assistant. The piece calls the idea interesting but possibly a gimmick, while noting personal AI companions are coming. Source-twitter
Former OpenAI Geopolitics Lead publishes observations on frontier AI military policies — A former OpenAI Geopolitics Team lead publishes observations on frontier AI companies and their military usage policies, drawing on experience as a former employee and international security researcher. The notes are shared in a Twitter post with a link provided. Source-twitter
Anthropic promotes Claude memory import for an easy switch — A teaser post on Twitter promotes switching to Claude by highlighting a memory import feature. The post directs users to claude.com/import-memory, signaling an easy transition to Claude’s memory capabilities. Source-twitter
Curated Awesome LLM Apps with AI Agents and RAG — A curated collection of LLM-powered apps built with RAG, AI Agents, and multi-agent systems, spanning OpenAI, Anthropic, Gemini, Google, and open-source models. It showcases practical demos—from coding helpers to email workflows—with well-documented projects and local-running options. The repository invites community contributions and highlights sponsorships. Source-github
Superset: IDE for the AI Agents Era on macOS — Superset markets a turbocharged terminal that lets developers run 10+ coding agents (e.g., Claude Code, Codex) locally. It provides per-task worktree isolation, parallel agent execution, centralized monitoring, and a built-in diff viewer to review changes, aiming to boost development throughput. The open-source project on GitHub positions itself as the central hub for managing AI agents on a single machine. Source-github
Sub2API Delivers Open-Source AI API Gateway for Subscriptions — Sub2API is an open-source AI API gateway that distributes and manages API quotas from AI product subscriptions (e.g., Claude, OpenAI, Gemini, Antigravity). It issues API Keys to users and handles authentication, billing, load balancing, and request forwarding. The platform offers multi-account management, API key distribution, precise token-level billing, smart scheduling, and per-user concurrency control; a demo and GitHub source are provided. Source-github
US Military Prefers Claude Over GPT, Users Resist Switch — The piece argues the US military is more addicted to Claude than to GPT, implying a deep dependency that actual users will resist. It warns that leaders who forced the switch will confront pushback, comparing this to Copilot being despised in many organizations. Source-twitter
OpenAI contract clash: all lawful uses vs safety stack — A Twitter thread argues OpenAI claims its contract permits ‘all lawful purposes’ while asserting protections reside in the deployment safety stack. The author warns that if the safety stack blocks a lawful use, the contract could be breached, raising enforcement concerns. The discussion also speculates about potential future contract disputes with the Pentagon over safety-language interactions. Source-twitter
Honor to adopt Deepseek — A Reddit post claims Honor would use Deepseek in its devices. The claim references an X status but provides no official confirmation or details. Source-reddit
Critics slam AI performance as underwhelming — A tweet on X expresses strong negative sentiment toward AI, calling its performance underwhelming. The post provides no specifics about the AI model or context, illustrating public frustration with AI quality. Source-twitter
OpenAI Employees Assigned to PR Part-Time Roles — An X post alleges that OpenAI staff are being assigned to part-time PR duties and pressed to publicly defend the company’s decisions. The poster expresses sadness and hopes the company hires full-time PR staff soon. The item highlights staffing and communications practices rather than AI research or product developments. Source-twitter

Generated by AI News Agent | 2026-03-01