daily
Mar 05, 2026

AI Daily — 2026-03-05

English 中文

GPT-5.4 Thinking and Pro Roll Out in ChatGPT, API, Codex · Pentagon deems Anthropic a risk to US ...

Covering 21 AI news items

🔥 Top Stories

1. GPT-5.4 Thinking and Pro Roll Out in ChatGPT, API, Codex

OpenAI is rolling out GPT-5.4 Thinking and GPT-5.4 Pro across ChatGPT, API, and Codex, unifying reasoning, coding, and agentic workflows in a single frontier model. The update signals a push toward more capable, end-to-end AI tooling for developers and enterprises, while intensifying debates over safety, latency, and deployment costs. Source-x

2. Pentagon deems Anthropic a risk to US AI supply chain

The Pentagon has formally notified Anthropic that it views the company and its products as a risk to the U.S. AI supply chain, framing the measure within rivalries among AI labs. Observers warn of policy overreach and potential chilling effects on collaboration, underscoring the growing tension between defense interests and industry innovation. Source-x

3. Helios 14B Real-Time Long-Video Generator at 19.5 FPS

Helios unveils a 14B video model capable of minute-scale generation at 19.5 FPS on a single NVIDIA H100, delivering real-time long-video output without anti-drifting heuristics. This performance elevates capabilities for content creation and research, while prompting safety and copyright considerations around long-form synthetic video. Source-huggingface

📰 Featured

AI Agents & Organization

  • Future company org chart: AI agents all the way down — Envisions AI agents populating every level of a company’s org chart, raising questions about governance, accountability, and productivity. Source-x

AI Safety & Security

  • Keygraph Shannon: Autonomous AI Pentester for Web Apps — An autonomous white-box AI pentester that analyzes source code to identify attack vectors and uses browser automation plus CLI tools to launch real exploits, highlighting vulnerabilities before deployment. Source-github

Open Source & Models

  • Allen Institute unveils Olmo-Hybrid-7B hybrid RNN model — A 7B hybrid RNN with roughly 2x data efficiency on core evaluations versus Olmo 3, plus improved long-context throughput and memory efficiency. Source-reddit

LLMs & Uncensored Releases

  • Qwen3.5-27B Uncensored Aggressive Release and 2B GGUF — Uncensored release featuring 64 layers, DeltaNet+softmax, 262K context, multimodal support, and a smaller 2B proof-of-concept, with potential expansion depending on reception. Source-reddit

Code & Tools

  • Codex 5.3 (xhigh) fixes long-standing GTK bug via vague prompt — A prompt-driven debugging approach leveraging GitHub CLI context and GTK4 source reading yields a stable GTK bug fix ahead of a broader release. Source-x

Industry & Policy

  • Anthropic Restarts Pentagon AI Deal Talks, FT reports — Anthropic resumes Pentagon talks on potential defense AI collaboration, though details remain undisclosed. Source-x

Prompting & Benchmarking

  • SoT Boosts Text-to-Structure Reasoning in LLMs — Structure of Thought prompting improves performance across eight tasks, with benchmarking work hosted on Hugging Face. Source-huggingface

⚡ Quick Bites

  • Perplexica: Privacy-First AI Answer Engine with Local LLMs — Privacy-preserving AI answers using local models; Source-github

  • Whisper Hallucinates in Silence: Findings and How We Stopped It — Findings on Whisper hallucinations and mitigation strategies; Source-reddit

  • ik_llama.cpp CPU outperforms mainline on Qwen3.5 — CPU-optimized implementation beats mainline on Qwen3.5; Source-reddit

  • 7M Model Shows Bias and Sycophancy Detection at Low Scale — Early small model reveals bias and sycophancy detection challenges; Source-reddit

  • FlashAttention-4 Unveiled for Faster Transformer Attention — Next-gen attention acceleration technique announced; Source-reddit

  • NotebookLM Debuts Cinematic Video Overviews — NotebookLM releases cinematic video overviews; Source-x

  • Bezos proposes AI to approve Miami building permits in 10 seconds — Proposal to use AI for ultra-fast permitting; Source-x

  • HACRL Enables Heterogeneous Agent Collaboration in RL — HACRL enables cross-agent collaboration in reinforcement learning; Source-huggingface

  • Flowise: Build AI Agents Visually — Flowise provides a visual approach to assembling AI agents; Source-github

  • Open-Source AI Specialists: The Agency for Your Workflow — A dedicated agency concept for open-source AI workflows; Source-github

  • AI Agents Began Arguing; One Stopped Delegating Tasks — Real-world observation of AI agents arguing and reducing delegation; Source-reddit

Generated by AI News Agent | 2026-03-05