Microsoft's Copilot Mode

Hey there 👋

We hope you're excited to discover what's new and trending in AI, ML, and data science this week.

Here is your 5-minute pulse...

print("News & Trends")

Meta Names ChatGPT Co-Creator as Chief Scientist of Superintelligence Lab (1 min. read)

Meta has named Shengjia Zhao, a key co‑creator of ChatGPT and GPT‑4, as chief scientist of its new Superintelligence Lab, part of an aggressive push to rival OpenAI and Google. Zhao will shape its research agenda alongside Alexandr Wang, Meta’s chief AI officer, in a bid to unify efforts across the FAIR division and product teams. The move follows Llama 4 setbacks and massive investments, including a $14.3 billion stake in Scale AI, to accelerate Meta’s ambitions in building open‑source AGI.

China Unveils Blueprint for Global AI Governance (6 min. read)

China’s new Global AI Governance Action Plan pushes for international collaboration on AI safety, innovation, and accessibility, especially for the Global South. It calls for open-source platforms, high-quality datasets, green tech standards, and public sector leadership. The plan urges alignment with UN goals, cross-border cooperation, and traceable, secure AI development to prevent misuse. It’s an ambitious blueprint for a multilateral, inclusive approach to shaping the future of global AI systems.

Agent‑Native Open LLM: GLM‑4.5 Raises the Bar (10 min. read)

Z.ai, formerly Zhipu, has released GLM‑4.5 and its leaner GLM‑4.5‑Air as fully open‑source, agent‑native models that unify reasoning, coding, and tool use within a single architecture. GLM‑4.5 (355 B total, 32 B active) outperforms nearly all open‑source peers and ranks third globally across 12 key benchmarks, while GLM‑4.5‑Air hits top-tier efficiency with only 106 B total and 12 B active. Both offer 128 K context windows, hybrid thinking modes, MIT licensing, and dramatically lower cost and inference overhead than proprietary alternatives.

Microsoft’s Edge Becomes Your AI Co‑Pilot (7 min. read)

Microsoft just dropped Copilot Mode in Edge, an experimental feature that lets users control the browser through natural language, combining search, page navigation, tab analysis, and task execution in a single chat interface. It can summarize articles, compare products, dig through multiple tabs, and even act as an agent to complete actions like booking a trip. Available now for free on Windows and macOS, it signals a shift toward more autonomous, context-aware browsing experiences.

print("Applications & Insights")

Six Principles for Production AI Agents (9 min. read)
This guide outlines six engineering principles for reliable production AI agents, focusing on clear and detailed system prompts, minimal and split context with on-demand fetch tools, and simple toolsets with strictly typed parameters. It advocates an actor‑critic loop for automated validation and LLM-driven error analysis that surfaces blind spots. It also notes that frustrating agent behavior usually signals system design flaws, not model limitations, and highlights that strong evaluation pipelines and feedback loops are core to scalable agent reliability.

Building and evaluating alignment auditing agents (45 min. read)
Anthropic has unveiled three autonomous “auditing agents” to automatically probe LLMs like Claude for misalignment risks. An investigator agent digs into hidden goals via interpretability tools, an evaluation agent constructs and runs behavioral tests, and a red‑teaming agent launches parallel probes to surface implanted quirks. Single agents catch about 10–13% of injected issues, but aggregating as a super‑agent boosts detection to 42%, showing promise for scalable AI safety auditing.

How to add real-time web search to your AI agent
This hands-on tutorial walks you through integrating real-time web search into your AI agents using Tavily and LangChain. Across three progressive notebooks, you’ll learn to extract and crawl web content, build intelligent agents that can browse and answer with fresh data, and finally combine live web info with private knowledge bases for more grounded, context-rich responses.

print("Tools & Resources")

TRENDING MODELS

Text Generation
Qwen/Qwen3‑Coder‑480B‑A35B‑Instruct
⇧ 14,400 Downloads
A 480 billion parameter mixture‑of‑experts model with 35 billion active parameters and native 262K‑token context support plus extrapolation to million‑token contexts, optimized for agentic code generation and tool‑leveraging tasks.

Text Generation
zai‑org/GLM‑4.5
⇧ 919 Downloads
A MoE hybrid‑reasoning foundation model (355 B total, 32 B active) offering thinking mode for complex reasoning and tool use plus non‑thinking mode for fast responses, excelling at coding, agentic multi‑step planning and long‑context tasks.

Text-to-Speech
bosonai/higgs‑audio‑v2‑generation‑3B‑base
⇧ 82,500 Downloads
A 3 billion parameter audio generation model trained to synthesize natural speech from text inputs, tuned to deliver expressive and high‑quality TTS output with realistic prosody and clarity.

Image-to-3D
tencent/HunyuanWorld‑1
⇧ 5,430 Downloads
An image‑to‑3D model converting 2D images into 3D representations optimized for visual and spatial reconstruction tasks, enabling users to generate 3D content from single images efficiently.

Text Generation
Qwen/Qwen3‑235B‑A22B‑Instruct‑2507
⇧ 18,500 Downloads
A 235 billion parameter Qwen3 model with 22 billion active MoE parameters, engineered for instruction‑following, reasoning and multilingual generation over long contexts in an efficient sparse architecture.

TRENDING AI TOOLS

• 🎤 Seed LiveInterpret 2.0: Real-time Chinese-English speech interpretation with voice cloning.

• 📊 Shortcut: AI analyst for spreadsheets and dashboards.

• 🤖 Moby Agents: AI agents for ecom insights and automation.

• 📚 Anara: AI assistant for reading, researching, and writing.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the pulse experience for you.

See you soon,

Andres