China's Kimi-K2 is Impressive and Open-Source

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...

Find out why 1M+ professionals read Superhuman AI daily.

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

1. Join the Superhuman AI newsletter – read by 1M+ people at top companies

2. Master AI tools, tutorials, and news in just 3 minutes a day

3. Become 10X more productive using AI

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI

print("News & Trends")

AWS Enters the AI Agent Race With Anthropic-Powered Marketplace (3 min. read)

AWS will launch an AI agent marketplace at the AWS Summit in New York City this week, allowing developers to list, sell, and monetize ready-made agents directly to AWS customers. In partnership with Anthropic, the platform will feature Claude-powered agents and offer a seamless way for enterprises to find task-specific tools. This positions AWS to compete more directly with Google and Microsoft in the growing AI agent economy, while giving startups a new channel for distribution within the AWS ecosystem.

Anthropic Rolls out Hooks in Claude Code to Control Agent Actions (12 min. read)

Anthropic’s Claude Code now supports hooks that let you inject custom shell commands before or after agent actions, giving precise control over workflows. You can log commands, enforce formatting, block risky operations like “rm ‑rf,” and trigger notifications or linting automatically. This feature ensures consistent, deterministic behavior by embedding policies into tooling rather than relying on LLM prompts. It’s a powerful upgrade for data scientists and ML engineers seeking robust, customizable agentic control in code workflows.

Kimi K2: Open-Source LLM Powerhouse Built for Scale and Precision (9 min. read)

Chinese startup Moonshot AI has released Kimi K2, a powerful open-source LLM using a 1 trillion parameter mixture of experts architecture with 32 billion active parameters. Trained with their custom MuonClip method for stability at scale, Kimi K2 comes in two flavors—Base for fine-tuning and Instruct for agentic tasks. It rivals GPT-4 on coding and reasoning benchmarks like LiveCodeBench and SWE-bench Verified, while offering 128K context, strong tool use, and a cost-efficient edge for real-world ML applications.

OpenAI Delays the Release of It’s Open Model, Again (3 min. read)

OpenAI has delayed the release of its open-weight model, originally promised by the end of June, with no new timeline set. This model was expected to rival offerings from Meta and Mistral by providing a publicly accessible alternative under an open license. CEO Sam Altman cited safety reviews and the irreversible nature of releasing weights as key reasons for the pause. The delay puts pressure on OpenAI as competitors accelerate their open-source AI efforts and capture community interest.

print("Applications & Insights")

Building with Llama 4 by Meta (Course)
This free one-hour course from DeepLearning.AI and Meta guides you through building multimodal and long-context applications with Llama 4. You’ll learn its Mixture-of-Experts architecture, use the official API to call models like Maverick and Scout, and apply image grounding, long-context reasoning (up to 10 million tokens), prompt optimization, and synthetic data generation. With 9 video lessons and 7 code examples, the course equips ML engineers to build robust, production-ready GenAI features across text and images.

The AI Agent Landscape: What Data Scientists Should Know (And Expect) (11 min. read)
This deep dive demystifies AI agents for data scientists, showing how they extend LLMs with memory, tools, and planning to solve tasks iteratively. It breaks down popular frameworks like LangGraph, and CrewAI, compares agent protocols (MCP, ACP, A2A), and explains why agents are more than hype. Real-world examples show how agents can 10x workflows in EDA and analysis by automating repetitive steps while keeping humans in the loop. A must-read for building smarter systems, not just better prompts.

12-Factor Agents: Patterns of reliable LLM applications (Video)
This video by Dex Horthy, explores building reliable AI agents. Horthy discusses applying software engineering principles, introducing 12 factors for robust LLM-based applications. Key themes include owning control flow, managing execution and business states, crafting prompts, handling errors, and building small, focused "micro-agents" as stateless functions. The goal is to integrate LLMs effectively into systems for high reliability.

The Ultimate Context Engineering Guide
This guide defines context engineering as architecting the full information environment for LLMs—not just prompt tweaks. It covers designing system prompts, dynamic context inputs like dates and tools, RAG and memory management, and agent workflows. It walks through a real multi-agent setup in n8n showing how to engineer context for reliability, cost efficiency, and iterative problem-solving. The guide emphasizes evaluation pipelines and warns about context drift, while previewing advanced topics like compression, safety, and automation.

print("Tools & Resources")

TRENDING MODELS

Text Generation
moonshotai/Kimi‑K2‑Instruct
⇧ 19.3 k Downloads
A 1 trillion‑parameter mixture‑of‑experts model with 32 B active parameters, fine‑tuned for instructive, conversational workflows. It combines high throughput and context awareness for scalable LLM applications.

Text Generation
HuggingFaceTB/SmolLM3‑3B
⇧ 27.1 k Downloads
Compact 3‑billion‑parameter model offering efficient text generation ideal for on‑device and lightweight deployments. It balances performance and resource use for versatile NLP tasks.

Image‑Text‑to‑Text
THUDM/GLM‑4.1V‑9B‑Thinking
⇧ 35.7 k Downloads
A multimodal model that accepts and generates text with image–text fusion, supporting tasks like captioning and visual question answering. It's geared for reasoning over mixed modalities.

Image‑to‑Image
black‑forest‑labs/FLUX.1‑Kontext‑dev
⇧ 254 k Downloads
A fast, open‑source text‑to‑image generator from the Flux.1 family, optimized for high‑quality visuals in creative pipelines. Built for artistic generation and inpainting.

Text Generation
mistralai/Devstral‑Small‑2507
⇧ 7.33 k Downloads
A 24 B‑parameter small variant from Mistral AI’s Devstral line, tailored for software engineering and coding tasks with agentic capabilities. It emphasizes efficiency and instruction following.

TRENDING AI TOOLS

• 🏢 The AI Office Suite: AI-powered docs, spreadsheets, and slides that write, analyze, and visualize in real time.

• 🎬 Google's Flow: Turns images into video with speech and sound, powered by Veo 3.

• 💻 Devstral: A 24B-parameter coding-focused LLM with agentic capabilities and 128K context.

• 🌐 Comet by Perplexity: An AI-native browser that automates search, summarizes pages, and follows context across tabs.

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