baidu-goes-open-source-with-ernie-45-meta-grabs-talent-from-openai-img-0

AI_Distilled #102: What’s New in AI This Week

Learn to Run and Deploy Open-Source LLMs with This Free Course

baidu-goes-open-source-with-ernie-45-meta-grabs-talent-from-openai-img-1

Join Open-Source LLM Zoomcamp to explore how to run, fine-tune, and deploy open-source large language models. During this short free course, you’ll discover the open-source LLM ecosystem, learn practical tools like Hugging Face, vLLM, and Llama Factory, and work with models like DeepSeek-R1.

Register now for free

Hello and welcome to this week’s AI roundup!

Here’s wishing our readers in the U.S. a very Happy Independence Day!

This week, we’re witnessing thrilling developments in the AI race. With China closing in on the AI gap through its open-source strategy and Meta poaching OpenAI’s employees, it looks like this summer is heating for the AI giants.

Dive in for the full scoop.

LLM Expert Insights,

Packt

In today's issue:

🧠 Expert Build Recap: “Build AI Agents Over the Weekend” drew hundreds to prototype real-world agent use cases with LangChain and Python.
🔮 Next Up—DeepSeek Demystified: Get ready for a live breakdown of DeepSeek’s architecture, strengths, and red flags.
🌍 Global Agent Meetups: From EUMAS to PRIMA, the best events this fall spotlight the future of multi-agent systems.
📦 ERNIE 4.5 Goes Open: Baidu drops 10 massive multimodal MoE models under Apache 2.0—toolkits included.
💸 Meta’s AI Flex: $14.3B Scale AI stake and a star-studded OpenAI exodus fuel Zuck’s Superintelligence Labs.
🤖 Gemini Powers Robotics & Devs: On-device robot control and CLI magic—Google is gunning for full-stack AI.
🛡️ Cloudflare vs AI Bots: With "Pay Per Crawl," Cloudflare strikes back at lopsided content scraping economics.
🛠️ Langfuse Gets Agentic: Multi-agent onboarding, DevOps-ready orchestration, and observability out of the box.

📈LATEST DEVELOPMENT

Here is the news of the week.

Baidu open sources ERNIE 4.5 model family

Baidu's ERNIE 4.5 is a newly open-sourced family of 10 large-scale multimodal AI models, featuring Mixture-of-Experts (MoE) architectures with up to 424B parameters. It features a heterogeneous modality structure designed for efficient cross-modal learning, enhancing performance in text, image, audio, and video tasks. Trained using the PaddlePaddle framework, ERNIE 4.5 achieves state-of-the-art results in instruction following, knowledge retention, and multimodal reasoning. All models are available under the Apache 2.0 license, accompanied by industrial-grade development toolkits. Read more.

Meta creates SuperIntelligence Labs, SamA calls it distasteful

Meta has successfully recruited several researchers from OpenAI, including Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai. These hires are part of Meta's strategy to assemble a world-class AI research team to drive its superintelligence ambitions. Read more.

OpenAI CEO Sam Altman called Meta’s $100 million-plus recruitment packages “distasteful,” insisting none of OpenAI’s top engineers have defected to Zuckerberg’s new Superintelligence Labs.

In another development, Meta has announced a $14.3 billion investment to acquire a 49% stake in Scale AI. This move is aimed atbolstering Meta's capabilities in AI data labeling and infrastructure, positioning the company to accelerate its AI development initiatives. Watch this at 22:39.

Google pushes it with robotics on device and Gemini CLI

Google DeepMind has introduced Gemini Robotics On-Device, is an AI model that runs directly on robots, eliminating the need for internet connectivity. It offers general-purpose dexterity and rapid task adaptation, enabling robots to perform complex tasks like folding clothes or assembling parts. The model adapts to various robot types and can learn new tasks with minimal demonstrations. AnSDK for developers is has also been made available for fine-tuning and testing. Read more.

Google has also released Gemini CLI, a free, open-source AI tool that integrates Gemini 2.5 Pro directly into developers' terminals. It supports natural language prompts for coding, content creation, and task automation, with generous usage limits. The CLI is extensible, integrates with Gemini Code Assist, and supports tools like Veo and Imagen for multimedia generation. Read more.

Cloudflare introduces pay per crawl feature, pushes for fair web-use

Cloudflare's latest Radar update reveals a growing imbalance between AI bots' bots scraping content and

genuine user referrals. For instance, Anthropic's Claude exhibits a 70,900:1 crawl-to-referral ratio, indicating extensive content access with minimal traffic return. This trend threatens publishers' revenue models, prompting Cloudflare to introduce tools like "Pay Per Crawl" and default AI bot blocking to help content creators manage and monetize AI-driven content usage. Read more.

Langfuse gets agentic onboarding

In its latest update, Langfuse introduces Agentic Onboarding and the Docs MCP Server, allowing developers to spin up multi-agent swarms with a single command, instrument them end-to-end, and hand them to DevOps for seamless production readiness. Read more.

EXPERT INSIGHTS

What We Built—and What’s Next

DeepSeek is an emerging open-source large language model (LLM) ecosystem that’s making waves by delivering GPT-4-level performance without the usual proprietary restrictions. Its flagship DeepSeek-V3 model offers results comparable to GPT-4 at only a fraction of the training cost, with model weights openly available to the community.

Under the hood, DeepSeek’s success stems from unique technical breakthroughs. Techniques like Multi-Head Latent Attention (MLA), Mixture-of-Experts (MoE) architecture, Multi-Token Prediction (MTP), and 8-bit floating point (FP8) precision training work in tandem to boost efficiency and scale. These innovations allow DeepSeek models to maximize throughput and minimize memory bottlenecks, enabling performance on par with leading closed models at dramatically lower cost. Equally important, DeepSeek’s open approach invites the global AI community to build upon these advances, accelerating progress toward more accessible AI.

Real-world use cases for DeepSeek already span a broad spectrum. Developers are using the specialized DeepSeek-Coder model for AI-assisted code generation in over 80 programming languages. Other DeepSeek variants excel at complex reasoning (solving math and logic problems) and multilingual natural language understanding, thanks to training on massive, diverse datasets rich in high-quality multilingual data. This versatility makes DeepSeek attractive to practitioners seeking cost-effective, cutting-edge AI solutions.

For those eager to learn more, Packt is hosting a one-day virtual summit "DeepSeek Demystified" on August 16 to explore these innovations. It’s a chance to hear insights from experts and see DeepSeek in action — interested readers can register here.

baidu-goes-open-source-with-ernie-45-meta-grabs-talent-from-openai-img-2

If scaling LLMs in production is on your radar, block time for the ML Summit 2025 and MCP Workshop. And there is 25% off our combined ticket with the discount code MCP25.

With the combined ticket, you’ll learn how to:

Build flexible pipelines that don’t fall apart under load
Utilize data infrastructure for AI: SQLMesh, DuckDB, and Apache Iceberg
Use Model Context Protocol (MCP) to keep your AI tools and LLMs separate

BOOK YOUR SPOT

Use code MCP25 at checkout to get 25% off

📈UPCOMING EVENTS

Upcoming Must-attend AI Agents Events

The world of AI agents is evolving rapidly, with agent-based architectures and autonomous systems taking center stage. From global conferences to hands-on developer meetups, the latter half of 2025 offers many opportunities to learn, network, and build with cutting-edge AI agent technologies. Here's a curated list of key events you won't want to miss:

1. EUMAS 2025 – European Conference on Multi-Agent Systems

Date: September 3–5, 2025
Location: Bucharest, Romania
Cost: TBA
Focus: Research on multi-agent systems

2. AI Agent Event 2025 – East Coast Edition

Date: September 29–30, 2025
Location: Herndon, VA, USA
Cost: $695 (Early Bird), $995 (Regular)
Focus: Real-world AI agent use cases across business and tech

3. PRIMA 2025 – Principles and Practice of Multi-Agent Systems

Date: December 15–21, 2025
Location: Modena, Italy
Cost: TBA
Focus: Research, principles, and applications of multi-agent systems
Website: prima2025.unimore.it

4. AI Agents Summit 2025

Date: TBA
Location: Online
Cost: TBA
Focus: Tools, use cases, deployment, innovation in agents
Website: aiagentsummit.com

What’s stopping you? Choose your city, RSVP early, and step into a room where AI conversations spark, and the future unfolds one meetup at a time.

Built something cool? Tell us.

Whether it's a scrappy prototype or a production-grade agent, we want to hear how you're putting generative AI to work. Drop us your story at [email protected] or reply to this email, and you could get featured in an upcoming issue of AI_Distilled.

📢 If your company is interested in reaching an audience of developers and, technical professionals, and decision makers, you may want toadvertise with us.

If you have any comments or feedback, just reply back to this email.

Thanks for reading and have a great day!