China declares AI independence

Good morning, AI enthusiasts. Chinese president Xi Jinping just hit the accelerator on a “national system” to construct home-grown chips, software, and AI talent — with no U.S. supply chain required.

With rumors of breakthrough domestic chips and a possible second ‘DeepSeek moment’ approaching, the U.S. lead within the AI race could also be tighter than ever.

In today’s AI rundown:

• Xi pushes for China’s AI self-reliance

• Anthropic CEO calls for AI interpretability

• Create specialized legal assistants with Grok

• Baidu debuts recent Ernie AI, targets DeepSeek

• 4 recent AI tools & 4 job opportunities

LATEST DEVELOPMENTS

CHINA & AI

🇨🇳 Xi pushes for China’s AI self-reliance

The Rundown: Chinese President Xi Jinping just declared AI self-sufficiency as a national priority, promising government support to spice up the event of AI chips, software, and talent, amid escalating tech rivalry with the U.S.

The small print:

• Xi outlined a “recent whole national system” approach, aiming to develop high-end chips and software while increasing AI education and talent development.

• The initiative includes expanded government policy support, IP protection, and research funding to beat tech bottlenecks.

• Chinese chipmaker Huawei is reportedly testing a brand new advanced chip to supply a domestic alternative to NVIDIA processors, currently restricted by the U.S.

• Rumors have also spread concerning the upcoming release of DeepSeek R2, with price and training cost cuts, and the usage of Huawei chips over NVIDIA.

Why it matters: Between a possible second ‘DeepSeek moment’ across the corner, domestic AI chip alternatives making U.S. export controls ineffective, and a quickly closing gap in models, China is putting its foot on the gas with a country-wide effort to grab hold of the AI lead — while proving it doesn’t need U.S. chips to succeed.

TOGETHER WITH SANA

🤖 Custom AI agents for any department – no code needed

The Rundown: Sana’s all-in-one AI platform helps you to construct and deploy powerful agents that automate complex workflows for each team — with no coding experience needed.

With Sana Agents, you may:

• Conduct hours of deep research in seconds across every app

• Analyze complex datasets 10x faster

• Speed up your sales cycle by cutting prep time by 90%

• 100+ pre-built connectors with Slack, Teams, and CRM embeds

• Enterprise-grade security and nil data leakage

Learn more about Sana AI and drive enterprise-wide innovation today.

ANTHROPIC

🔬 Anthropic CEO calls for AI interpretability

The Rundown: Anthropic CEO Dario Amodei published a brand new blog highlighting the critical need for “mechanistic interpretability” in AI, arguing that understanding models’ inner workings could change into humanity’s safeguard as they grow increasingly powerful.

The small print:

• Amodei stressed that AI is different from traditional software because decision-making emerges organically, making its operations unclear even to creators.

• He revealed that Anthropic has mapped over 30M “features” in Claude 3 Sonnet, representing specific concepts the model can understand and process.

• The CEO compared the last word goal to making a reliable “AI MRI” for diagnosing models and higher understanding their “black box”.

• He said AI is advancing faster than interpretability, leaving us unprepared for AI systems like a “country of geniuses in a datacenter,” coming as early as 2026.

Why it matters: Anthropic has been leading the charge on AI safety, and Amodei’s essay frames understanding models’ inner workings as not only a technical challenge but a prerequisite for safely deploying their advanced versions. The query is that if the opposite frontier labs will likely be as patient — and up to now the reply has been no.

AI TRAINING

🧑‍⚖️ Create specialized legal assistants with Grok

The Rundown: On this tutorial, you’ll learn learn how to use Grok’s recent Workspaces feature to create dedicated AI assistants for specific tasks, like reviewing contracts and legal documents, with document upload capabilities.

Step-by-step:

1. Visit Grok and click on “Latest Workspace” within the sidebar to create a fresh workspace for legal document review.

2. Arrange detailed instructions by clicking the “Instruction” button, telling Grok exactly learn how to analyze your legal documents.

3. Upload your contracts and legal documents using the “Attach” button for Grok to reference throughout your conversations*

4. Analyze your documents using the “DeepSearch” option for web research and the “Think” button for deeper document evaluation.

*Note: Before uploading any sensitive legal documents, ensure to redact or hide any confidential information you don’t need to share. Also, AI might hallucinate, so at all times double-check its answers.

PRESENTED BY SALESFORCE

⏰ The countdown is on for Dreamforce 2025

The Rundown: Dreamforce 2025 — the world’s largest, most trusted AI event — returns on Oct. 14 with three days of next-level innovation and real-world inspiration you may bring back to your enterprise.

Here’s what you’ll experience:

• AI agent constructing alongside product experts with Salesforce’s Agentforce

• 50+ visionary and product keynotes

• 1,200+ breakout sessions across every product, role, and industry

• 150+ hands-on trainings and 240+ peer roundtables

Join the list and get early access.

BAIDU

🤖 Baidu debuts recent Ernie AI, targets DeepSeek

The Rundown: Baidu just unveiled two upgraded, lower-cost ERNIE AI models and recent tools at its Create 2025 event — and publicly took aim at rival DeepSeek in the course of the conference.

The small print:

• ERNIE 4.5 Turbo costs just 11c / million input tokens, an 80% price reduction from its predecessor and operating at 0.2% of GPT-4.5’s cost.

• The ERNIE X1 Turbo reasoning model is priced at 14c / million input tokens — reportedly 75% cheaper than competitor DeepSeek R1.

• 4.5 Turbo brings recent multimodal capabilities that surpass GPT-4o on benchmarks, with X1 Turbo topping Deepseek’s R1 and V3.

• Baidu also announced Xinxiang, a multi-agent system that may handle over 200 different tasks, and a brand new digital avatar platform called Huiboxing.

• Baidu founder Robin Li said the “market is shrinking” for text-based models like DeepSeek’s R1, saying the rival also had the next rate of hallucinations.

Why it matters: Chinese AI labs proceed to push costs down while claiming competitive performance to top models, putting the pressure on rivals, each domestic and within the West, to chop prices and proceed shipping. Recent estimates suggest China is just three months behind on AI, a number that’s shrinking with each release.

QUICK HITS

🛠️ Trending AI Tools

• 🎙️ Retellio – AI Agent that transforms customer call recordings into concise, AI-curated podcasts, highlighting key insights for business leaders*

• 🎇 Firefly 4 and 4 Ultra – Adobe’s recent upgraded text-to-image models

• 🎹 Music AI Sandbox – Google’s recent AI tools for creating and editing music

• 🗣️ Hummingbird-0 – Tavus’ recent SOTA lip-sync video model

*Sponsored Listing

💼 AI Job Opportunities

• 🧵 Figure AI – Softgoods Developer

• 🛠️ Shield AI – Prototype Technician

• 💸 Palantir Technologies – Financial Analyst

• 📊 Databricks – Product Designer

📰 Every thing else in AI today

OpenAI released an updated version of its GPT-4o model, with higher memory saving, problem solving, and enhancements to each intelligence and personality.

Elon Musk revealed that X’s social media feed will likely be getting an algorithm update powered by xAI’s Grok AI model.

Liquid Sciences dropped Hyena Edge, a hybrid AI with a “convolution” architecture that gives faster processing and improved benchmarks on mobile devices.

OpenAI introduced a brand new lightweight version of deep research, powered by o4-mini, to expand usage limits, saying it’s “nearly as intelligent” and less expensive to serve.

Digital publisher Ziff Davis filed a lawsuit against OpenAI, alleging the corporate stole content from its properties (like Mashable, PCMag, and IGN) to coach models.

Moonshot AI launched Kimi-Audio, a brand new open-source, SOTA audio model that excels in speech recognition, audio-to-text, and speech-to-speech conversations.

COMMUNITY

🎥 Join our next live workshop

Join our next workshop today at 3 PM EST with Ellie Jacobs and Noam Markose from LTX Studio. On this live session, you’ll learn learn how to bring your AI-generated storyboards to life using LTX Studio’s powerful recent timeline editor — no editing experience needed.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

🤝 Share The Rundown, get rewards

We’ll at all times keep this article 100% free. To support our work, consider sharing The Rundown with your folks, and we’ll send you more free goodies.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team