This week, DeepSeek just released an open-weight model that matches the best of the West for pennies on the dollar, while Salesforce's revenue jump confirms what everyone whispered but nobody wanted to say: AI agents are finally replacing human seats. Let’s dive in ⬇️

In today’s newsletter ↓
🧧 Deepseek's V3.2 Speciale is a serious problem for U.S. AI labs
📉 Salesforce succeeds at replacing humans with AI
📰 Google Discover's AI headlines create a clickbait crisis
👽 Kling 2.6 introduces infinite video backgrounds
🎯 Weekly Challenge: The "Devil's Advocate" Bot

The moat just got a lot shallower. Deepseek, the Chinese AI lab that has been quietly nipping at the heels of Western giants, just released Deepseek V3.2 Speciale. This isn't just another open-source model; benchmarks suggest it matches or exceeds GPT-5.1 performance on reasoning and coding tasks, all while being 10x cheaper to run.

In head-to-head comparisons, the "Speciale" variant is posting numbers that shouldn't be possible for an open-weights model. On the critical AIME 2025 math benchmark, it reportedly scores 96%, edging out GPT-5 and GPT-5.1. Even more striking is its strong performance on the SWE-bench Verified coding test.

While Google's Gemini 3 Pro still holds a slight lead in massive-context multimodal tasks, DeepSeek has effectively commoditized "intelligence" at the reasoning layer. For U.S. labs, the nightmare scenario isn't just that they are being caught; it's that they are being undercut by a competitor running on a fraction of the compute budget.

While OpenAI and Google race to build $100 billion data centers, Deepseek has taken a different path: extreme optimization. The "Speciale" architecture uses a novel Mixture-of-Experts (MoE) approach that activates only a tiny fraction of parameters per token.

This allows it to run on consumer-grade hardware while delivering "frontier-class" logic. For developers paying exorbitant API fees to U.S. providers, the math has suddenly shifted.

The release has set off alarm bells in Washington and Silicon Valley. If a Chinese lab can produce SOTA (State of the Art) models using restricted hardware, the U.S. export control strategy may be failing.

The "capability gap" between closed U.S. models and open global models has effectively vanished. For enterprise CTOs, the choice is no longer "good vs. cheap." It's now "expensive and locked-down" vs. "performant and open." Expect a pricing war to follow immediately.

Marc Benioff is taking a victory lap. After facing intense criticism, and internal morale crises, over major layoffs earlier this year, Salesforce's latest earnings report offers a cold, hard validation: the strategy worked. The company raised its annual revenue forecast, citing explosive growth in its "Agentforce" platform.

The quiet part is now being said out loud. The revenue surge correlates directly with the deployment of autonomous AI agents that handle customer service and sales qualification, tasks previously performed by entry-level humans.

Benioff's controversial past comments about not needing to hire additional software engineers to scale are proving prescient. Enterprises are buying outcomes, not seats, and Salesforce is successfully selling the "human-free" tier of customer relationship management.

Wall Street is eating it up. Bloomberg reports that margins have hit historic highs as the cost of goods sold (COGS) drops. By replacing headcount with compute, Salesforce has decoupled revenue growth from hiring. It's a grim milestone for labor markets but a gold rush for shareholders, signaling that the "AI replacement" phase of the hype cycle has officially begun.

Is this a one-time efficiency bump or the new normal? If Salesforce can maintain quality while slashing staff, every other SaaS CEO will be forced to follow suit. We are watching the first real-time case study of white-collar automation at scale.

🗳️ Cornell study reveals AI chatbots sway voters by 10 points

🛑 Verge slams Google Discover for AI clickbait nonsense

🎥 Kling 2.6 Pro introduces seamless "infinite video backgrounds"

🤝 Meta strikes major content deals to feed AI models

📉 Microsoft fights claims about lowering AI sales targets

🚀 North Dakota details AI optimizing spacecraft propulsion

Here’s what to do:

🔥 Step 1: Pick a topic you feel strongly about (e.g., "Remote work is better," "Pineapple belongs on pizza," or a specific policy).

🗣️ Step 2: Prompt ChatGPT or Claude: "I believe [X]. Act as a world-class debater and persuasion expert. Your goal is to change my mind. Use data, logic, and Socratic questioning to dismantle my argument. Do not be agreeable."

⚔️ Step 3: Engage in a 5-minute debate. Don't go easy. See if the AI can find cracks in your logic you didn't know existed.

📝 Step 4: Review the chat. Did it change your mind? Or did it just make you angry? Note how it tried to persuade you, was it emotional appeals or hard data?

Share your debate logs with us (if you dare).

From Chinese labs breaking benchmarks to Salesforce breaking the hiring link, the rules of the game are being rewritten daily. Have you adjusted to our new AI reality yet? Hit reply and let us know your thoughts.

Zoe from Overclocked