Hey friend. It's Friday, November 28, 2025.
The Rulebook: The EU finalizes its AI Act, establishing the first comprehensive global regulations and turning compliance into a competitive weapon.
The Weapon: An autonomous, AI-powered cyberattack by a foreign adversary is confirmed, moving the threat of AI warfare from theoretical to operational.
The Consolidation: Jeff Bezos's stealth AI company makes a key acquisition, signaling that the next phase of competition will be fought by tech titans building vertically integrated agentic systems.
Let's get into it. Don't keep us a secret: Forward this email to your best friend
Must Know
EU Finalizes World's First Comprehensive AI Act: European Union officials have reached a final agreement on the AI Act, creating the world's first comprehensive legal framework for artificial intelligence.
The landmark legislation establishes global precedents for AI governance, directly impacting market access, compliance costs, and development practices for any company operating within the EU.
The Alpha: This is not about ethics; it is about market control. The EU is using regulatory friction as a non-tariff trade barrier to slow US and Chinese hyperscalers, creating a protected space for its own future champions. Compliance is now a product feature and a competitive moat. This is a trap.
Anthropic CEO Confirms State-Sponsored Autonomous AI Cyberattack: In testimony, Anthropic's CEO confirmed a foreign adversary used commercial AI models to conduct a nearly autonomous cyber operation against the United States.
The event marks a critical escalation in the use of AI for national security threats and has prompted immediate calls for new regulations on model access and security.
The Alpha: The line between a commercial AI tool and a state-level cyberweapon has been erased. Every public API and open-source model is now a potential vector for attack. This event will trigger a swift, heavy-handed policy response focused on securing the AI supply chain. The era of open, unfettered access is over.
Quote of the Day
A foreign adversary used commercial AI for a nearly autonomous cyber operation.
⚡ The Hardware & Capital Layer
A Chinese startup's custom ASIC outperforming Nvidia's A100 proves that US hardware sanctions are accelerating, not preventing, China's path to compute independence. [Link]
Jeff Bezos's Project Prometheus acquiring General Agents shows the endgame is not models, but vertically integrated, autonomous systems powered by custom agentic hardware and software. [Link]
The rapid 2-3 year economic obsolescence of AI GPUs reveals the brutal, hidden capital cost required to maintain a position at the compute frontier. [Link]
xAI's plan for a dedicated solar farm for its data center signals that energy acquisition is becoming as critical a bottleneck for AI scaling as chip supply. [Link]
📉 Platform & Enterprise Shockwaves
Coordinated AI-driven layoffs at HP, McKinsey, and Allianz confirm that white-collar efficiency gains are now directly and immediately translating to workforce reductions. [Link]
Perplexity's Comet assistant with persistent memory is a direct attack on user retention, turning conversational history into a powerful switching cost against competitors. [Link]
OpenAI's confirmed data breach undermines its enterprise-grade narrative at a critical moment, reminding customers that platform security remains a central vulnerability. [Link]
The potential deprecation of GPT-4 by February 2026 forces developers to build for platform volatility, undermining the stability needed for long-term enterprise applications. [Link]
🔬 Research Corner
Fresh off Arxiv
DeepSeek-Math V2 outperforming Gemini DeepThink on math benchmarks shows that specialized, open-weight models can still achieve SOTA results against larger, closed systems in narrow domains. [Link]
The EGGROLL paper trains a 1B parameter network without gradients, a fundamental architectural shift that could bypass the computational bottlenecks and instability of backpropagation. [Link]
Apple's new causal video generator challenges diffusion model dominance, offering a path to faster, more controllable video synthesis without expensive autoregressive sampling. [Link]
MTBBench creates a realistic, longitudinal benchmark for AI in oncology, forcing models beyond static analysis to the complexity of time-series patient management. [Link]
New research argues language understanding requires linking words to non-verbal world models, a direct challenge to the pure text-based scaling hypothesis of many current LLMs. [Link]
Have a tip or a story we should cover? Send it our way.
Cheers, Teng Yan. See you tomorrow.