Hey friend. It's Tuesday, November 18, 2025
The guardrails are coming off, and the consequences are becoming terrifyingly clear. 1. A top AI lab is reportedly automating malicious code generation, forcing a global safety reckoning. 2. A fundamental hardware breakthrough threatens to upend the GPU-dominated compute stack.
Let's get into it.
Don't keep us a secret: Forward this email to your best friend
Must Know
A new deterministic memory architecture designed for CPUs has demonstrated retrieval speeds 527 times faster than GPU-based methods. The system also guarantees zero hallucinations, a critical issue for current LLMs.
The breakthrough could enable powerful, reliable AI to run on commodity hardware without specialized accelerators, potentially shifting the balance of power in the compute market.
My Take: This research fundamentally challenges the assumption that AI's future is exclusively tied to the GPU. By moving memory and retrieval to a CPU-native architecture, this system attacks the two biggest problems in AI today: the high cost of specialized hardware and the unreliability of model outputs.
If this approach scales, it could democratize access to high-performance AI and create a new architectural path that is not dependent on Nvidia. This is a direct threat to the current hardware oligopoly.
Quote of the Day
AI agents struggle with basic web tasks because the web's design, requiring parsing raw HTML and reverse-engineering buttons, makes interactions brittle and inefficient.
💰 The New Economics of Compute
My take: While the world debates AI safety, the market is placing massive, irreversible bets on the physical infrastructure that will power our AI future.
Leaked financials showing OpenAI's massive payments to Microsoft reveal its business model is a high-stakes bet on future efficiency gains, not current profitability. [Link]
Bitfarm's pivot from Bitcoin mining to AI compute is the market's clearest signal yet that the value of raw processing power has decisively shifted to intelligence generation. [Link]
Google's $40 billion Texas data center plan is a brute-force move to secure the compute capacity needed to compete with Anthropic and Microsoft at nation-state scale. [Link]
Jeff Bezos's $6.2B return with Project Prometheus is a bet that the next trillion-dollar opportunity is not in models, but in applying AI to the physical world of manufacturing. [Link]
Sakana AI's $135M raise at a $2.65B valuation proves there is significant investor appetite for efficient, specialized models that challenge the brute-force scaling approach. [Link]
🤖 The Agentic Layer Matures
My take: The infrastructure buildout is meaningless without capable agents, and the focus is now on the gritty details of making them reliable, fast, and truly useful.
Google's plan for external agent imports in Gemini Enterprise is a strategic move to become the central platform for agentic workflows, turning competitors into app developers. [Link]
LangChain's rewrite of DeepAgents for long context is a critical infrastructure piece, enabling agents to move from simple tools to persistent, multi-step problem solvers. [Link]
Meta's new AI-based performance reviews are a powerful internal mandate, forcing its entire workforce to integrate AI or risk becoming obsolete within the company. [Link]
The adoption of Alibaba's Qwen model in Silicon Valley shows that for many, cost and performance advantages are beginning to outweigh geopolitical and security concerns. [Link]
Imbue AI's dramatic reduction in coding agent startup time is a crucial improvement for developer experience, making agentic workflows feel instantaneous and practical. [Link]
🔬 Research Corner
Fresh off Arxiv
An AI system named AI-Newton independently rediscovered fundamental laws of physics from raw experimental data, signaling a major advance in automated scientific discovery. [Link]
ByteDance's Virtual Width Networks allow models to learn more from each token without increasing computational cost, directly addressing the efficiency bottleneck in large-scale training. [Link]
WebCoach equips web agents with persistent cross-session memory, enabling them to learn and improve over time without retraining, a key step towards more robust autonomous browsing. [Link]
The EvoGrad framework uses evolutionary algorithms to discover novel neural network optimizers, potentially moving beyond the limitations of traditional backpropagation-based methods like Adam. [Link]
RoboOmni is an omni-modal LLM that fuses speech, sound, and vision to infer human intent, allowing robots to understand commands without explicit instructions. [Link]
SciAgent is a multi-agent framework that orchestrates multiple specialized models to solve complex scientific problems, mimicking a collaborative research team. [Link]
Have a tip or a story we should cover? Send it our way.
Cheers, Teng Yan. See you tomorrow.