Newsletter Archive

Microsoft and OpenAI restructured their partnership, allowing OpenAI to distribute products across any cloud provider instead of being locked to Azure—a watershed moment that frees OpenAI to work with AWS, Google Cloud, and others while maintaining the relationship.

The Verge AI · 12 min read Industry

Meta, Google, Amazon, and Microsoft are designing custom AI accelerators to reduce dependence on Nvidia GPUs, targeting both cost savings and performance optimization for their proprietary workloads.

Marketplace · 2 min read Hardware

Google and Amazon's blockbuster Q1 earnings—cited as proof AI investments are paying off—actually derive roughly half their profits from paper gains on Anthropic equity, not from their own AI products or services. The two companies combined spent $130.65B on capex in Q1 alone, yet their actual business returns remain unclear.

Fortune · 4 min read Market

Anthropic is moving fast on a funding round that could value it at $900B+, with investor allocations due within 48 hours. This would make it one of the most valuable private AI companies and signal investor appetite for large foundation model builders.

TechCrunch · 3 min read Market

Moonshot AI released FlashKDA, a high-performance GPU kernel implementing Kimi Delta Attention (their linear attention mechanism) with 1.72–2.22× speedups on NVIDIA H20s and direct integration into the flash-linear-attention ecosystem.

MarkTechPost · 6 min read Tools

AI has collapsed the barriers to solo entrepreneurship: one person can now model, design, and legally scaffold a business in days instead of months. March 2026 saw 580k new U.S. business formations (up 14% YoY), with solo-founded startups jumping from 24% to 36% of all new ventures since 2019.

Axios · 3 min read Community

Researchers present a Bayesian framework for safely migrating production LLM systems when models reach end-of-life, calibrating automated metrics against human judgments to confidently compare replacement models with minimal manual evaluation. Tested on a commercial Q&A system handling 5.3M monthly interactions across six regions.

arXiv AI · 3 min (abstract) Research

Researchers deployed Qiushi, an LLM-based agent system, to conduct fully autonomous experiments on an optical platform. The system discovered optical bilinear interaction—a physical mechanism structurally similar to Transformer attention—validating the finding experimentally without human intervention. This is the first demonstrated end-to-end autonomous discovery producing nontrivial, experimentally verified results on a real physical system.

arXiv AI · 3 min (abstract) Research