Superhuman hit 200K requests per second—and load balancing became the bottleneck.

Power-of-Two Choices isn't fancy, but it works. Instead of round-robin, you sample two nodes per request and pick the one with lower load. Superhuman and Databricks built a 200K QPS inference platform using this technique, handling grammar correction at scale for Coda. At that throughput, traditional routing explodes. Infrastructure gets unsexy when it actually has to work.

🚀 Databricks shipped the MCP Marketplace, turning agent data access into plug-and-play.

MCP Marketplace is a governance layer for Model Context Protocol servers—think AppStore for agentic data sources. Agents can now discover and integrate external data sources with built-in control mechanisms. This matters because agents without reliable, governed access to real-time data are just expensive chatbots. Databricks is betting that enterprise AI runs on composable infrastructure, not monolithic models.

🏗️ Voice agents are eating the middle of the call center stack.

Autonomous Tool Control capability means agents can now operate software interfaces, not just generate text. Combined with voice-based reasoning and configurable inference budgets, this unlocks real customer service automation. Attio (AI CRM) and Zillow voice agents show the pattern: if you can coordinate tools + voice + reasoning, you don't need humans for tier-1 support. The economic pressure is immediate.

💰 Frontier labs are hiring; everyone else is pretending to do AI.

Anthropic is growing 10x year-over-year while traditional tech announces layoffs. The divergence is stark. Real AI builders are scaling fast. Everyone else is doing “AI-washing”—sprinkle Claude or GPT into a product, announce it as transformation, then cut costs. The industry's doing natural selection in real-time.

🔬 Open-weight models have already won machine translation.

Proprietary translation (DeepL) vs open-weight models is no longer a fair fight. Open models are cheaper, faster to deploy, and good enough. The economic moat evaporated. This pattern repeats: whenever a task becomes commodified, open weight wins. Builders should stop licensing proprietary APIs for solved problems.

🎯 Still no Claude Mythos broad release, GPT-5.5 is live but real-time reasoning is the real story, and Zyphra is quietly shipping open-source MoE models.

That's the brief. Full pages linked above. See you tomorrow.