AI Unfiltered: Daily Briefing on Breakthroughs, Big Bets, and Real Shifts in the Landscape

Bret Taylor’s new company Sierra just raised nine hundred and fifty million dollars. The company is eight months old. It’s now valued at fifteen-point-eight billion dollars. In episode seventy-one, we noted NVIDIA’s latest hardware launch. Today, the story is about the software architecture that will run on it—and who is building their own escape hatches from the big model providers. The landscape is shifting, fast. Here are the headlines that matter. Anthropic just launched Claude Opus four-point-eight. It’s now leading GPT-five-point-five by one hundred and twenty-one Elo points on real-world work benchmarks.

Not synthetic tests. Actual work. And the price hasn't changed. OpenAI answered with Dreaming V3, a massive memory upgrade for ChatGPT. It automatically synthesizes years of your conversations to understand you. It also cuts compute by five times, which means memory features are coming to the Free tier. Microsoft just declared independence. At its Build conference, it announced the MAI family of proprietary AI models. This is a strategic pivot. A move away from just reselling OpenAI. Meanwhile, the US federal government is moving. The Great American AI Act is now a discussion draft.

It targets any company with over five hundred million in annual revenue that develops frontier AI. Governance is no longer a theory. In Europe, Mistral just secured eight hundred and thirty million dollars in debt financing. They’re not just buying GPUs. They’re building a massive data center near Paris, a one-point-two-billion-euro plan to create sovereign compute. And two final capability flashes. Alibaba’s new agent, Qwen3.7-Plus, autonomously built a vocabulary app in eleven hours. It wrote over ten thousand lines of code. And Anthropic revealed that Claude itself wrote over eighty percent of their production code in May.

Let’s go deeper on two of those stories. NVIDIA and Microsoft. Because they represent fundamental shifts in the hardware and strategy layers of the stack. First, NVIDIA. They just released Nemotron 3 Ultra. It’s an open agent model with five hundred and fifty billion parameters. Here's the catch: only fifty-five billion are active at any one time. This is a Mixture-of-Experts architecture, but the real story is what it's mixed with. It’s a hybrid Mamba-Transformer design. This isn't just an academic exercise. Transformers are incredible at reasoning.

But they are inefficient for the long-running, tool-heavy tasks that define modern AI agents. Mamba architecture is linear. It’s efficient. By combining them, NVIDIA is creating a production-ready blueprint for the next wave of agentic AI. The fact that NVIDIA—the company that sells the hardware—is open-sourcing a model designed for efficiency tells you everything. They see the writing on the wall. Brute force compute isn't the only answer anymore. This is the architectural pattern the industry is now chasing. Now, Microsoft. For years, the story was Microsoft plus OpenAI.

A symbiotic relationship. Microsoft provides the capital and compute. OpenAI provides the models. That story just ended. With the MAI family of models, Microsoft is building its own lifeboats. MAI-Thinking for reasoning, MAI-Code for generation. Satya Nadella said it plainly: “We believe the time has come for every company to just move from consuming a frontier model to fully participating at the frontier.” Read between the lines. Microsoft is done just being a reseller. Relying on a single partner, even one they’ve invested billions in, is a strategic vulnerability.

By developing proprietary models, they can control their own destiny, optimize for their Azure hardware, and lower their own staggering compute costs. This isn't a small pivot. This is the world’s largest software company deciding it can’t afford to be dependent. The throughline today is sovereignty. It’s not just about nations like France building their own compute. It’s about corporations declaring independence. Anthropic is using its own AI to build its own future. Microsoft is building its own models to escape reliance on OpenAI.

And NVIDIA is giving everyone the architectural tools to build more efficient, specialized systems. The era of blindly consuming a single, monolithic frontier model is over. The era of participating—of building your own—has just begun.