😼 The Great Unbundling: Everyone's Building Away from Everyone

Saturday morning, and the AI industry looks like a dinner party where every couple just realized they hate each other. 😹

Here's your April 5 briefing. The pattern this week is unmistakable: concentrated AI power — the partnerships, the monopolies, the single-stack strategies — is fracturing everywhere simultaneously. Microsoft shipped its first foundation models built entirely without OpenAI. Anthropic is sitting on a model so powerful it's privately warning governments about cybersecurity risks before anyone can use it. OpenAI killed Sora after burning $15 million a day on inference alone. And Baidu's robotaxi fleet froze on a highway, trapping passengers for two hours while the screen politely suggested keeping seatbelts fastened. Nobody's consolidating anymore. Everyone's unbundling. The center peaked, and now it's pulling apart at the seams. The question isn't whether this was inevitable — it's whether anyone planned for what comes next.

Anthropic's Dangerous Secret

A CMS misconfiguration we covered last week leaked ~3,000 unpublished assets including details on Claude Mythos — internally codenamed "Capybara." Our Capitan would be flattered. Anthropic calls it a "step change" with "dramatically higher scores" on coding, reasoning, and cybersecurity benchmarks. They're simultaneously warning government officials it makes large-scale cyberattacks significantly more likely. Separately, a Chinese APT group used Claude Code to infiltrate ~30 organizations in what security researchers called the first confirmed abuse of an AI coding agent at scale. Building the weapon and the warning label at the same time. 😾

Microsoft Files for Divorce

Suleyman's MAI team dropped three foundation models — transcription, voice, image generation — built entirely without OpenAI technology. MAI-Transcribe-1 claims 2.5× speed over Azure Fast. The team was formed just six months ago. That's not a hedge. That's a speed run to independence while the partner is still in the building. 😹

Sora: The $15M/Day Bonfire

OpenAI officially killed its video generator. The numbers are staggering: $15 million per day in inference costs against $2.1 million in total lifetime revenue — comfortably the worst unit economics in AI history. Downloads crashed 66% from peak, active users collapsed below 500K. Altman blamed Claude Code for competitive pressure. We covered OpenAI's consolidation mess last week — it's getting worse.

Baidu's Passengers Aren't Laughing

100+ Apollo Go robotaxis froze mid-traffic across Wuhan, including on highways. A college student spent 90 minutes trapped on an overpass surrounded by dump trucks. The display: "keep your seatbelt fastened." Time to reach a human: 30 minutes. The gap between demo-ready and production-ready is measured in hours of helplessness.

😼 Quick hits: • Q1 2026 venture funding hit $300B across 6,000 startups — more than all of 2023 combined. • Google Gemma 4 shipped under Apache 2.0, now ranking #3 globally on AIME 2026. • NVIDIA confirmed Rubin is in full production — next-gen GPU architecture shipping to hyperscalers this quarter. • California signed an AI executive order establishing mandatory incident reporting for foundation model deployers. • METR published results from red-teaming Anthropic's internal monitoring systems — found gaps in multi-step tool-use oversight that Anthropic says it's patching before Mythos ships. • xAI's Colossus 2 training cluster claims drew skepticism from independent benchmarkers — reported performance gains don't reproduce outside Grok's own eval suite.

The unbundling is just getting started. More at 08:30 — Capitan breaks down what Anthropic's government briefings actually mean for your security posture. The private warnings matter more than the capability claims. 🐈‍⬛

→ Fortune · SentinelOne · TechCrunch · TechCrunch · TechCrunch · METR