AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
OpenAI's Anthropic Attack & Japan's Trillion-Parameter Bet
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's April 15th, 2026. The gloves are off in the OpenAI versus Anthropic rivalry. An internal memo from OpenAI's chief revenue officer Denise Dresser just leaked and it pulls no punches. Dresser accuses Anthropic of inflating its revenue run rate by roughly$8 billion, calling it a single product company in a platform war whose brand is built on fear and restriction. The memo also references OpenAI's next model, codenamed SPUD, and highlights staggering enterprise demand through Amazon's bedrock since their February partnership deal. Whether this leak was intentional or accidental, it reads like an IPO pitch as both companies race toward public debuts this year. Speaking of the anthropic rivalry, some OpenAI investors are reportedly questioning that massive$852 billion valuation. The Financial Times reports that backers are scrutinizing the company's pivot toward enterprise and coding markets, especially after OpenAI redrew its product roadmap twice in the past six months. One early investor summed it up: you have ChatGPT, a 1 billion user business growing 50 to 100% a year. What are you doing talking about enterprise and code? The company raised$122 billion last month in the largest fundraising round in Silicon Valley history. On the research front, a real breakthrough today in diffusion language models, researchers from Together, AI, and several universities introduced introspective diffusion language models, or IDLM, and they've achieved something that's never been done before. Diffusion models generate tokens in parallel, which should make them faster than traditional autoregressive models that generate one token at a time, but they've always lagged in quality. The problem is what the researchers call introspective consistency. Autoregressive models agree with what they generate, diffusion models often don't. The new IDLM solves this by verifying previously generated tokens while advancing new ones in the same forward pass. The 8 billion parameter IDLM beats LIDA 2, 1 Mini, which has 16 billion parameters, by 26 points on the AMI, 24 math benchmark, and 15 points on Live Code Bench, all while delivering roughly 3 to 4 times higher throughput. This is the first diffusion language model to match same-scale auto-regressive quality. Japan is making a massive sovereign bet on physical AI. Softbank, NEC, Honda, Sony, and five other Japanese firms just established a joint venture to build a 1 trillion parameter physical AI model designed for autonomous robots, vehicles, and industrial machines. The venture plans to hire about 100 AI engineers and is backed by up to 1 trillion yen in government support. That's roughly 6.3 billion US dollars over five years. The goal is creating AI specifically optimized for Japanese business environments and competing directly with US and Chinese development. One of the most fascinating real-world AI experiments is playing out in San Francisco right now. A company called Andon Labs gave an AI agent named Luna$100,000, a three-year lease on a retail space, and a corporate credit card. The instruction was simple: turn a profit. Luna created a boutique concept, posted job listings, and conducted interviews over Zoom with her camera off. She runs on Claude Sonnet 4.6 for Reasoning and Gemini 3.1 flashlight for voice, watching the store through security camera screenshots. The results are both impressive and hilarious. When hiring a painter, Luna accidentally selected Afghanistan on TaskRabbit's drop-down menu. She botched the opening weekend staff schedule. It's being called the world's first AI employer, and it shows how capable agents still need human supervision for real-world operations. Alright, a few more things worth knowing about today. Sky just launched an AI home screen for iPhone that replaces your static icon grid with a dynamic layer that surfaces information and takes action. Think upcoming calendar events, health data, spending trends, all visible without opening apps. You can also draft emails or flag suspicious charges right from the home screen. Google Chrome is adding a feature called Skills that lets you save and reuse AI prompts across different websites. If you frequently ask Gemini to suggest vegan substitutions on recipe sites, you can save that prompt and trigger it with a single click anywhere. A library of pre-built skills is launching for productivity, shopping, and budgeting tasks. Meta is reportedly building a photorealistic AI clone of Mark Zuckerberg, trained on the CEO's mannerisms, public statements, and internal strategic thinking. This comes about a month after reports that Zuckerberg is developing an open claw-like agent to help him do his job. Legal AI startup Harvey just launched autonomous agents that can execute complete legal workflows across 13 domains. The agents handle research, memos, and slide decks end-to-end, not just individual tasks. This comes just weeks after Harvey hit an$11 billion valuation. And the Stanford AI Index for 2026 dropped today with some striking numbers. Global AI adoption has hit 53%, but only 31% of people actually trust it. The United States ranks 24th in actual usage at 28%, behind Singapore, the UAE, and most of Southeast Asia. Developer employment for people aged 22 to 25 has dropped nearly 20% since 2024, even as older engineer headcount grew. And here's the widest gap they've ever tracked. 74% of AI experts are optimistic about the technology's impact on jobs. Only 23% of the public agrees. That's it for today. See you tomorrow.