Microsoft just stopped being OpenAI's distribution partner. Loudly.

Yesterday in AI | Thursday, June 4, 2026

Microsoft just stopped being OpenAI's distribution partner. Loudly.

This episode covers a day when the biggest players in AI all moved at once: one tech giant showed up to its developer conference with 7 new in-house models, an autonomous agent, and hardware that looks nothing like anything shipping today. Google told investors it can't build fast enough and asked for $80 billion to fix it. And new data on tech layoffs reveals that CEOs have been quietly misattributing the cause - the actual number is smaller than almost anyone is saying. Plus, Anthropic spent a year studying 832 real criminals who tried to use Claude as their hacking toolkit, and what they found will change how you think about AI-enabled threats.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Thursday, June 4th, and Microsoft just declared its AI independence at build. Google is asking investors for $80 billion because it literally can't keep up with demand, and new data is quietly dismantling the story most tech CEOs have been telling about AI and layoffs. Let's get into it. Let's start with the biggest story of the day, Microsoft Build 2026. For years, Microsoft's AI story was really OpenAI's story. Every headline was Microsoft partners with OpenAI to do X. At Build 2026 this week, that changed. Microsoft dropped seven new in-house MAI models spanning reasoning, coding, image generation, voice, and transcription. The headliner is MAI Code 1 Flash, a model built to turn plain descriptions into source code for apps and websites, now baked into GitHub Copilot, Microsoft's AI coding assistant, and Visual Studio Code. They also previewed MAI Thinking One, a reasoning model designed for complex tasks at low token costs. Microsoft says it worked with McKinsey to fine-tune one of these models and beat GPT-4.5 at 10 times the cost efficiency. That's a number we're sitting with. The economics here are pretty clear. When you're paying a third party every time the model answers a query, building your own model starts looking a lot more attractive. Microsoft is betting its own stack running on Azure is the better margins play as demand grows. But models were just part of it. Microsoft also announced Scout, their first always-on autonomous agent, built on OpenClaw, Microsoft's internal system for running AI agents, software that acts autonomously on your behalf. Scout lives in Microsoft 365 and can reschedule meetings for you without being asked twice, handle multi-step tasks end-to-end, and will be priced on usage rather than a flat subscription. It'll initially require GitHub Copilot access. Then there's Project Solara, a platform that Microsoft calls Agent First Devices. The reference design is a desktop hub next to your PC that responds to voice commands, plus a wearable badge with a fingerprint button that wakes an agent with one press. Microsoft is handing these designs to hardware partners to build on. Oh, and they announced a quantum computing chip they expect to make commercially viable by 2029. Just a casual week. Microsoft spent this week telling developers it can stand on its own. OpenAI spent the same week telling everyone else they can use Codex. OpenAI announced that Codex now has 5 million weekly active users, up from 3 million in April, and they launched six new role-specific plugins designed to pull in every kind of knowledge worker, not just engineers. We're talking a sales plugin with Salesforce and Hubspot Pre-Wired, a creative production plugin with Figma and Canva built in, a data analytics plugin connected to Snowflake and Tableau, plus equity investing and investment banking plugins. In total, 62 tools and 110 skills bundled across these plugins, ready out of the box with no coding required. They also launched a sites feature. Codex can now build and host interactive web apps, dashboards, and project hubs your team can share via a simple URL. And an annotations update lets you point to a specific cell in a spreadsheet or line in a document and say fix that part, rather than regenerating the whole thing. What OpenAI is building with Codex looks more and more like a universal work platform, the kind of thing every employee in a company routes their day through. Whether your team actually builds the habits to use it is a different question, and honestly, a harder one. While Microsoft was wiring up its agent platform and OpenAI was handing Codex plugins to every job function, Alphabet was doing something a little different, asking investors for $80 to $85 billion. Google published an investor presentation this week explaining why. The headline, Demand for their AI products is meaningfully exceeding available supply. They have 2.5 billion monthly users of AI overviews in search, over 1 billion monthly users of AI mode, and 900 million monthly Gemini app users. And they're saying we can't keep up. The plan is a $180 to $190 billion infrastructure budget for 2026, almost entirely for data centers and Google's proprietary AI chips. They're balancing cash flow, debt, and equity to avoid over-leveraging while building fast enough to hold position. What stands out is the framing. Google isn't leading with a product roadmap. They're saying the constraint is physical infrastructure and they're willing to dilute shareholders to solve it. That's the argument every major player is making right now. Models are becoming commodity. Physical compute is the moat. All that money moving into AI infrastructure keeps raising the same question, who exactly is winning here? A new analysis digs into the layoff numbers, and the answer is more complicated than the headlines suggest. Big tech CEOs have spent the past year crediting AI for mass layoffs. The pattern has been convenient. Cut thousands of jobs, cite AI, and the market rewards you for being strategically ahead. But CGC, which tracks U.S. labor data, put an actual number on it. In 2025, 1,206,374 layoffs were announced across the U.S. economy. Of those, only 54,836 were explicitly attributed to AI. That's 4.5%. The rest? Post-pandemic overstaffing corrections. Interest rate effects, slower growth. The same forces that drive layoffs in any economic cycle. What executives figured out is that saying AI makes cuts sound strategic rather than reactive. Analysts reward it. Shareholders reward it. So the language caught on fast. This matters because how we respond to AI job displacement depends on what we think is actually causing it. Retraining programs, education investment, policy proposals, all of those get built around a diagnosis. If that diagnosis is being shaped by corporate messaging more than actual data, the responses probably are too. And while companies are figuring out how to talk about AI's economic impact, Anthropic published research this week about a more immediate problem, the people actively using AI to cause harm. Not theoretical threat modeling, real banned users. They mapped the findings to the mitre attack framework, which is the standard taxonomy the security industry uses to describe attack techniques. What did they find? Hackers are using Claude across every phase of an attack, scoping out targets, writing malware, manipulating people into handing over access, and finally breaking in. What Anthropic found is that AI is compressing the timeline from someone with intent to someone with operational capability. Things that used to require weeks of specialized knowledge now get done in hours with decent prompting. This matters for a couple of reasons. First, it gives defenders actual behavioral data instead of hypotheticals. Second, it directly explains why the Trump EO we covered yesterday put so much emphasis on AI-enabled cyber defense. When Anthropic is publishing reports like this, and when the government has already been briefed on what Mythos, Anthropic's most powerful and still restricted AI model, is capable of doing offensively, those voluntary cybersecurity frameworks start to feel a lot harder to ignore. One more, and it's the one that made Tesla investors nervous. OpenAI is getting back into robotics. Sam Altman announced a new division called OpenAI Robotics, focused on building general purpose robots for industrial infrastructure, factories, warehouses, power grids, data centers. Two job listings went live immediately, an actuator design engineer to build the robot's joints and muscles, and a DAX station engineer to build the sensor network, essentially the robot's nervous system. OpenAI actually had a robotics program well before ChatGPT took over the company's identity. They trained a robotic hand called Dactyl to solve a Rubik's Cube back in 2019, then shelved the whole program in 2021 when language models became the main event. Now it's back, with a specific industrial focus. The market noticed Tesla's stock dropped 4.57% on the announcement. Elon Musk's humanoid robot vision through Tesla's Optimus program just got a well-funded direct competitor with deep model capabilities already in place and an explicit mandate to automate the kinds of jobs that currently need skilled workers on industrial sites. We're not there yet, but the competition for physical AI is now very much on. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdayanaai.news. Or you can find me on LinkedIn, X, or Blue Sky. And if you like this podcast and want to see it continue, please be sure to take a minute to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.