AI in 15 — April 16, 2026

Google just taught a robot dog to read a pressure gauge with ninety-three percent accuracy. Last generation? Twenty-three. That's not incremental improvement. That's a robot going from failing the eye exam to acing it overnight.

Welcome to AI in 15 for Thursday, April 16, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Thursday, Marcus. Great lineup today. Google DeepMind gives Boston Dynamics' Spot robot a serious brain upgrade. A scheduling company blames AI for going closed source and the internet isn't buying it. A federal judge rules your AI chats can be used against you in court. Anthropic now wants your passport. A shoe company becomes an AI company and its stock jumps six hundred percent. And the Stanford AI Index delivers some uncomfortable truths we haven't covered yet. Let's go.

Spot the robot dog gets Gemini-powered vision and can now inspect factories alone.

A federal court says your conversations with Claude are not privileged. At all.

And Allbirds sells its shoes, buys GPUs, and Wall Street loses its mind.

Marcus, Google DeepMind dropped Gemini Robotics-ER 1.6 yesterday, and this one is a concrete, real-world deployment story. What's the headline capability?

They call it agentic vision. The model combines visual reasoning with code execution so robots can interpret physical instruments, analog gauges, digital readouts, sight glasses, with what they're calling sub-tick accuracy. The jump from ER 1.5 to 1.6 on instrument reading went from twenty-three percent to ninety-three percent success rate. In a single generation. That's not a tune-up, Kate. That's a completely different level of capability.

And Boston Dynamics is the launch partner here.

Spot is now running ER 1.6 for autonomous facility inspections. Marco da Silva from Boston Dynamics said it enables Spot to operate completely autonomously in industrial environments. Reading gauges, navigating hazards, monitoring conditions, no human in the loop. The model can also natively call tools like Google Search, other vision-language-action models, or custom functions. So it's not just seeing, it's reasoning about what it sees and taking action.

This feels like one of those moments where foundation models stop being a research curiosity and start doing actual work.

That's exactly what it is. Industrial inspection is a massive market. Think oil refineries, power plants, manufacturing floors. These are environments where you need regular checks but where sending humans is expensive, slow, and sometimes dangerous. A robot that can autonomously walk through a facility, read every gauge, detect anomalies, and file reports? That's a real product solving a real problem. And it's available now through the Gemini API.

The pace of improvement is what gets me. One model generation, four-x improvement on the core task.

And ER 1.6 also improved on safety compliance, six percent better than Gemini 3.0 Flash on text safety and ten percent better on video scenarios. DeepMind is clearly positioning Gemini as the reasoning backbone for the entire robotics stack. This is Google playing the long game while everyone else fights over chatbots.

From robots to retreats. Cal.com, the open-source scheduling platform, announced yesterday it's going closed source after five years. And Marcus, the reason they gave is raising eyebrows.

Their CEO said, and I'm quoting, being open source is increasingly like giving attackers the blueprints to the vault. The argument is that AI coding tools have changed the security calculus. Attackers can now point AI at public codebases and systematically find vulnerabilities at a speed that wasn't possible before. They specifically referenced Anthropic's Mythos Preview finding that twenty-seven-year-old OpenBSD bug we covered last week.

Hacker News was not sympathetic.

Three hundred and eighteen points, a hundred and sixty-eight comments, and the consensus was pretty clear. Security through obscurity was debunked decades ago. Multiple commenters pointed out that open source projects actually benefit from community vulnerability reports. One developer said AI-powered security scanning helped them find and fix every legitimate vulnerability in their project. If they'd been closed source, those same flaws would have been exploited silently with nobody telling them.

So the community thinks this is a business decision wearing a security costume.

The most upvoted take was that Cal.com took a convenient excuse to do something they wanted to do anyway because open source SaaS businesses are hard. And look, they're not wrong that AI-powered vulnerability discovery is real. We've covered that extensively. But the answer to better offensive tools isn't hiding your code. It's using those same tools defensively. Closing source doesn't eliminate vulnerabilities. It just means fewer people are looking for them.

The irony of citing Anthropic's security research to justify closing your code is pretty rich.

One AI company's defensive research becoming another company's excuse to go proprietary. You can't make this stuff up.

Now for a story that should concern anyone who's ever typed something sensitive into an AI chatbot. A federal judge in New York just ruled that conversations with AI are not protected by attorney-client privilege. Marcus, walk us through the case.

United States v. Heppner. Bradley Heppner, former chair of a bankrupt financial services company, was charged with securities and wire fraud. He'd used Claude to prepare case reports to share with his attorneys, but he did it on his own, not at his lawyer's direction. Judge Jed Rakoff's reasoning was blunt on multiple grounds. First, Claude is not an attorney. That alone disposes of the privilege claim. Second, the communications weren't confidential because the platform's privacy policy explicitly states it collects data on inputs and outputs.

Wait, so everything you type into a cloud-based AI chatbot is potentially discoverable in court?

That's exactly what this ruling establishes. And it's a question of first impression nationwide, meaning it's the first federal ruling on this. Lawyers across the country are now warning clients that their AI conversations could be subpoenaed. The judge noted the result might have been different if Heppner had used Claude at his attorney's specific request. But for everyone else? Treat every AI conversation as if it could end up in a courtroom.

This has implications way beyond this one case.

Massive implications. Every business executive brainstorming legal strategy with ChatGPT. Every founder asking Claude about regulatory compliance. Every employee discussing workplace disputes with an AI. All of it is potentially discoverable. And it strengthens the case for local, on-device AI models when you're dealing with anything sensitive. If the model runs on your hardware and never sends data to a third party, the calculus changes entirely.

Speaking of Anthropic and trust, they quietly started requiring government ID from some Claude users this week. Marcus, what do we know?

Anthropic updated its help page to announce that some users will need to submit a government-issued photo ID and a live selfie through a third-party provider called Persona. This isn't universal. It triggers for specific advanced features, platform integrity checks, or when safety and compliance requirements kick in. Acceptable documents are passports, driver's licenses, or national identity cards.

And the reaction?

On Hacker News, a hundred and fifteen points, eighty-eight comments, mostly negative. Users flagged Persona's own privacy policy, which reportedly allows training its models on verification data and sending information to up to seventeen subprocessors. Some developers called it a disaster for adoption. The sharpest criticism framed it as a step toward a future where programming a computer in any meaningful way requires total identification and permission.

And this lands the same week Claude had a major outage affecting over seven thousand users. As we reported yesterday, the outage hit Claude.ai, the API, and Claude Code all at once during peak hours. If you're asking people for their passport, the service needs to actually work.

The timing is terrible. Seven thousand users affected, login failures, API errors, the incident lasted nearly three hours. Hacker News commenters noted this is becoming a pattern around peak US hours. One person built a real-time tracker for Claude's reliability. Another quipped that once AGI is achieved, they'll reach the fabled superhuman two nines of uptime. Look, Anthropic is building incredible technology. But the combination of ID requirements, reliability issues, and the quota frustrations we covered Monday is testing user loyalty.

Okay Marcus, I saved the best for last in this section. Allbirds. The shoe company. Is now an AI company. And its stock went up six hundred percent.

Peak 2026, Kate. Allbirds sold its footwear assets and branding for thirty-nine million dollars a few weeks ago. Then announced a fifty million dollar convertible financing facility to pivot into GPU-as-a-Service and AI-native cloud solutions. The stock surged five hundred and eighty-two percent in a single day, bouncing between two dollars and twenty-four dollars before settling around twelve-eighty after hours.

GPU-as-a-Service. From a shoe company.

They're rebranding as NewBird AI, according to Bloomberg. Hacker News immediately compared it to Long Island Iced Tea Corp rebranding as Long Blockchain Corp in 2017. And honestly, the comparison is apt. Fifty million dollars is a laughably small amount to build AI data center infrastructure. Engadget's headline was a sign of a totally normal and healthy economy. My personal favorite comment was someone who said they'd like to buy some sneakers and also some compute capacity to run AI inference.

Is this a canary in the coal mine for an AI bubble?

It's certainly a data point. When a company with no AI assets, no AI expertise, and no AI revenue can see its stock jump six hundred percent by announcing an AI pivot, that tells you something about investor psychology. We talked Monday about the market demanding proof from AI companies. Apparently that discipline doesn't extend to former footwear brands. This is the kind of thing that happens in the late stages of a hype cycle. Whether we're late-stage or mid-stage, I'll leave that to the market to decide.

Quick hit on the OpenAI Agents SDK. They shipped a significant update yesterday with sandboxing and a frontier model harness for enterprise developers.

The sandboxing feature lets agents operate in controlled environments, siloed workspaces where they can access files and code for specific operations while protecting the broader system. This is a critical missing piece for enterprise agent deployment. Without sandboxing, giving agents file system access is a security nightmare. Python support first, TypeScript coming later. It's OpenAI's bid to become the default platform for production agents, competing directly with LangChain, CrewAI, and Anthropic's agent tooling.

And Google launched a native Gemini desktop app for Mac. Option plus Space to summon it from anywhere, screen sharing built in.

It's free, available to anyone thirteen and older on macOS 15. Direct challenge to Claude's desktop app and Apple Intelligence. The screen-sharing capability means it can analyze whatever you're looking at and provide contextual help. Early reactions were mixed. You can't access past chats without enabling data sharing, no model selection beyond speed modes, and Google's on-device Gemma models aren't available through it. But Google's distribution advantage is enormous. Anyone with a Google account is a potential user.

Thursday big picture. Marcus, robots are reading gauges autonomously, courts are reading our AI chats, shoe companies are buying GPUs, and Anthropic wants our passports. Connect the dots for me.

The common thread is that AI is leaving the sandbox. It's entering physical spaces with robots, entering courtrooms through legal precedent, entering financial markets through speculative pivots, and entering identity systems through verification requirements. Every one of these stories is about AI crossing a boundary it hadn't crossed before. And at every boundary crossing, the stakes get higher and the margin for error gets smaller.

And the institutions, courts, markets, companies, are scrambling to figure out the rules as they go.

Judge Rakoff didn't have case law to cite. Boston Dynamics is writing the playbook for autonomous industrial robots in real time. The market apparently hasn't figured out how to distinguish a real AI company from a shoe company with a press release. We're in the messy middle, Kate. The technology works. The frameworks for governing it don't. Not yet.

The messy middle. I think that's where we'll be for a while.

That's your AI in 15 for Thursday, April 16, 2026. See you tomorrow.