AI in 15 — April 06, 2026

Google's smallest new AI model now runs on your iPhone, no cloud, no subscription, no internet required. And over the weekend, developers proved it can turn on your flashlight, open your maps, and write code, all from a two-billion parameter model running in under one and a half gigs of memory.

Welcome to AI in 15 for Monday, April 6, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Monday, Marcus. Big week ahead. Gemma 4's smallest model goes viral running on iPhones and Raspberry Pis, and we need to talk about what that means. Investors are fleeing OpenAI for Anthropic on the secondary market. Microsoft's legal team admits Copilot is quote for entertainment purposes only. Japan bets six billion dollars on robots to fill jobs nobody wants. OpenAI moves Codex to token-based pricing ahead of its IPO. Iran threatens to annihilate a thirty-billion-dollar AI data center in Abu Dhabi. And Bloomberg declares vibe coding has officially gone mainstream. Let's get into it.

Gemma 4's tiny model runs on your iPhone and the developer community loses its mind.

Investors dump OpenAI shares and race to buy Anthropic equity.

And Microsoft's own terms of service say Copilot is for entertainment only.

Marcus, we covered the Gemma 4 release on Friday, but something happened over the weekend that changes the story entirely. This model is now running on phones.

The E2B model, that's the two-billion parameter variant, is now available through the Google AI Edge Gallery app on both iOS and Android. iPhone 15 Pro or newer. And the Hacker News post about it hit five hundred and thirty-four points, which tells you the developer community sees this as a big deal. Users were posting videos of it running on their phones with genuinely useful results.

And it's not just chat. People are doing agentic things with it?

That's the breakthrough. Developers demonstrated mobile action tool calls where the model can turn on your flashlight, open maps, perform device actions, all running locally with zero cloud dependency. One developer built a real-time AI system with audio and video input and voice output on an M3 Pro Mac using the same model. On a Raspberry Pi 5, it hits a hundred and thirty-three prefill tokens per second. And it supports a hundred-and-twenty-eight-thousand token context window, over a hundred and forty languages, and multi-step agent workflows.

All in under one and a half gigs of memory. That's less than most photo editing apps.

And here's the comment from the Hacker News thread that I think captures the moment perfectly. The original poster wrote, quote, it is my firm belief that the only realistic use of AI in the future is either locally on-device for almost free, or in the cloud but way more expensive than it is today. Given everything we covered last week about Anthropic cracking down on subscription usage and the economics of cloud AI falling apart, that framing feels prophetic.

So we're watching two futures diverge in real time. Expensive cloud AI for heavy workloads, and free local AI for everything else.

And Google is betting massively on owning the local side. Apache 2.0 license means anyone can build commercial products on top of this. No restrictions, no royalties. They're trying to become the Android of on-device AI.

Now for a story that says a lot about where the AI industry stands right now. Investors are actively dumping OpenAI shares and rushing to buy Anthropic. Marcus, what's driving this?

The LA Times piece that hit Hacker News this weekend laid it out pretty starkly. OpenAI's valuation sits at eight hundred and fifty-two billion. Anthropic closed its Series G in February at three hundred and eighty billion. That was the second-biggest private financing round in tech history at thirty billion dollars. And bankers now expect Anthropic's IPO to value the company between four hundred and five hundred billion, with a raise exceeding sixty billion.

So the smart money thinks the gap is closing fast.

Or has already closed. The Hacker News comments were brutal toward OpenAI. The top comment said, quote, I don't know how OpenAI screwed this up. They had the best tech, the largest installed base, the best brand recognition. And somehow they got hubristic and sloppy and failed to iterate on the core product. Anthropic is approaching nineteen billion in annualized revenue. OpenAI has passed twenty-five billion. But the trajectory favors Anthropic.

Both companies eyeing IPOs. This secondary market shift is basically a preview of how public markets will react.

And it's not just sentiment. Developer mindshare has shifted meaningfully. Claude Code's adoption, the enterprise trust Anthropic has built, the interpretability research we covered yesterday showing they actually understand what's happening inside their models. These aren't superficial advantages. OpenAI still has more revenue, but Anthropic has momentum, and in tech, momentum is everything.

Okay, this next one is genuinely funny. Microsoft's Copilot terms of service say, and I'm quoting directly, Copilot is for entertainment purposes only. Marcus, entertainment purposes only?

The full quote is even better. Quote, Copilot is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don't rely on Copilot for important advice. Use Copilot at your own risk. End quote. The terms were last updated October 2025, and they also state Microsoft makes no warranty of any kind about Copilot and can't promise responses won't infringe copyrights or trademarks.

This is the product they've embedded in Office 365, Windows, and every enterprise workflow they sell. The product they charge thirty dollars a month per seat for.

The cognitive dissonance is extraordinary. Marketing says it's your AI-powered productivity assistant. Legal says it's a toy. Microsoft told reporters the language is quote legacy and will be updated. But that raises an obvious question. If you've been selling this aggressively for over a year, why does your legal page still say it's for entertainment?

One Hacker News commenter nailed it. Quote, go to copilot dot com and ask a question. You can see from the answer that it is clearly for entertainment only.

Ouch. But the serious point is about liability. Every AI company is walking this tightrope. They want enterprise contracts and consumer trust, but their lawyers won't let them actually stand behind the output. Microsoft just got caught with the gap between marketing and legal wide open for everyone to see. A hundred and twenty-two points on Hacker News means a lot of enterprise buyers saw this story.

Japan is taking a completely different approach to AI than the West. Instead of replacing workers, they're building robots for jobs that literally have zero human applicants. Marcus, this is fascinating.

Japan's demographic crisis is well known. The population is aging rapidly, the workforce is shrinking, and there are entire categories of jobs with no applicants. So the government announced in March its aim to capture thirty percent of the global physical AI market by 2040, backed by six point three billion dollars in investment under Prime Minister Takaichi.

And they're not starting from scratch. Japanese manufacturers already make seventy percent of global industrial robots.

That existing dominance is key. FANUC just announced a collaboration with NVIDIA to advance physical AI, enabling robots that interpret voice commands and generate Python code autonomously. The market is projected to grow from three hundred million in 2025 to nearly seven billion by 2035. And here's what I find most interesting. The Hacker News discussion highlighted the contrast with Western deployment. One commenter said, quote, meanwhile in the US they're replacing artists, writers, and teachers.

Japan uses AI to fill gaps. The West uses it to cut costs. Very different social contracts.

And Japan's early confrontation with demographics may end up being an advantage. They're building the infrastructure and expertise for embodied AI while the rest of the world is still arguing about chatbots.

Quick hit on OpenAI. Codex moved to token-based pricing as of April 2, replacing the old per-message credit system. What's the real story here, Marcus?

On the surface it's a transparency improvement. Token-based pricing aligns Codex with standard API economics. But the Hacker News thread with nearly two hundred points caught the subtext immediately. One commenter wrote, quote, Anthropic bundled Claude Code with the subscription because OpenAI bundled ChatGPT with Codex. Now OpenAI is unbundling. IPO must be around the corner.

Investors want clean revenue attribution per product line.

Exactly. You can't go public with opaque credit systems and bundled products. Wall Street wants to see what each product earns independently. The unbundling is an IPO preparation move, pure and simple. And it comes as competition from Claude Code intensifies.

Now for something that sounds like a thriller plot. Iran's Revolutionary Guard released a video threatening quote complete and utter annihilation of the thirty-billion-dollar Stargate data center being built in Abu Dhabi. Marcus, how seriously should we take this?

The video included satellite imagery of the facility and night vision footage revealing its full extent. The Stargate UAE project is a one-gigawatt compute cluster within a planned five-gigawatt campus spanning nineteen square kilometers. But here's the broader context. Iranian military channels circulated a list of eighteen US tech companies designated as potential targets, including Apple, Google, Meta, Microsoft, and NVIDIA.

Whatever you think of the threat's credibility, it exposes a real strategic problem.

The AI industry's rush to build massive data centers in the Middle East, drawn by cheap energy and friendly regulations, creates geopolitical risk that simply didn't exist when compute was concentrated in Virginia and Oregon. Insurance costs and security requirements for Gulf AI infrastructure are going up after this. The concentration of critical AI compute in a region with active military tensions is a strategic vulnerability that the industry has been ignoring.

And finally, vibe coding hit Bloomberg's front page. The mainstream has arrived.

Bloomberg published a major piece examining how non-developers in creative industries are now building tools through natural language instructions to AI. But the coverage wasn't all rosy. A companion piece explored how AI assistants may be causing more burnout, not less. And Fortune emphasized that trust is the real bottleneck. Meanwhile, a viral Hacker News post with seven hundred and twenty-seven points documented the honest reality. Initial excitement, then architectural challenges AI alone can't solve.

So we're entering the trough of disillusionment for AI coding?

The honeymoon is ending. The hard questions about reliability, architecture, and long-term maintenance are surfacing. That's actually healthy. The tools are genuinely useful, but the gap between demo and production is where the real work lives.

Monday big picture. AI runs on your phone for free, investors are reshuffling billions between AI companies, and the physical world is starting to catch up with the digital one. Marcus, what's the thread?

The AI industry is fragmenting into distinct lanes. On-device AI is becoming real and free. Cloud AI is getting more expensive and more scrutinized. Physical AI is finding its first genuine market in Japan. And the financial markets are placing very different bets on who wins each lane. The era of one company dominating all of AI is over. What we're watching now is specialization. And the winners won't be the companies that try to do everything. They'll be the ones that pick a lane and own it.

Pick a lane and own it. Good advice for a Monday morning.

Now if only the AI companies would take it.

That's your AI in 15 for Monday, April 6, 2026. See you tomorrow.