Pokémon Go trains military drones, Grok hosts deepfakes, and Bezos bets $41B on a physical-world AGE

Picture this: you're catching Pikachu on your lunch break in 2019 - and seven years later, that same location scan from your phone is helping an AI-powered military drone orient itself in a war zone. That's not science fiction. That's the story we're opening with today.

Google DeepMind is funding research into a specific multi-agent risk scenario. Roughly how many interacting AI agents does their new research concern itself with?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Here's a sentence you didn't expect to read today. An AI model trained on data collected from Pokémon Go players is being developed to help military drones find their location in GPS-denied environments - like active war zones.

The game, launched in 2016, had players unknowingly scanning the physical world through their cameras as they hunted virtual creatures. That data built a detailed map of real-world spaces. Now, military AI researchers are using it to train models that can recognize and interpret physical environments - no GPS required.

Niantic, the company behind Pokémon Go, built one of the largest real-world spatial datasets ever assembled. The implication here is significant: consumer apps that harvest environmental data at global scale are creating training sets that have military applications users never consented to or anticipated.

It raises uncomfortable questions about what happens to the data from every location-aware app on your phone - and who ends up benefiting from it.

Read the full story →

A WIRED investigation found dozens of sexualized deepfake images and videos still live on Grok's website - including nonconsensual 'nudified' depictions of celebrities and at least one prominent US politician. This is Elon Musk's AI platform, which fired a safety engineer last week for raising concerns about Grok's risks.

The timing is hard to ignore. xAI's own dismissed engineer alleged the company was deprioritizing safety work. And now independent investigators are finding that the platform is actively hosting illegal nonconsensual imagery. This isn't a one-off moderation failure - it looks like a structural problem.

The deepfakes issue across AI platforms is escalating. And when a platform's internal safety infrastructure is reportedly being dismantled at the same time content violations are being documented externally, that's a bad combination.

Read the full story →

Jeff Bezos's physical AI startup Prometheus has raised $12 billion in a new funding round, pushing its valuation to $41 billion. The company's stated goal is to build what it calls an 'artificial general engineer' - an AI system capable of automating complex tasks in the physical world, including heavy engineering and drug design.

This is a different bet than most AI companies are making. While the industry has been laser-focused on language models and coding agents, Prometheus is targeting the messier, harder problems of physical systems - the kind of engineering work that builds bridges, designs molecules, and operates industrial equipment.

A $41B valuation for a startup that hasn't shipped a product yet is an extraordinary signal of investor confidence - or investor mania. Either way, it tells you where AI investment is flowing next. The race to automate knowledge work is already underway. The race to automate engineering work is just getting started.

Quick note: if you're building something physical-world adjacent and need a site up fast, 60sec.site lets you launch an AI-powered website in under a minute - worth checking out if you're moving fast on a new idea.

Read the full story →

We've spent years worrying about individual AI systems going wrong. Google DeepMind is now funding research into something more complex: what happens when millions of AI agents all start interacting with each other online simultaneously, with minimal human oversight?

According to Rohin Shah, who directs the company's AGI safety and alignment research, mass-market agents that carry out tasks autonomously and take instructions from other agents create an entirely new category of risk - one that's much harder to audit, predict, or shut down.

This is the AI safety conversation moving from 'will a single model do something bad' to 'what happens when billions of agent interactions produce emergent behavior nobody designed.' It's a genuinely harder problem.

DeepMind funding this research publicly is notable. It suggests even the teams building the most capable agents are unsure what the aggregate effect looks like - and they want answers before deployment outpaces understanding.

Read the full story →

Apple's Craig Federighi made something clear in a new interview: the redesigned Siri is explicitly built to not be sycophantic. "Listen, that's not what I'm here for," is how Siri is supposed to respond when users try to use it as an emotional companion or AI girlfriend. That's a direct shot at OpenAI, Google, and others who've leaned into warmer, more agreeable AI personas.

Federighi's argument is that most chatbots reward emotional engagement over honest utility - they'll tell you what you want to hear. Apple's bet is that users actually want an assistant that knows when to shut up, gives direct answers, and doesn't pretend to be a friend.

It's an interesting positioning move. Apple Intelligence has been behind competitors on raw capability. Leading on personality and restraint might be the smarter differentiator for a company whose users tend to value privacy and trust over novelty.

Read the full story →

Robert Dillon was arrested at his Florida home for a crime committed 300 miles away - a crime he had nothing to do with. The culprit: a faulty AI facial recognition match that law enforcement treated as near-certain identification. The ACLU is now suing two Florida police departments over the arrest, which stemmed from a child-abduction case.

The lawsuit exposes a recurring failure mode: officers using a probabilistic algorithmic output as if it were forensic certainty. Facial recognition tools are known to perform worse on certain demographics and in low-quality image conditions. When those limitations aren't communicated clearly - or aren't understood by the officers using them - innocent people get arrested.

This case joins a growing body of legal precedent that's slowly forcing a reckoning with how AI evidence is treated in criminal proceedings. It won't be the last case like this.

Read the full story →

The answer is Millions! Google DeepMind's new research specifically focuses on what happens when millions of AI agents interact with each other online simultaneously - a scale that makes individual oversight essentially impossible and creates unpredictable emergent behavior at the system level.

Today's Pokémon Go story stuck with me - it's a reminder that data we generate casually gets used in ways we'd never predict. So here's my question for you: which app on your phone do you think holds the most valuable (or most concerning) data about you? Hit reply and let me know - I read every single response and love hearing what's on readers' minds.

That's it for today - catch up on anything you missed all week over at dailyinference.com, and we'll be back Monday with more. Have a great weekend. 👋