Intro
Mythos, the hacker model too capable for the public, accessed by guessing the website
You can’t make this up. Last month, Anthropic delayed the public release of its newest model, Claude Mythos, claiming it was so powerful it could be used to uncover economy-wrecking cybercrime attacks and would therefore remain unreleased until it could help shore up some of the internet’s problems. Too bad they didn’t start with the model itself.
Evidently, some AI “enthusiasts” correctly guessed the URL and gained early access. Sigh.
Meanwhile, some enterprising genius figured out how to get the McDonald’s website support bot to pay for Claude Code access. Maybe not surprising: The Pentagon still has Anthropic blacklisted.
Tech Giants
Microsoft + OpenAI: 1+1=3
I’m no financial analyst, but when Om Malik breaks down Microsoft’s growing AI revenue into words I understand, I can start to see why folks are so worried about a financial bubble.
Microsoft buys equity in OpenAI (an investment) → OpenAI gives that money right back to Microsoft for compute (revenue) → meanwhile the OpenAI investment increases in value (the return) — and just like that $1 turns into $3. This is happening with Anthropic on Google and Trainium on Amazon too.
Om recalled the dot-com days when Lucent loaned money to companies to buy products. It worked great until those companies couldn’t pay. Rumors of OpenAI’s CFO expressing concern over growth not covering compute costs are starting to sound familiar.
Advertising
Can AI read your site? Probably not all of it, and now Adobe lets you see why
Adobe released survey results that bolster the argument that AI-sourced retail traffic is growing and arrives with significantly higher intent: 42% higher conversion rates and 48% longer time on site.
The company also released a Chrome Extension to help benchmark your site’s AI accessibility. Across U.S. retail sites, it found a staggering 25% of content isn’t accessible by AI at all.
How is this possible? ChatGPT’s browser agents struggle with commonly used browser technologies like JavaScript, infinite scroll, and video. These formats aren’t machine-friendly, so LLMs just skip them. Now you can see why.
AI
Meta’s new “AI Connectors” are a textbook example of using AI to help customers to spend more
If you’ve ever tried to spin up an ad campaign on a new platform from scratch, you’ll know how complicated it can be. Understanding a platform’s workflow, specific parameters, and the campaign learning thresholds is confusing, and it all differs between platforms, even campaign types. For instance, Meta needs 50 conversions per week to exit the learning stage on app performance marketing campaigns. TikTok only requires 25. That kind of knowledge either lives as institutional wisdom — learned the hard way by ad experts — or buried somewhere in the depths of the platform documentation.
These intricacies cause friction that can slow or prevent advertisers from running successful campaigns. Translation: it hurts sales. That’s why Meta released an open beta of “AI Connectors,” a set of tools promising to help streamline the process of creating and managing ad campaigns on Meta. The tools appear to be an LLM with access to a Meta MCP server, designed to help with reporting, management, and measurement of ongoing ad campaigns. The idea: you get a chatbot with knowledge of your campaigns and the technical know-how to get the best results. It’s like having the foremost expert on the Meta platform personally help you run your campaign.
Virtually every company is exploring how to use MCPs and LLMs to enable their clients, and this is a great example of a flywheel at work: Meta helps clients run better campaigns, clients get better ROAS, clients spend more with Meta. How much more effective would you be if Meta’s best engineer and best account manager were assigned full-time to your single account?
This is how Meta can simultaneously lay off 10% of its workforce, eliminate an additional7% of planned hirings, but still plan for growth. Whether it works is another question.
Also, they're now building robots?
Amanda's Take
The next iPhone won't come from Apple
Apple just turned 50. Over five decades it went from garage startup to the most valuable company on earth, and the iPhone is the product that cemented that legacy. But when discussing Tim Cook’s retirement on a recent episode, the Hard Fork crew called out something striking Apple hasn’t invented the next defining personal device.
Sure, there’s Apple Vision Pro — cool and technically impressive — but it hasn’t become a mainstream product. It’s ultimately a bust. But the takeaway isn’t that Apple can’t innovate, it’s that it doesn't need to.
The iPhone is still the dominant personal device. It owns the most valuable users, the strongest ecosystem, and, critically, the distribution layer that everything else depends on. Why would Apple disrupt its own advantage? Now Cook is stepping down and handing the keys to John Ternus, Apple's head of hardware engineering — a guy who has spent his career obsessing over what the device is, not what it thinks. That tells you where Apple’s priorities remain.
The crack in that logic is AI. Siri has been “getting an overhaul” for years and still hasn’t meaningfully evolved. Meanwhile, users are increasingly doing AI-native tasks through apps like ChatGPT or Perplexity AI that sit on top of the OS, not within it. Your iPhone has become a gateway to AI, but not an AI-native experience itself. Apple is letting other companies own the interaction layer.
The market structure underneath all of this matters. Globally, Android dominates with ~70% share, largely driven by affordability and emerging markets. In the U.S. and other premium markets, iOS holds close to 60% share, and iOS users also spend dramatically more on apps. Android has comfortably settled into second place — profitable, but never a real threat to Apple’s dominance
But we may have a new contender. Reports recently surfaced that OpenAI is actively building a smartphone with real supply chain partners lined up and mass production targeted for 2028. Sam Altman posted the same day the news broke that it "feels like a good time to seriously rethink how operating systems and user interfaces are designed."
If Apple keeps sleepwalking on Siri while OpenAI ships a device where ChatGPT is the interface, we might finally see a legitimate category threat to the iPhone — the first since Android launched in 2008.