Good morning. It's Wednesday, October 14, and we're covering Anthropic's fight against AI model theft, why advanced AI remains out of reach for most people, and Google's latest frontier model release.

Plus: yesterday's poll results on AI agent deployment, a guest essay on the limits of AI creativity, and a new Real or AI challenge.

Google released Gemini 4 Argon, a new frontier model built for difficult coding, science, legal, finance and cybersecurity work. It is available through Google’s developer tools and to Gemini Pro and Ultra subscribers, with a one-million-token limit for long, multi-step jobs.

Anthropic is reportedly nearing a deal with NTT Data for its first Japanese data center, a project in Osaka that could draw up to 500 megawatts of power and cost about $6.3 billion. The companies aim to sign by the end of 2026, with part of the facility operating as early as 2028. (Paywall)

Palantir CTO Shyam Sankar told CBS’s Face the Nation that some AI researchers have started smoking because they believe the technology will cure cancer before the habit harms them. He offered the story as an example of extreme faith in AI; the report did not establish how common the behavior is. (Paywall)

AI made code cheap. It did not make changes safe to ship. Agentic Change Management gives teams the controls around agent output: Review every change with codebase context, prioritize a queue hundreds of PRs deep, understand big diffs in minutes, and secure the whole codebase.

Bring Agentic Change Management to your PR today.

We’re living in unprecedented times on the internet.

We’re being sold a future of infinite self-optimizing content, one-shot brand campaigns, and fully automated taste. A new class of founders are leaning into this vision, claiming you can prompt your way into building a world class brand. → Read the full article here.

Who will actually capture the economic value AI creates? Will today's leading AI labs maintain their advantage? And are we moving too quickly toward systems we don't fully understand?

In this exclusive interview, Matthew Berman sits down with Christian Catalini, an MIT economist and expert on digital platforms and AI, to explore the economic forces shaping the next phase of artificial intelligence.

For existing members: Watch the Interview →
For nonmembers: Unlock with Frontier Pass →

Anthropic says new defenses reduced by roughly 90% to 99% the amount of Claude reasoning data stolen through distillation attacks. Distillation is a shortcut in which another company sends millions of requests to a stronger model, collects its answers and reasoning, then uses that material to train a smaller model to copy its abilities.

Anthropic now limits where Claude’s hidden reasoning can move, uses classifiers to block extraction attempts and bans networks of fraudulent accounts—sometimes within two hours. The tradeoffs are real: some developers had to change their integrations, a small share of legitimate requests can be blocked and Anthropic now charges for blocked requests in areas where its false-positive rate is below 0.1%. → Read the full article here.

AI researcher Andrej Karpathy estimates that about six billion people have barely used a large language model, while only around 5,000 people inside leading labs have access to the most capable systems. Box CEO Aaron Levie says that gap explains why most people still think of AI as a chatbot even as a small group uses agents to work for hours and change large codebases.

The capability is not the same as deployment. Companies still have to connect their data, set permissions, control costs and decide who is responsible when an agent makes a mistake. Levie expects extensive change management before agent teams become normal outside software and AI research. → Read the full article here.

Satlyt Raises $8 Million for Satellite AI: Startup funds software that shares computing across satellites, with NASA-related tests planned next.

Google Play Adds Private AI-Built Apps: U.S. adults can install personal Android apps built with Google AI Studio without public Play Store listings.

Character.AI Files for Chapter 11: Character.AI plans a $20 million Disney asset deal; its chatbot platform could close if no buyer emerges.

SpaceXSI Bundles X, Grok and Cursor: Five subscription tiers combine X Premium, Grok and Cursor, ranging from free to $200 monthly.

Amex Adds Permission Checks for AI Purchases: Its new framework records budgets and product rules before AI agents receive payment credentials.

Here's how you voted: Caution won, with nearly half saying AI agents aren't ready, while 21% support deployment with human oversight.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team