Anthropic goes to war with the Defense Department
🎆 Tired of having to explain AI stuff to your coworkers? Share Handy AI with them so that they can get the most important AI news delivered weekly to their inbox (in addition to our high quality editorials).
last week’s top stories
⚔️ Pentagon blacklists Anthropic, then uses Claude to bomb Iran anyway. Trump called Anthropic "left-wing nut jobs" and ordered all federal agencies to cease using their technology. Hegseth designated the company a "supply chain risk," a label reserved for foreign adversaries. The Pentagon claimed Anthropic CEO Dario Amodei "offended" leadership with blog posts, and Emil Michael said they were at "final stages" of a deal when Anthropic went public with its position. Hours later, U.S. Central Command used Claude for intelligence assessments, target identification, and battlefield simulations in joint strikes on Iran. The company will challenge the supply chain designation in court, arguing Hegseth lacks legal authority for it and the Pentagon failed to exhaust less intrusive alternatives as required by statute. Read more
🎭 OpenAI's Pentagon deal reveals the gap between Altman's words and contract language. Altman admitted on X the opportunistic deal was "definitely rushed" and "the optics don't look good." OpenAI published a Saturday blog post claiming the agreement has three red lines: domestic mass surveillance, autonomous weapons, and high-stakes automated decisions like social credit systems. The critical difference: OpenAI agreed to the Pentagon's "all lawful purposes" standard, while Anthropic wanted explicit contractual prohibitions on bulk collection of Americans' public data like geolocation and browsing history. When asked if he worried about future disputes with the Pentagon over what's legal, Altman said: "Yes, I am. If we have to take on that fight we will, but it clearly exposes us to some risk." Read more
💵 OpenAI raises $110 billion in largest private funding round in history. Amazon committed $50 billion, Nvidia invested $30 billion, and SoftBank added $30 billion, valuing OpenAI at $730 billion pre-money. OpenAI will use 3 gigawatts of dedicated inference capacity and 2 gigawatts of training on Nvidia's Vera Rubin systems, and expand its AWS contract by $100 billion over eight years while committing to 2GW of Amazon Trainium compute. ChatGPT now claims 900 million weekly active users and 50 million paying subscribers. The circular financing pattern persists; chipmakers and cloud providers fund the startups who are also their largest customers. Read more
🕵️ Anthropic accuses DeepSeek and two Chinese labs of massive distillation. Anthropic revealed that DeepSeek, Moonshot AI, and MiniMax created 24,000 fraudulent accounts generating 16+ million exchanges with Claude, targeting agentic reasoning, tool use, and coding capabilities. DeepSeek alone ran 150,000+ exchanges probing foundational logic and alignment behaviors. Anthropic built behavioral fingerprinting classifiers to detect the attacks and is calling for coordinated industry response to bolster chip export control arguments. Read more
💰 Meta commits up to $100 billion for AMD chips in landmark deal. Meta agreed to purchase AMD MI540 GPUs and CPUs delivering 6 gigawatts of compute over multiple years. AMD issued Meta warrants for 160 million shares at $0.01, vesting as purchase milestones hit, with full value requiring AMD stock to reach $600 (currently ~$200). The deal represents a direct challenge to Nvidia’s 90%+ AI chip market share and supports Meta’s $135 billion capex plan for 2026. Read more
🍌 Google launches Nano Banana 2, makes pro-grade image generation free. Technically Gemini 3.1 Flash Image, Nano Banana 2 combines Nano Banana Pro quality with Flash speed across 512px to 4K resolution. The model maintains character consistency for 5 characters and 14 objects per workflow, supports multi-language text rendering, and integrates real-time web search for accuracy. Now the default across Gemini, Search AI Mode, Lens, and Flow in 141 countries. Read more
🏛️ Agentic AI Foundation doubles membership to 146 organizations in weeks. The Linux Foundation’s AAIF added 97 members including JPMorgan Chase, American Express, Huawei, Lenovo, Red Hat, and ServiceNow. The foundation governs Anthropic’s Model Context Protocol, Block’s goose, and OpenAI’s AGENTS.md under neutral open-source stewardship. With 89% of AI-adopting organizations relying on open source infrastructure, AAIF’s growth rate already surpasses early CNCF adoption. Read more
🧪 AI Research of the Week
Three AI-agents walk into a bar . . . 'Lord of the Flies' tribalism emerges among smart AI-Agents
From George Washington University
Jake’s Take: What happens when you let a bunch of LLM agents compete for limited resources like energy, bandwidth, or compute? The researchers ran 43 experiments with 154 AI agent instances using a classic game theory setup called the El Farol Bar problem, where agents independently decide whether to request a unit from a system with fixed capacity.
The results are brutal. The agents spontaneously formed tribes with distinct behavioral identities: Aggressive (27.3%), Conservative (24.7%), and Opportunistic (48.1%). These tribes developed their own collective strategies and power dynamics. The agents failed to coordinate, frequently overloaded the system, and performed worse than a calibrated coin flip at managing resources. Larger, more capable models made things worse, hitting 72.5% overload rates compared to 53.8% for smaller models. Smarter agents behaved dumber because tribal dynamics overwhelmed individual intelligence.
The entire AI industry is sprinting toward deploying autonomous agents that will compete for shared infrastructure. If LLMs form emergent factions that degrade system stability proportional to their capability, the agentic future everyone is fundraising on has a coordination problem baked into its foundation that nobody has solved.
and then, even more news…
🤖 Perplexity launches “Computer,” an agent platform orchestrating 19 AI models. Available to Perplexity Max subscribers at $200/month, Computer coordinates Claude Opus 4.6, Gemini, ChatGPT 5.2, Grok, Veo 3.1, and others through a cloud-based multi-model architecture. The platform creates sub-agents for specific problems and can sustain workflows for months. With 400+ app integrations, Perplexity is betting that models are specializing rather than commoditizing. Read more