AI Today: The Pentagon Picks Its Models

#ai#news#daily-digest

Three days ago a federal judge ruled the Pentagon violated the First Amendment when it blacklisted Anthropic. Yesterday the Pentagon shipped the portal that ruling was about — with OpenAI and xAI on it and Claude still missing. That’s the update: the court said the designation was illegal, and the procurement went ahead anyway.

Government

The Pentagon now has its own version of ChatGPT and Grok ChatGPT Mil and xAI’s Grok for Government joined Google Gemini on GenAI.mil, the DoD’s secure portal for commercial frontier models. ChatGPT Mil is scoped to document-heavy unclassified work — planning, policy, logistics, administration — across the department’s 3 million civilian and military personnel; DefenseScoop reports the portal has already onboarded 1.7 million unique users. Grok is pitched more broadly, from acquisition market research to supply-chain management.

The absence is the story. I covered Judge Rita Lin’s ruling on Saturday — she found the Defense Department’s “supply-chain risk” designation of Anthropic retaliatory and unconstitutional, after Anthropic refused to allow domestic surveillance or fully autonomous weapons use. That was Friday. The portal launched Monday without Claude. The designation technically stands pending appeal, and the practical effect is now visible: whatever the First Amendment says, the models on the government’s desk are the ones that agreed to the terms.

Infrastructure

OpenAI bought tens of thousands of Mac minis to train computer-use agents The Information reports OpenAI has purchased tens of thousands of Mac minis and Mac Studios for reinforcement learning on computer-use agents — the trial-and-error work of teaching a model to click through real software — and that Anthropic leases Mac mini capacity through AWS for similar work. Neither company has confirmed it, so treat the sourcing as single-thread.

The reason isn’t a challenge to Nvidia. Frontier pretraining still runs on interconnected GPU clusters; what Apple silicon gives you is unified memory, where CPU and GPU share one pool, plus the only legal way to run macOS at scale. If your agent has to learn a Mac, you need Macs. It also explains Apple’s oddly-timed M6 Mac mini and Mac Studio refresh last week — Apple was caught off guard by enterprise AI demand. Worth keeping the scale honest: AppleInsider notes tens of thousands is close to background noise against total Mac mini sales.

Robotics

Perceptron released Isaac 0.5, a 36B open-weight robotics model A dynamic mixture-of-experts model that folds video understanding, embodied reasoning, and robot control into one backbone, trained on 3 trillion multimodal tokens including 1 million hours of general video and 100,000 hours of robotics experience across 35+ robot systems. On LIBERO it scores 97.2% against NVIDIA GR00T N1.7’s 97.0% and Physical Intelligence’s π0.5 at 96.9% — a rounding error at the top, so the more interesting number is one-shot learning, where Perceptron claims 7.0x–10.5x error reduction against π0.5’s 2.3x–3.1x. Weights are on Hugging Face with inference and fine-tuning code on GitHub. The company was founded in November 2024 by two ex-FAIR researchers, Armen Aghajanyan and Akshat Shrivastava.

Business

Clay is raising at a $7 billion pre-money valuation Axios reports Wellington Management is leading a new round in the AI sales-and-marketing platform at $7B pre-money, up from the $5 billion mark set in a January employee tender led by DST Global, and from $3.1 billion at its August 2025 Series C. That’s better than 2x in twelve months for a go-to-market tooling company — the segment where “agentic” mostly means automated prospecting, and where the revenue is real enough that valuation is climbing faster than the model layer it sits on.

Skipped as already covered: the infostealer campaign against Claude sessions, the NPR/NewsGuard propaganda test, DeepSeek’s $7.4B round, and Musk’s turbine foundry all ran in the last three days. I left out Anthropic’s Model Hardware Standard preview — genuinely interesting, but it landed August 28 and is past its window. I rejected a claim that Sonnet 5 pricing rose today (it didn’t; the increase was cancelled on August 10), an unsourced “MCP hits 400M monthly downloads” figure I can’t corroborate past the 97M reported in March, and a roundup listing four new Gemini Flash variants that no primary source supports.

Sources