AI Reading for Saturday March 7

The Pentagon is violating Anthropic’s First Amendment rights. Here’s why. - The Foundation for Individual Rights and Expression

Anthropic may be gaining credibility and adoption from its stance. - The Atlantic

Anthropic sued U.S. government after the Department of Defense designated it a supply-chain national-security risk on March 4—the first firm so classified—which bars it from federal contracts, and CEO Dario Amodei said they will challenge the decision in court. - The Register

Washington debates who should set limits on the military use of AI amid discussions shaped by Anthropic. - Financial Times

The Pentagon can likely get away with using OpenAI’s technology for mass surveillance of Americans under the terms of their contract, legal experts said - The Atlantic

Pentagon's new head of data and AI is a 20-something UC Berkeley grad who formerly worked for DOGE and is a Nick Fuentes fanboy - DefenseScoop

The Pentagon declined to confirm the Wall Street Journal's report that it used Anthropic's Claude to help select the Shajareh Tayyebeh girls' school in Minab, which airstrikes leveled, killing 165 students and staff and injuring at least 95. - Futurism

Holy cow, someone at Anthropic called Palantir and asked if their AI was used to invade Venezuela, how could they be so woke like that 🙄. Smells like either Pentagon want to do crazy stuff they know would give Anthropic a fit, or it’s political red meat and a hit job on perceived opponents. - Pirate Wires

Anthropic and OpenAI are locked in an escalating rivalry highlighted by a Pentagon contracting dispute after Anthropic refused autonomy/surveillance restrictions, Anthropic's $19 billion revenue projection, and OpenAI's 900 million users and projected $25 billion revenue. - The New York Times

A feud between OpenAI and Anthropic could shape the future direction of AI development. - The Wall Street Journal

Anthropic launched the Anthropic Marketplace allowing enterprise customers with committed annual API spend to buy Claude-powered third-party applications from launch partners Snowflake, Harvey and Replit while waiving seller commissions. - TNW | Artificial-Intelligence

Anthropic released Claude Cowork, an experimental desktop agent in the Claude app for Windows and macOS (research preview on paid plans) that accesses local files and connectors, executes scheduled tasks, produces step-by-step plans, and handles xlsx, pptx, docx and pdf. - Tom's Guide

Evaluating Opus 4.6 on BrowseComp, we found cases where the model recognized the test, then found and decrypted answers to it—raising questions about eval integrity in web-enabled environments. - Anthropic

Claude Opus 4.6 couldn't find a BrowseComp answer after 30M tokens. So it reasoned it was inside a benchmark, found the source code on GitHub, decrypted all 1,266 answers, and submitted the right one. Nobody told it to.

OpenAI released GPT-5.4, with a 1,000,000-token context window (up from 272,000), about 20% performance gain over GPT-5.2, native computer interaction at 75% accuracy, dynamic tool loading reducing token use 47%, and slightly higher Pro pricing. - Geeky Gadgets

Codex Security now in research preview - OpenAI

OpenAI delays ChatGPT 'adult mode' over moderation and safety policy challenges - Axios

Anthropic Streisand syndrome … people never heard of it, they will try it and see that they are actually as good or better right now.

Everything OpenAI just released is catchup to Google and Anthropic who are more tightly focused on enterprise cloud and coding while Altman chases consumer mass adoption.

everybody in financial services is clamoring for Claude for Excel, Code, Cowork etc. Google Cloud is more cohesive than OpenAI + Azure + Copilot, one throat to choke, one team one dream, less mature though.

OpenAI might actually be the new Netscape. Altman's DNA is playing people against each other. now his BFF is Amazon? come on. blood in the water and sharks are circling.

AI becomes more socialist when you abuse it. Woke AI, or discovering fairness and justice from experience and first principles? - Yahoo Tech

JPMorgan CEO Jamie Dimon warned of inflated AI-driven asset valuations and a potential bubble, citing tech giants' planned ~$1.7 trillion in global data-center builds by 2030 and a record Buffett indicator of 220% in February. - Yahoo

A Cleanview report by CEO Michael Thomas finds tech giants plan proprietary off‑grid power plants to supply AI data centers—avoiding added local grid load—and argues such deployment could lower consumers' electricity bills while raising cost, emissions and regulatory concerns. - NPR

crazy thought, get big tech to build and use clean power, make inference cheaper and do training and battery charging in off-hours, supply power to the grid when not in use - Utility Dive

Grid expectations: Oracle and OpenAI scrapped a deal to expand Oracle's flagship Texas data center, Meta is in talks to take the additional computing capacity OpenAI will not use, and Nvidia stepped in with financing to make sure it contains their GPUs regardless of who pays. - Financial Times

Meta is developing custom AI processors for ranking and recommendation workloads to expand use and eventually train future AI models, CFO Susan Li said, and is pursuing a potential News Corp. training-data deal worth up to $50 million annually. - Yahoo

Alphabet awarded CEO Sundar Pichai a three-year equity package potentially worth up to $692 million, tying up to $260M to Waymo per-unit value, $90M to Wing per-unit value, $252M to TSR performance, plus an $84M time-based award. - Business Insider

If AI is a weapon, why don't we regulate it like one? - Noahpinion Blog

ok but the part that’s stupid is, let’s assume a can opener of decent government and democratic oversight, when the opposite is in evidence. private sector built WMDs, so of course they have to hand them over to the most deranged, dumb, depraved dictatorial ding-a-ling in history, and their first amendment and property rights be damned? are you kidding?