AI Reading for Tuesday February 17

Supply Chain Kill Chain: The Pentagon threatened to terminate a $200 million contract with Anthropic after the company refused to allow its AI to be used for mass surveillance or fully autonomous weapons, citing the need for unrestricted military support. - The New Republic

Defense Secretary Pete Hegseth is "close" to cutting business ties with Anthropic and designating the AI company a "supply chain risk" — meaning anyone who wants to do business with the U.S. military has to cut ties with the company, a senior Pentagon official told Axios. - Axios

Anthropic says, you can't do these 2 things, DoD says, then no one can use Anthropic even for other stuff. You just did a deal with Salesforce and now you want to tell them they can't use Claude Code anymore … a lot of stuff might get delayed or prices might go up from this sort of nuclear option, you must follow our policy or we will seriously mess up your business.

India puts on a huge show to showcase and attract AI investment. Modi’s favorite entrepreneur Adani Enterprises plans to invest $100 billion by 2035 to build renewable-powered AI-ready data centers in India, aiming to expand capacity to 5GW and establish a $250 billion infrastructure ecosystem through partnerships with Google and Flipkart. - Reuters

If you can't beat'em, join'em. Infosys teams with Anthropic to build agents. - Bloomberg

OpenAI, Google DeepMind, and Anthropic are using advanced mathematics to evaluate the reasoning capabilities of their AI models, shifting from pattern recognition to complex problem-solving as a primary metric for progress. - Financial Times

IMO leaderboard - imobench.github.io

Anthropic deployed autonomous AI agents to execute complex legal, marketing, and sales tasks, triggering market volatility and concerns over the automation of specialized white-collar professions. - The Washington Post

Anthropic researchers are mapping the internal neural structures of the Claude AI model using mechanistic interpretability to understand concept representation and improve safety. - The New Yorker

Evercore ISI and Bank of America warn that rising AI capital expenditures, including Amazon's projected $200 billion 2026 spend, could push hyperscalers into negative free cash flow as aggregate debt exceeds cash holdings. - Fortune

High electricity demand from AI infrastructure is increasing energy costs, contributing to inflationary pressures, reducing consumer discretionary spending, and slowing projected economic growth. - Financial Times

NVIDIA says Blackwell Ultra delivers 50x higher throughput per megawatt and 35x lower cost per token than Hopper, with Microsoft, CoreWeave, and Oracle deploying the platform for low-latency agentic AI and long-context reasoning. - NVIDIA Blog

Micron is executing a $200 billion U.S. expansion while currently meeting 50% to 66% of demand for key customers as artificial intelligence increases memory chip requirements. - The Wall Street Journal

Researchers identified side-channel vulnerabilities in LLMs, including ChatGPT and Claude, where efficiency techniques leak timing patterns that allow adversaries to infer conversation topics with 90% precision and recover personally identifiable information from encrypted traffic. - Schneier on Security

Face/Off, for OnlyFans: Web site unmasks names of cam girls, presents a serious privacy risk to sex workers, some who may not want stalkers, harassers, or employers to discover their profiles. The site’s creator claimed to 404 Media that millions of searches are done each month on the site. - 404 Media

Jeff Dean on the Latent.Space AI Engineer podcast - Latent Space

GitHub and Microsoft Research released a technical preview of Agentic Workflows, which integrates AI agents into GitHub Actions to automate repository management tasks like issue triaging and documentation updates through sandboxed execution. - The Register

Fork this: Gentoo abandons GitHub over pushing Copilot AI, attendant copyright concerns - The Register

Cohere released Tiny Aya, a family of 3.35-billion-parameter open-weight models supporting over 70 languages for offline use, trained on 64 H100 GPUs and including regional variants for Africa and Asia. - TechCrunch

Manus, now a Meta company, launched Manus Agents to integrate AI capabilities directly into messaging apps, beginning with Telegram, allowing users to perform tasks using voice, images, and files through Manus 1.6 Max or Lite models. - Manus

Yes, You Can Vibe-Code. Here’s How to Get Started. - The Wall Street Journal

Seventy-one percent of CIOs anticipate budget cuts if AI investments lack measurable ROI by 2026, while 85% expect their compensation to be tied to performance outcomes amid reports of vendor remorse and governance challenges. - The Register

Wedbush analyst Dan Ives characterized the recent 20% selloff in enterprise software stocks as a "disconnected trade," arguing that high switching costs protect companies like Salesforce and ServiceNow from immediate AI disruption. - Yahoo

Structurally justified, temporarily overstated? Large language models and AI agents are replicating 80% of vertical software capabilities, causing market cap losses for incumbents like FactSet as competition shifts toward proprietary data and regulatory defensibility. - X (formerly Twitter)

Founder Mode: Workday co-founder Aneel Bhusri replaced Carl Eschenbach as CEO on February 9 to lead AI-focused enterprise software growth, while the company confirmed that fourth-quarter results are expected to align with previous financial guidance. - Yahoo

There is truth to AI risks to e.g. Adobe

Supply chain risk, you are dependent on AI, at some point they may have pricing power, look at what you are doing and build their own solution like Amazon Basic Illustrator
UI risk, users start using product via prompts, AI can tell them Canva is better. Then you become a dumb pipe, people access through AI, easy to switch, no pricing power
also Adobe is already getting disrupted by cheap Canva and Figma, if you raise prices too much people vibe-code something for the 10% of the product they actually use. Or buy 1 seat that agents talk to, per-seat licensing can become problematic
In case of Adobe, there may be far fewer creatives, people do stuff with AI, it can destroy demand, it can make fewer people 10x as productive, you sell fewer seats/licenses

Bank of America CEO Brian Moynihan dismissed AI job loss concerns, citing workforce growth following the 1960s computer revolution and describing the technology as an augmentation of human capabilities that requires worker retraining. - Fortune

Amazon, Meta, and Salesforce are reducing corporate roles as AI automates professional tasks, with over 55,000 layoffs projected by late 2025 and increasing political focus on taxing robot labor and universal basic income. - UnHerd

Professional experts in medicine, law, and finance are training AI models through reinforcement learning, fueling a $17 billion service industry led by firms like Mercor. - CNN