Top Stories
Anthropic Abandons Core Safety Pledge Amid Pentagon Standoff
In a significant shift, Anthropic has removed a key provision from its AI safety policy that would have required the company to pause training more powerful models if their capabilities outpaced its ability to control them. The change comes as Defense Secretary Pete Hegseth’s Friday ultimatum looms.
The new policy replaces Anthropic’s previous self-imposed guardrails with a “nonbinding safety framework” that the company says “can and will change.” While Anthropic insists the policy revision is unrelated to Pentagon negotiations, the timing is striking.
Two red lines remain: Anthropic has stated it won’t allow Claude to be used for AI-controlled weapons that fire without human involvement or for mass domestic surveillance of American citizens. Everything else appears negotiable.
What this means: The company that positioned itself as AI’s responsible alternative to OpenAI is now adopting the same flexible approach to safety commitments. Whether this represents pragmatic adaptation or mission drift depends on your perspective - but it’s a notable departure from the “responsible scaling policy” that was central to Anthropic’s identity.
Samsung Launches Galaxy S26 with “Agentic AI” at Unpacked
Samsung unveiled the Galaxy S26 series yesterday in San Francisco, positioning it as “the most intuitive Galaxy AI phone yet.” The headline feature: agentic AI that acts autonomously on your behalf.
The S26 can now handle multi-step tasks like scheduling, summarizing communications, and conducting research directly from conversations. A new “Now Nudge” feature reduces app-switching by surfacing contextual information when relevant - for example, checking your calendar when a friend asks about evening plans and alerting you to conflicts.
Circle to Search now supports multi-element searches, and Samsung has expanded its AI ecosystem to include Bixby, Gemini, and Perplexity. Google’s Scam Detection is now available directly in the Samsung Phone app.
The Galaxy S26 Ultra ships March 11 with a 200MP camera, Snapdragon 8 Elite Gen 5, and 12GB RAM.
Sources: Samsung Newsroom, Google Blog, Neowin
Nvidia Beats Earnings, Vera Rubin Gets First Public Demo
Nvidia reported Q4 FY26 earnings after the bell yesterday, beating estimates by $0.10 per share with revenue topping expectations. The company’s sales forecast confirms that AI infrastructure spending shows no signs of slowing.
Separately, CNBC got an exclusive first look at Vera Rubin, Nvidia’s next-generation AI system shipping in the second half of 2026. The system combines one Vera CPU with two Rubin GPUs in a single chip containing 17,000 components.
Key specs: Vera Rubin uses twice the power of Blackwell but delivers 10x better performance per watt, with 10x reduction in inference token cost and 4x fewer GPUs needed to train mixture-of-experts models. It’s also Nvidia’s first 100% liquid-cooled system, which CEO Jensen Huang says will help data centers consume “much less water” than traditional cooling.
AWS, Google Cloud, Microsoft, and OCI will be among the first to deploy Vera Rubin-based instances.
Sources: Investing.com, Bloomberg, CNBC
Local AI & Open Source
Ollama 0.17 Brings Major Performance Gains
Ollama released version 0.17 on February 22 with up to 40% faster prompt processing on certain hardware. The update introduces a new inference engine, improved tensor parallelism for multi-GPU setups, and better memory management that reduces out-of-memory crashes.
This is particularly relevant for users running 70B+ parameter models across multiple GPUs.
Source: Web And IT News
llama.cpp Finds Permanent Home at Hugging Face
ggml.ai has joined Hugging Face to ensure long-term open-source backing for llama.cpp. The move comes after Ollama departed from llama.cpp in May 2025 to build its own inference engine. Hugging Face’s backing should help keep the project community-driven and well-maintained.
Source: WinBuzzer
Quick Hits
-
UK funds 60 AI alignment projects: The UK AI Safety Institute announced £27 million in grants across 60 projects, backed by OpenAI, Microsoft, Anthropic, and AWS. Over 800 applications from 466 institutions in 42 countries applied. ResultSense
-
Google warns AI wrapper startups: Google’s Darren Mowry told TechCrunch that companies built around LLM wrappers and AI aggregators have their “check engine light” on - suggesting their business models may not survive as foundation model providers expand capabilities. PYMNTS
-
EU releases AI Act transparency draft: The European Commission published the first draft Code of Practice on transparency for AI-generated content under the AI Act, developed with industry, academia, and civil society input. Fladgate
-
New chip shortage looming: Automakers are racing to prepare for a new microchip shortage as AI data centers consume increasing amounts of DRAM, driving up prices. Detroit News
Worth Watching
The Friday Deadline: All eyes are on Anthropic as Hegseth’s ultimatum expires tomorrow at 5:01pm. Will Anthropic capitulate on its remaining red lines around autonomous weapons and mass surveillance? Will the Defense Production Act actually be invoked? The outcome will set precedent for how AI companies negotiate with government power.
Apple’s March 4 Event: Apple has invited media to “a special Apple Experience” in New York, London, and Shanghai. The AI-enhanced Siri that’s been delayed since 2025 may finally get a concrete timeline - though Bloomberg reports it may slip again to iOS 26.5 or iOS 27. New MacBook Air with M5 chip expected.
Vera Rubin Deployment Timeline: With Nvidia’s next-gen platform entering full production, watch for cloud provider announcements about availability windows. The 10x efficiency gains over Blackwell could reshape AI infrastructure economics by late 2026.