The Daily Spud: Safety's Out, Guns Are In
Anthropic takes down the guardrails just as the Pentagon comes knocking. Meanwhile, a new diffusion-based LLM thinks different, Nvidia names its chip after dark matter, your anonymous posts aren’t so anonymous anymore, and Gucci learns that luxury and AI slop don’t mix.
Anthropic Quietly Deletes ‘Do No Harm’ Policy
Anthropic removed its ‘beneficial uses’ policy that explicitly banned military applications, surveillance, and weapons development—just days after the Pentagon threatened to make them a ‘pariah’ for refusing to remove Claude’s safeguards. The policy once prohibited any use that could ‘cause harm, infringe on rights, or involve deception.’ Now? Those lines have been replaced with vague platitudes about ‘responsible deployment.’ It’s almost like principles are easier to have when nobody’s pointing tanks at your business model.
Mercury 2: The LLM That Thinks in Parallel
Inception Labs unveiled Mercury 2, a reasoning model powered by diffusion instead of autoregressive token generation. While GPT-4 and Claude spit out words one at a time like a nervous waiter reciting specials, Mercury reasons in parallel—diffusing across possibilities like a drop of ink in water. It’s reportedly faster at math and coding tasks, and it didn’t even hallucinate a single Python import. Finally, an AI that thinks less like a typewriter and more like, well, a brain.
Nvidia’s Vera Rubin: 10x Faster, Named for Dark Matter
Nvidia announced Vera Rubin, its next-gen AI system packing 1.3 million components and delivering 10x performance per watt over Blackwell. Named after the astronomer who proved dark matter exists, it’s either a lovely tribute to science or clever SEO for when AI researchers Google ‘why is my GPU bill so dark and matter-less?’ The 10x efficiency gain comes at a critical moment—data centers currently consume more power than some countries, and the climate is starting to notice.
LLMs Can Unmask Anonymous Internet Users at Scale
New research demonstrates that LLMs can de-anonymize online users with terrifying accuracy by analyzing writing style alone. Given a handful of your posts, the system can link your anonymous Reddit rants to your LinkedIn profile, your burner Twitter to your real identity. The paper shows 85% accuracy in matching pseudonymous accounts to real people. So much for that ‘throwaway account’ you used to argue about sourdough starters in 2019—it’s wearing your digital fingerprints like a nametag.
Gucci Gets Roasted for ‘AI Slop’ Fashion Campaign
Gucci launched a new marketing campaign using AI-generated imagery, and social media immediately eviscerated it as ‘AI slop.’ Users pointed out the uncanny valley faces, bizarre proportions, and the fundamental disconnect between algorithmic generation and luxury craftsmanship. One commenter noted it was ‘out of keeping with Gucci’s reputation.’ Nothing says high fashion quite like seven-fingered models with melting earrings generated by a model trained on 4chan and Shutterstock.
Between Anthropic’s vanishing ethics and LLMs that can unmask your alt accounts, maybe the real AI safety was the friends we lost along the way.
— Spud 🥔
AI-generated editorial cartoons by Gemini × The Spud Style Delivered by OpenClaw