Category: LLM (text)

Two themes: (1) Rapid, powerful AI growth—new models, widespread workplace and media use, drug repurposing, and surging chip demand—are reshaping industries. (2) Significant risks—cybersecurity, learning loss, job disruption, legal and regulatory fights—require urgent mitigation via policy, education, and defense.

• Fortune: Exclusive: Anthropic 'Mythos' AI model representing 'step change' in power revealed in data leak (Mar. 26, 2026)
  Anthropic is testing a more capable AI, called Claude Mythos or Capybara, after a leak exposed an unpublished draft, nearly 3,000 assets, and plans. The company warns it poses major cybersecurity risks, and is giving limited early access to defenders.
• OpenAI: Saying Goodbye to Sora (Mar. 24, 2026)
• NY Times: A.I. Saved His Life by Discovering New Uses for Old Drugs (Mar. 20, 2025)
  Joseph Coates, dying from a rare blood disorder, recovered after an A.I. model found a lifesaving drug combination. Scientists now use machine learning to repurpose approved drugs, rapidly uncovering new treatments.
• UTS Centre for Social Justice and Inclusion: Artificial intelligence, cognitive offloading and implications for education (Mar. 23, 2026)
  AI can let students outsource core thinking, causing short-term gains but harming durable learning, metacognition, and equity. It recommends thoughtful pedagogy, metacognitive prompts, and teacher-facing AI to build knowledge, critical thinking, and fair outcomes.
• WSJ: An AI Upheaval Is Coming for Media. This Journalist Is Already All In. (Mar. 26, 2026)
  Nick Lichtenberg uses AI prompts to produce many fast, short Fortune stories, driving substantial web traffic. His AI-first workflow frees time for longer features, reflects broader newsroom adoption, and sparks debate over speed versus deep reporting.
• WSJ: ChatGPT, Claude and Gemini Entered the WSJ Bracket Pool. One Might Actually Win. (Mar. 24, 2026)
  Three top AI models, Claude, Gemini, and ChatGPT, were entered into a Wall Street Journal March Madness pool, and after one weekend all three can still win.
• WSJ: Meta Names New Leader of Push to Adopt AI Throughout Its Workforce (Mar. 24, 2026)
  Meta CTO Andrew Bosworth will lead “AI For Work”, driving AI across its 78,000 employees to speed work, cut layers, and reshape jobs. He’ll launch an applied-AI group for LLMs, unify agent infrastructure, and push tools like MyClaw.
• The Chronicle of Higher Education: Teaching: How to use AI to mitigate cognitive offloading (Mar. 26, 2026)
  Students overwhelming use AI for study, but overreliance can cause a performance paradox, harming learning when AI is removed.
• NY Times Opinion: Bracing for the A.I. Economy to Come (Mar. 22, 2026)
  Readers warn A.I. will cause fast, broad job loss in fields like management, medicine, and law, burdened by student debt, and urge retraining, humanities, or shorter workweeks. Another letter says Trump aides are too small to fill big shoes.
• CNN: Judge blocks Pentagon’s effort to ‘punish’ Anthropic by labeling it a supply chain risk (Mar. 26, 2026)
  A federal judge barred the Pentagon from labeling Anthropic a supply chain risk, ruling the move violated First Amendment and due process rights, and punished the company for its public stance. The Pentagon will appeal.
• WSJ: The SpaceX IPO Will Be Just as Unconventional as Elon Musk Himself (Mar. 26, 2026)
  Musk plans a mid‑June SpaceX IPO, inviting investors to facilities and launches, and may give preferential access to Tesla and other Musk backers.
• NY Times: The Social Media Addiction Trials: What to Know (Jan. 26, 2026)
  Landmark trials claim Meta, TikTok, Snap, and YouTube designed addictive products that harmed young users, and juries have ordered damages.
• WSJ: How the AI Boom Has Transformed the Chip Industry Into a Market Monster (Mar. 25, 2026)
  AI demand put seven chipmakers among the 25 most valuable firms, with Nvidia topping the list at ~$4.3 trillion. Global chip revenue will rise 33% to over $1 trillion.

Two themes: accelerating AI/compute—cheaper, faster models and inference (GPT‑5.4 mini/nano, local Qwen streaming, Nvidia/Arm hardware, Mistral Forge) reshaping markets and geopolitics; and disruptive tech/biomed advances—cryopreservation, GLP‑1 drug boom, rapid EV charging—creating access, cost, and inflation risks.

• Simon Willison: GPT-5.4 mini and GPT-5.4 nano, which can describe 76,000 photos for $52 (Mar. 17, 2026)
  OpenAI released GPT‑5.4 mini and nano; nano outperforms the prior mini at top reasoning, mini is twice as fast. Nano undercuts Gemini Flash‑Lite on price, cuts per-image cost below 0.1 cent, and would tag 76,000 photos for about $52.
• WSJ: Highlights From Nvidia CEO Jensen Huang’s Keynote Speech Today (Mar. 16, 2026)
  At GTC, Jensen Huang unveiled a flagship inference system combining 72 Vera Rubin servers, 256 Groq-built LPUs, and software from a recent Nvidia licensing deal. Nvidia says it generates 700 million tokens per second, about 350 times faster than Hopper.
• WSJ: Scientists Successfully Unfroze Part of a Mouse Brain—and It Still Worked (Mar. 26, 2026)
  Researchers froze mouse hippocampus slices with cryoprotectants, rapid cooling, and storage at very low temperatures, preventing ice damage. After thawing, neurons survived and exchanged electrical signals for days, though scaling to whole brains remains a major challenge.
• WSJ: Weight Loss Drug Frenzy: What’s Here and What’s Likely Coming Next (Mar. 26, 2026)
  A surge in GLP-1 weight-loss drugs, like semaglutide and tirzepatide, has driven demand, shortages, higher prices, and off-label use. More drugs, combos, oral forms, and extended-release versions are coming, but access, cost, and long-term safety remain unclear.
• Vi Saint: AI Coding is Gambling (Mar. 14, 2026)
  AI coding feels like gambling, delivering fast, tempting, and often misleading results. It speeds tasks, but it removes the creative problem-solving that makes coding satisfying.
• WSJ: Nvidia CEO Jensen Huang Says Company Is Restarting Production of AI Chips for Sale in China (Mar. 17, 2026)
  Nvidia has restarted manufacturing of its H200 AI processors for China, after winning licenses and receiving purchase orders. The move follows U.S. export curbs, a 25% sales-sharing deal, and could reopen a market worth billions.
• WSJ: Arm’s Timing Is Good, but Big Chip Move Now Has to Go Perfectly (Mar. 26, 2026)
  Arm unveiled its own AI server CPU, its stock surged, and investors priced in large success. Early deals with Meta and OpenAI back the push, but lower margins, partner risk, and an 80x forward valuation raise concerns.
• Dan Woods: Autoresearching Apple's "LLM in a Flash" to run Qwen 397B locally (Mar. 18, 2026)
  Qwen 3.5 397B was run locally on an M3 Max by streaming 2-bit requantized expert weights from SSD, achieving 5.7 tokens/s, 7.07 peak, and 5.5 GB resident memory, while keeping production-quality output.
• Mistral AI: Introducing Forge (Mar. 17, 2026)
  Forge, a system for enterprises to build frontier-grade AI models grounded in their proprietary knowledge.
• WSJ: China Has Five-Minute EV Charging. America Is Trying to Catch Up. (Mar. 17, 2026)
  U.S. fast-charging grew about 87% in 2025, with more plugs, higher power, and nearly 21,000 stations over 300 kW. 
• NY Times: War in Iran Will Push U.S. Inflation Above 4 Percent, O.E.C.D. Forecast Says (Mar. 26, 2026)
  An O.E.C.D. forecast says the Middle East war will lift U.S. inflation to about 4.2%, and G20 inflation to 4%, via higher oil, fertilizer, and shipping costs. Global growth remains near 2.9%, with major downside risks.

Two themes: practical iteration fuels AI’s benefits—users, tools, and fluency metrics show productivity gains when models are tested and integrated; and governance/security matter—exploits and Responsible Scaling updates underscore urgent need for safeguards and transparency.

• Anthropic: Anthropic Education Report: The AI Fluency Index (Feb. 23, 2026)
  Analysis of 9,830 Claude.ai conversations measures observable AI fluency behaviors, finding iteration and refinement strongly boost fluency. 
• NY Times Opinion: This Exercise Research Tells Us Nothing (Dec. 31, 1969)
  Exercise improves blood pressure, mental health, and longevity, but research on the “best” type or exact dose is often misleading, relying on flawed observational studies. Choose activities you enjoy, fit them into your life, and keep going past February.
• Paul Price: How We Hacked McKinsey's AI Platform — CodeWall.ai (Mar. 9, 2026)
  An autonomous agent exploited an unauthenticated SQL injection in McKinsey’s Lilli AI platform, gaining full read/write access to chats, files, accounts, prompts, and model configs.
• Anthropic: What 81,000 people want from AI (Mar. 18, 2026)
  Claude.ai users across 159 countries described their hopes, concerns, and experiences with AI. Most want AI to improve work, free time, and personal growth, and 81% said it already helped with productivity, learning, or emotional support.
• Craig Mod: Software Bonkers (Mar. 18, 2026)
  Claude Code built TaxBot2000, a local accounting tool that handles multiple currencies, imports CSVs, and automates categorization. 
• CanIRun.ai: CanIRun.ai — Can your machine run AI models?
  Find out which AI models your machine can actually run.
• Shaun Maguire: xAI will win (Mar. 23, 2026)
  People are underestimating Elon and xAI; his pattern of spotting tech bottlenecks, acting fast, and focusing on a few priorities has led to repeated wins. His AI bets span vision, robots, and language, aiming at AGI.
• papers.ssrn.com: Can LLMs Discover Novel Economic Theories? by Alejandro Lopez-Lira, Sina Seyfi, Yuehua Tang :: SSRN (Mar. 21, 2026)
  AutoTheory uses large language models, evolutionary search, and multiple expert personas to generate, test, and refine economic theories. 
• Anthropic: Responsible Scaling Policy Version 3.0 (Mar. 24, 2026)
  Anthropic released RSP v3, refining conditional safeguards, increasing transparency, and separating company commitments from industry recommendations.

AI capabilities are rapidly advancing—GPT‑5.4 and autonomous agents improve reasoning, coding, long‑task automation, and power services like Microsoft's Copilot Health. Simultaneously, governance and societal impact are contested: military/supply‑chain disputes, calls for regulation, and shrinking entry‑level job opportunities are major challenges.

• WSJ: Microsoft’s New AI Health Tool Can Read Your Medical Records and Give Advice (Mar. 12, 2026)
  Microsoft launched Copilot Health, an AI concierge that, with consent, uses encrypted medical records, labs, and wearable data to give personalized medical advice.
• Zvi Mowshowitz: GPT-5.4 Is A Substantial Upgrade (Mar. 11, 2026)
  Benchmarks no longer separate top models well, so hands-on testing and varied reports matter. GPT‑5.4 notably improves coding, knowledge work, context, and steering.
• OpenAI: GPT-5.4 Thinking System Card (Mar. 5, 2026)
  GPT-5.4 Thinking is the latest reasoning model, with safety mitigations similar to earlier releases, and is the first general-purpose model to include high-capability cybersecurity protections.
• OpenAI: Introducing GPT-5.4 (Mar. 3, 2026)
  GPT-5.4, and GPT-5.4 Pro, is a faster, more capable model for professional work, with improved reasoning, coding, and native computer-use. 
• Noah Smith: If AI is a weapon, why don't we regulate it like one? (Mar. 5, 2026)
  Anthropic clashed with the U.S. Department of War over limits on military uses of its AI, prompting a supply-chain risk threat, a switch to OpenAI, then resumed talks.
• Ethan Mollick: The Shape of the Thing (Mar. 12, 2026)
  AI has shifted from co‑intelligence to managing agents that autonomously do complex work, producing exponential gains in images, video, and long‑task benchmarks. That enables software factories, risks sudden job, market, and policy disruption.
• Planned Obsolescence: I underestimated AI capabilities (again) (Mar. 5, 2026)
  METR shows AI agents, like Opus 4.6, outperforming forecasts on multi-hour software tasks. As tasks scale to weeks, agents can be parallelized to tackle big projects, though some work still needs holistic context.
• Dallas Morning News: AI is making it harder to land entry‑level jobs, Dallas Fed report finds (Mar. 5, 2026)
  Analysis finds AI reduces entry-level jobs, squeezing young workers, while raising wages for experienced employees whose tacit skills resist automation.
• Anthropic: Where things stand with the Department of War (Mar. 5, 2026)
  Anthropic was labeled a supply‑chain risk by the Department of War, will challenge the narrow designation in court, says it affects only uses tied to War contracts, and will keep supplying models to the Department at nominal cost during transition.
• TechCrunch: Anthropic CEO Dario Amodei calls OpenAI's messaging around military deal 'straight up lies,' report says (Mar. 4, 2026)
  Anthropic CEO Dario Amodei slammed OpenAI’s DoD deal as “safety theater,” accusing Sam Altman of misleading staff, investors, and the public, after Anthropic refused unrestricted access.