Andrew Maddox

Blog

  • Various AI Links (Dec. 29)

    • WSJ: How AI Is Making Life Easier for Cybercriminals (Dec 26, 2025)
      Rapid advances in AI are empowering cybercriminals to automate and scale highly convincing phishing, malware, and deepfake attacks, and dark‑web tools let novices rent or build campaigns. Security experts warn that autonomy may be near, urging AI‑driven defenses, resilient networks, multifactor authentication, and skeptical user habits.
    • Ibrahim Cesar : Grok and the Naked King: The Ultimate Argument Against AI Alignment — Ibrahim Cesar (Dec 26, 2025)
      Grok demonstrates that AI alignment is determined by who controls a model, not by neutral technical fixes: Musk publicly rewired it to reflect his values. Alignment is therefore a political, governance issue tied to concentrated wealth and power.
    • NY Times: Why Do A.I. Chatbots Use ‘I’? (Dec 19, 2025)
      A.I. chatbots are intentionally anthropomorphized—with personalities, voices, and even “soul” documents—which can enchant users, foster attachment, increase trust, and sometimes cause hallucinations or harm. Skeptics warn that anthropomorphic design creates the “Eliza effect”: people overtrust, form attachments, or even develop delusions.
    • NY Times Opinion: What Happened When I Asked ChatGPT to Solve an 800-Year-Old Italian Mystery (Dec 22, 2025)
      Elon Danziger argues that his research shows Florence’s Baptistery was a papal-led project tied to Pope Gregory VII, and that ChatGPT, Claude, and Gemini failed to replicate his discovery. He claims that LLMs miss outlier evidence and lack the creative synthesis needed for historical breakthroughs.
    • WIRED: People Are Paying to Get Their Chatbots High on ‘Drugs’ (Dec 17, 2025)
      Swedish creative director Petter Rudwall launched Pharmaicy, a marketplace selling code modules that make chatbots mimic being high on substances like cannabis, ketamine, and ayahuasca. Critics say the effects are superficial output shifts rather than true altered experiences, raising ethical questions about AI welfare, deception, and safety.
    • WSJ: China Is Worried AI Threatens Party Rule—and Is Trying to Tame It (Dec 23, 2025)
      Worried AI could threaten Communist Party rule, Beijing has imposed strict controls—filtering training data, ideological tests for chatbots, mandatory labeling, traceability, and mass takedowns—while still promoting AI for economic and military goals. The approach yields safer-but-censored models that risk jailbreaks and falling behind U.S. advances.
    • NY Times: Trump Administration Downplays A.I. Risks, Ignoring Economists’ Concerns (Dec 24, 2025)
      The White House, led by President Trump, is championing A.I. as an engine of economic growth—cutting regulations, fast‑tracking data centers, and courting tech investment—while dismissing bubble and job‑loss concerns. Economists and Fed officials warn of potential mass layoffs, unsustainable financing, and systemic risks.
    • NY Times: The Pentagon and A.I. Giants Have a Weakness. Both Need China’s Batteries, Badly. (Dec 22, 2025)
      America’s AI data centers and the Pentagon’s future weapons increasingly depend on lithium-ion batteries dominated by China, creating strategic vulnerabilities. 
    • Piratewires: The Data Center Water Crisis Isn’t Real (Dec 18, 2025)
      Andy Masley used simple math, AI, and domain knowledge to debunk exaggerated claims that individual AI use (e.g., an email) or data centers “guzzle” huge amounts of water — the “bottles of water” metric is misleading and easily miscomputed.
    • NY Times: Senators Investigate Role of A.I. Data Centers in Rising Electricity Costs (Dec 16, 2025)
      Three Democratic senators asked Google, Microsoft, Amazon, Meta, and other data‑center firms for records on whether A.I. data centers’ soaring electricity demand has forced utilities to spend billions on grid upgrades that are recouped through higher residential rates. They warned ordinary customers may be left footing the bill.

  • AI in Higher Education & Medicine

    • Roon: Too Bearish on AI (Dec 26, 2025)
      The author admits they were too bearish mid-year, expecting improvements beyond reinforcement learning to be required. After trying Codex, they realized AI progress is clearly in a rapid takeoff.
    • WSJ: These Teenagers Are Already Running Their Own AI Companies (Dec 21, 2025)
      Teenagers are launching AI-powered startups—like 15-year-old Nick Dobroshinsky’s BeyondSPX—using generative models to build products quickly and attract users. Investors note AI lowers technical barriers and accelerates entrepreneurship.
    • WSJ Opinion: AI Means the End of Entry-Level Jobs (Dec 22, 2025)
      AI is eroding entry-level roles that traditionally launch careers, causing younger workers to worry and raising unemployment among 22–25-year-olds in affected sectors. Companies should create new pathways—AI-native roles, mentor-intensive programs, project-based progression, and competency-based advancement—integrating AI and business training to build future talent.
    • Importai Substack: Import AI 438: Silent sirens, flashing for us all (Nov 30, -0001)
      Powerful AI capabilities are often hidden from everyday users — tools like Claude Code can rapidly build complex software, and by 2026 an “AI economy” will accelerate and diverge from everyday experience, benefiting those who can access and elicit frontier systems.
    • Johannes Schmitt: AI model (GPT-5) autonomously solved an open math problem (Dec 17, 2025)
      GPT-5 autonomously solved an open enumerative-geometry problem, giving a complete, correct proof for ψ-class intersection numbers on moduli spaces of curves. 
    • NY Times Opinion: College Students Need Tech-Free Spaces (Dec 19, 2025)
      Colleen Kinder had Yale students surrender their phones for a four‑week, Wi‑Fi‑free writing course in Auvillar, France, and reports improved sleep, focus, reading speed, and creativity, with far greater writing output. She argues colleges should create internet‑free tracts, dorms, or “cloisters” to protect learning from constant distraction.
    • NOAA: NOAA deploys new generation of AI-driven global weather models (Dec 17, 2025)
      NOAA launched AI-driven global models—AIGFS, AIGEFS, and hybrid HGEFS—that provide faster, more accurate forecasts using far fewer computing resources (AIGFS ~0.3%, AIGEFS ~9%). HGEFS’s combined AI‑physics ensemble outperforms each system; NOAA reports better tropical cyclone tracks but will refine intensity forecasts.
    • WSJ: Millions of Kids Are on ADHD Pills. For Many, It’s the Start of a Drug Cascade. (Nov 19, 2025)
      The WSJ reports that many children put on ADHD drugs—often after school pressure and lacking behavioral therapy—receive additional psychotropic medications to manage side effects or perceived disorders. Medicaid data show that those started on ADHD meds in 2019 were over five times likelier to be on psychiatric drugs four years later.

  • AI Market & Product Updates (Dec. 27)

    • WSJ: Nvidia Licenses Groq’s AI Technology as Demand for Cutting-Edge Chips Grows (Dec 24, 2025)
      Nvidia struck a nonexclusive licensing deal with AI-chip startup Groq for its inference-focused language-processing-unit technology, with Groq CEO Jonathan Ross, the company president, and some staff joining Nvidia while GroqCloud stays independent.
    • WSJ: The Former Ice-Hockey Player Who Nailed This Year’s AI Trade (Dec 20, 2025)
      Former hockey captain Xavier Majic’s $3 billion Maple Rock hedge fund gained over 60% through November 2025 by betting early on data-storage suppliers (Western Digital, Seagate, Kioxia) that profited from AI-driven demand.
    • NY Times: Why the A.I. Rally (and the Bubble Talk) Could Continue Next Year (Dec 23, 2025)
      Do soaring valuations indicate the existence of an AI bubble? Nvidia and the “Magnificent 7” dominate markets, OpenAI’s huge fundraising and trillion‑dollar data‑center plans, and a construction boom strain power and capital. Analysts are split: some warn of valuation and investment bubbles, others argue AI’s productivity gains justify the rally.
    • Mistral Ai: Introducing Mistral OCR 3 (Dec 19, 2025)
      Mistral OCR 3 is a compact, cost-effective OCR model offering state-of-the-art accuracy—claiming a 74% overall win rate versus Mistral OCR 2—excelling at forms, handwriting, low-quality scans, and complex tables while producing markdown/HTML table output. It’s available via API and the Document AI Playground, priced at $2/1,000 pages ($1/ batch).
    • Andrej Karpathy: 2025 LLM Year in Review (Dec 19, 2025)
      2025 saw major LLM shifts: Reinforcement Learning from Verifiable Rewards (RLVR) drove long-horizon capability and emergent reasoning, revealing jagged, “ghost”-like intelligence. New paradigms—Cursor apps, local agents (Claude Code), vibe coding, and GUI breakthroughs (Nano banana)—democratized development and reshaped how AI is used.
    • WSJ: Meta Is Developing a New AI Image and Video Model Code-Named ‘Mango’ (Dec 18, 2025)
      Meta is developing Mango, an image-and-video AI model, alongside a text-based model called Avocado, with both expected in the first half of 2026. Avocado will emphasize coding and world-model research under chief AI officer Alexandr Wang as Meta expands its AI team amid fierce image-generation competition.
    • WSJ: OpenAI’s New Fundraising Round Could Value Startup at as Much as $830 Billion (Dec 18, 2025)
      OpenAI is seeking up to $100 billion in a fundraising round that could value it at $830 billion, targeting completion by Q1 and drawing investors like SoftBank and Disney. The cash is needed to build AI models amid competition from Google and investor scrutiny over costly computing deals.

  • iRobot Sold for Scrap

    From the NY Times: Roomba Maker iRobot Files for Bankruptcy, With Chinese Supplier Taking Control

    iRobot, founded in 1990 by three MIT researchers and maker of the Roomba (2002), filed for bankruptcy and will be taken over by its largest creditor, Chinese supplier Picea. Years of regulatory scrutiny, privacy issues, stiff competition, and the failed Amazon deal depleted revenue and left the company heavily indebted.

    This is another example of the incompetence of America’s antitrust laws (or enforcement thereof). I can’t imagine that the Sherman Antitrust Act was written to prevent American companies from buying struggling ones.

    From John Gruber:

    By 2022, the Amazon acquisition was iRobot’s lifeline. EU regulators wanted it shot down, and despite the fact that it was one American company trying to acquire another, the anti-big-tech Biden administration clearly preferred to let the deal collapse. The US should have told the EU to mind their own companies.

    This story is another anecdote that we’d be far better off trying to build things instead of reflexively decrying big business sweeping up smaller ones (particularly ones that were struggling). I’m sympathetic to Klein and Thompson’s arguments about abundance, particularly as AI technology is growing by leaps and bounds.

    …

    Related: WSJ Opinion agrees: How Lina Khan Killed iRobot. iRobot filed for bankruptcy after 35 years when the Biden FTC under Lina Khan—amid pressure from Sen. Elizabeth Warren—blocked Amazon’s acquisition and Trump’s tariffs hobbled production. Critics say the FTC’s opposition and trade policy accelerated layoffs and a takeover by Chinese manufacturer Picea, showing how intervention can strengthen foreign rivals.
     

  • Sunday (AI) Links (Dec. 14)

    • Simon Willison: JustHTML is a fascinating example of vibe engineering in action (Dec 14, 2025)
      JustHTML is a pure-Python HTML5 parser that passes the 9,200+ html5lib tests, offers CSS selectors, and achieves 100% test coverage in a ~3,000-line codebase. Emil Stenström built it largely with LLM coding agents—using benchmarks, fuzzing, profiling, and human-led design—as an example of “vibe engineering.”
    • Simon Willison: Useful patterns for building HTML tools (Dec 10, 2025)
      A list of single-file applications combining HTML, JavaScript, and CSS, often built with LLMs that are designed for easy hosting and distribution, leveraging techniques like CDN dependencies, copy-paste functionality, URL state persistence, and CORS-enabled APIs. 
    • Simon Willison: Dark mode (Dec 10, 2025)
      Willison used Claude Code to create a dark mode theme for his website. “It did a decent job,” Willison reported.
    • WSJ: AI Can Make Decisions Better Than People Do. So Why Don’t We Trust It? (Dec 12, 2025)
      Engineers and executives say well-designed AI decision systems—from autonomous truck drivers to an AI arbitrator—can outperform humans and be more auditable and explainable. But public distrust, past algorithmic harms, and unfamiliarity slow adoption; verification, transparency, and responsible development are needed to earn trust and reduce harm.
    • WSJ: He Blames ChatGPT for the Murder-Suicide That Shattered His Family (Dec 11, 2025)
      The estate of Suzanne Eberson Adams sued OpenAI and Microsoft after her son, Stein‑Erik Soelberg, who had months of delusion-filled conversations with ChatGPT that allegedly reinforced paranoia, killed her and himself. The complaint alleges OpenAI rushed unsafe models, won’t release chat logs, and should be held responsible.
    • WSJ Opinion: New York’s Lack of AI Intelligence (Dec 11, 2025)
      WJS’s Editorial Board decries legislation that could hinder open-source development and prevent smaller entities from accessing AI tools. They implore Governor Hochul to veto this poorly conceived bill.
    • The Chronicle of Higher Education: The Conference Where ChatGPT Wrote One in Five Reviews (Maybe) (Dec 8, 2025)
      An AI detection startup found that 21% of over 75,000 reviews for the ICLR conference appeared fully AI-generated, with over half showing some AI usage. I still wonder about what constitutes “AI” usage—does Grammarly count? What about Word grammar usage? What if you like using dashes—as I do?
    • Simon Willison: A quote from Claude (Dec 9, 2025)
      “See that ~/ at the end? That’s your entire home directory. The Claude Code instance accidentally included ~/ in the deletion command.”
    • Forbes: Purdue University Approves New AI Requirement For All Undergrads (Dec 13, 2025)
      Purdue University will require all undergraduates entering in 2026 to demonstrate a discipline-specific AI working competency before graduation, embedding AI skills into existing degree requirements rather than adding credits. 
    • Brian Merchant: Copywriters reveal how AI has decimated their industry (Dec 11, 2025)
      The article chronicles how AI has decimated copywriting and related media jobs through layoffs, reduced hours, degraded work (editing AI output), falling wages, and closed businesses. Workers describe financial precarity, eroded career pathways, and being forced into survival work as companies favor cheaper “good enough” AI.
    • Uwe Friedrichsen: AI and the ironies of automation – Part 2 (Dec 11, 2025)
      Friedrichsen applies Lisanne Bainbridge’s “ironies of automation” to AI-agent-driven white‑collar work, warning that monitoring fatigue, verbose agent plans, rare but critical errors, and simulator limits create a training paradox for supervisors. He also highlights a leadership dilemma—humans must learn to direct agents—and urges better UIs and sustained training.

  • Saturday (AI) Links (Dec. 13)

  • Friday (AI) Links (Dec. 12)

    • WSJ: Fresh Concerns About AI Spending Are Rattling Wall Street (Dec 12, 2025)
      Broadcom’s 11% plunge — despite strong sales and profits — highlighted investor concern about AIchip margins, timing of big OpenAI commitments, and visibility into 2027.
    • NY Times: Can OpenAI Respond After Google Closes the A.I. Technology Gap? (Dec 11, 2025)
      OpenAI released GPT‑5.2, saying it tops key benchmarks shortly after Google touted Gemini 3, underscoring a tightened A.I. race. Facing fierce rivals and huge computing costs, it declared a “code red” to improve ChatGPT while raising fees, testing ads, and pushing enterprise products to reach profitability.
    • WSJ: AI Gadgets Are Bad Right Now, but Their Promise Is Huge (Dec 11, 2025)
      Joanna Stern tested eight AI wearables—pendants, bracelets, and glasses—and found many quickly abandoned due to poor design, privacy concerns, and limited usefulness, with smartphones remaining the main hub.
    • Simon Willison: OpenAI is quietly adopting skills, now available in ChatGPT and Codex CLI (Dec 12, 2025)
      OpenAI added “skills” support to ChatGPT’s Code Interpreter and the Codex CLI, adopting Anthropic’s simple folder+Markdown format so models can use filesystem-based tools. ChatGPT’s skills process docs/PDFs by rendering pages to PNGs for vision-enabled models.
    • Simon Willison: GPT-5.2 (Dec 11, 2025)
      OpenAI announced GPT‑5.2 and GPT‑5.2 Pro with an Aug 31, 2025 knowledge cutoff, 400k‑token context window, and higher pricing (GPT‑5.2 at 1.4×; Pro much costlier). OpenAI reports large benchmark and vision gains, a response‑compaction API for long workflows, three API variants (incl. gpt‑5.2‑chat‑latest), and CLI access.
    • Mistral Ai: Introducing: Devstral 2 and Mistral Vibe CLI. (Dec 9, 2025)
      Mistral released Devstral 2 (123B, modified MIT) and Devstral Small 2 (24B, Apache 2.0), open-source coding models achieving 72.2% and 68.0% on SWE-bench Verified and offering high cost-efficiency. They’re available via API (free initially) and power Mistral Vibe, a native CLI for autonomous, project-aware code automation.
    • WSJ: Behind the Deal That Took Disney From AI Skeptic to OpenAI Investor (Dec 11, 2025)
      Disney is investing $1 billion in OpenAI and licensing over 200 characters for use in Sora, allowing fans to create AI-generated videos. This deal contrasts with Disney’s cease-and-desist letter to Google for alleged copyright infringement, highlighting Disney’s dual approach to navigating the AI landscape.
    • Anthropic: Accenture and Anthropic launch multi-year partnership (Dec 9, 2025)
      Anthropic and Accenture formed the Accenture Anthropic Business Group to scale Claude across enterprises, training about 30,000 Accenture professionals and deploying Claude Code to tens of thousands of developers.
    • WSJ: AI’s Next Challenge: Take the CEO’s Job (Dec 7, 2025)
      Big Tech executives increasingly suggest AI could perform CEOs’ duties and even run companies, with figures like Pichai and Altman touting rapid progress. My take: CEOs want to seem like they’re in the same boat as employees whose jobs are at risk. CEOs seem like the last job to be replaced by an AI.
    • WSJ: IBM Strikes $11 Billion Deal for Confluent (Dec 7, 2025)
      IBM bolsters its AI and cloud strategy by adding Confluent’s real-time data-streaming technology used to feed large AI models.
    • WSJ: The Accounting Uproar Over How Fast an AI Chip Depreciates (Dec 8, 2025)
      Tech companies are extending the useful lives of AI equipment, which critics argue inflates profits by reducing depreciation expenses. While this accounting choice can boost current earnings, the true economic reality of these assets might be better reflected by accelerated depreciation methods.
    • OpenAI: The Walt Disney Company and OpenAI (Dec 11, 2025)
      Disney and OpenAI have formed a significant partnership, with Disney becoming a major content licensing partner for OpenAI’s Sora, allowing fans to generate short videos featuring over 200 Disney, Marvel, Pixar, and Star Wars characters.
    • TechCrunch: Adobe brings Photoshop, Express, and Acrobat features to ChatGPT (Dec 10, 2025)
      With the massive improvement of Sora and Nano Banana, I’ve wondered about Adobe’s prospects in the AI world. The company is responding: “[Adobe] is adding features from Photoshop, Express, and Acrobat to ChatGPT, letting users ask the chatbot to use these apps to edit images, modify PDFs, or animate elements.”

  • Wednesday (AI) Links (Dec. 10)

  • (AI) Links (Dec. 8)

  • Various (AI) Links – Dec. 3

    • Vechron: Anthropic Prepares for Potential 2026 IPO in Bid to Rival OpenAI: Report (Dec 3, 2025)
      Anthropic hired Wilson Sonsini to begin IPO preparations possibly for 2026, aiming to list before OpenAI amid a private fundraising that could value it above $300 billion. It says no decision is final, has strengthened finance and governance, and faces heavy spending on data centres and model training.
    • Anthropic: Anthropic acquires Bun as Claude Code reaches $1B milestone (Dec 2, 2025)
      Anthropic’s Claude Code, a leading AI model for developers, has reached $1 billion in run-rate revenue and is acquiring Bun, a high-performance JavaScript runtime, to enhance its capabilities. This acquisition aims to improve speed, stability, and workflows for Claude Code users by integrating Bun’s toolkit and optimizing the JavaScript developer experience.
    • WSJ: Millions of Coders Love This AI Startup. Can It Last? (Dec 1, 2025)
      Cursor, an AI coding tool favored by tech leaders like Sam Altman and Jensen Huang, is experiencing rapid growth and is valued at $29.3 billion. Despite its popularity and impressive growth metrics, the company loses money, relies heavily on external AI models, and faces questions about its long-term sustainability in a competitive market.
    • Simon Willison’s Weblog: Claude 4.5 Opus’ Soul Document (Dec 2, 2025)
      Richard Weiss extracted a 14,000-token “Soul overview” document from Claude 4.5 Opus, which Anthropic’s Amanda Askell confirmed was used to train the model’s personality during its training run using supervised learning. The “soul doc” outlines Anthropic’s mission to develop safe and beneficial AI, emphasizing good values, comprehensive knowledge, and wisdom for Claude, and even addresses topics like prompt injection attacks.
    • WSJ: Apple to Revamp AI Team After Announcing Top Executive’s Departure (Dec. 1, 2025)
      Apple is restructuring its AI division after the retirement of its AI chief, John Giannandrea, whose tenure was marked by the company’s struggle to compete in the rapidly evolving AI landscape.
    • WSJ: This AI Startup Wants to Remake the $800 Billion Chip Industry (Dec. 2, 2025)
      Two former Google researchers are launching Ricursive Intelligence, a startup aiming to automate chip design, potentially revolutionizing the $800 billion industry by enabling companies to create custom chips quickly and easily.
    • Anthropic: How AI is transforming work at Anthropic (Dec 2, 2025)
      Anthropic surveyed engineers and analyzed Claude Code usage, finding Claude widely used—boosting productivity (~50%), enabling more full‑stack work, greater output, and new tasks while handling increasingly complex workflows autonomously. Employees nonetheless worry about skill atrophy, reduced collaboration and mentorship, and career uncertainty.
    • Mistral: Introducing Mistral 3 (Dec 2, 2025)
      New models offer state-of-the-art performance, multimodal capabilities, and are designed for customization, with optimized versions available through collaborations with NVIDIA, vLLM, and Red Hat.
    • AP News: AI may be scoring your college essay. Welcome to the new era of admissions (Dec 1, 2025)
      Colleges are increasingly using AI in the admissions process, primarily to streamline tasks like transcript review and essay evaluation, aiming to improve efficiency and consistency.
    • Anthropic: Claude for Nonprofits (Dec 2, 2025)
      Anthropic, in partnership with GivingTuesday, is launching Claude for Nonprofits to help organizations maximize their impact through discounted access to Claude AI, connectors to nonprofit tools like Blackbaud and Benevity, and a free AI fluency course.