Andrew Maddox

Blog

  • 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.

  • Google Gemini 3.0

    Google released version 3.0 of Gemini on November 18 and a few days later released Nano Banana Pro, their image generating model. Both were quickly declared to be some of the best models, although Anthropic later released version 4.5 of their topline Opus model, reclaiming leadership by some accounts. 

    This chart from METR shows LLM task success rates for a longer task duration. The present leader is GPT 5.1 Codex Max at more than 30 minutes of non-supervised work. The previous version of Gemini (2.5) was able to work for ~10 minutes, so I expect that Gemini 3.0 will be near the leading edge in the coming weeks.

    Progress continues for each of the major AI companies, and the developments in 2025 alone are remarkable. Gemini, with Google’s data-center and intellectual heft, is clearly well-positioned despite it’s earlier hallucinations and AI issues. OpenAI is reportedly alarmed by these tools and have declared “code red” to improve their product quickly.

    Below, you’ll find some of the more interesting examples I’ve seen of Gemini / Nano Banana Pro.

    …

    Solving Homework

    From Andrej Karpathy:

    Gemini Nano Banana Pro can solve exam questions *in* the exam page image. With doodles, diagrams, all that.

    Gemini 3 Pro also scores a 100% on the Best in High School Math (AIME 2025) evaluation, a competitive high school math benchmark.

    What does this mean? Are you a student struggling to understand your math assignment? Simply snap a photo of your assignment, and Gemini will complete the work for you. This could be helpful for understanding and checking work. But this could also be a very easy way to cheat, as I’ve read some reports that Gemini is able to replicate an individual’s handwriting style (meaning someone wouldn’t have to manually copy the work). For instructors, the idea of a student mastering take-home assignments may no longer be a viable indicator of learning.

    …

    Infographics

    From Grigory Sapunov, Visualizing Research: How I Use Gemini 3.0 to Turn Papers into Comics:

    I asked Nano Banana Pro to generate a graphic novel telling the story and explaining the most important concepts based on a summary I provided. Here is the result:

    Grigory’s post is fun and features a variety of comics with differing styles that communicate core ideas of very technical research. This, of course, won’t supplant the importance of original research, but it has the possibility of translating technical knowledge to the masses.

    Google also provided some inspiration on ways to use Nano Banana Pro: (Prompt: Create an infographic about this plant focusing on interesting information.)

    The prompt and image input are simple, and the generated infographic is interesting and full of helpful details. And best of all, it took a human very little time to create.

    What does this mean? Are you in the business of communicating interesting but ultimately difficult-to-understand research? You can generate visuals to convey key ideas within complex research to media and students. Are you in the business of creating infographics? I suspect that if you’re exceptionally talented, you’ll continue your work without much interruption. But for middling designers, your business may dry up as people find more cost-effective ways to create supporting graphics.

    Have you used Gemini 3.0 Pro yet? What are your observations of where the tool exceeds?

  • Various (AI) Monday Links (Dec. 1)

  • Tuesday (AI) Links (Nov. 18)

    • WSJ: These Small-Business Owners Are Putting AI to Good Use (Nov 15, 2025)
      Small businesses are adopting generative AI tools to streamline operations, improve customer service, and boost marketing efforts. Examples include using AI for financial analysis, automating customer service responses, and generating website code, leading to potential cost savings and reduced hiring needs.
    • NY Times: A.I. Chatbots Are Changing How Patients Get Medical Advice (Nov 16, 2025)
      Frustrated with the medical system’s shortcomings, patients are turning to AI chatbots for health advice, reshaping doctor-patient relationships, with some patients using AI-generated information to challenge or bypass their doctors. My take: if patients feel dismissed or in need of solutions, they’ll turn to alternative sources. If anything, this is a call for humility and research in medical sciences.
    • Futurism: People Are Having AI “Children” With Their AI Partners (Nov 15, 2025)
      A new study reveals that some users of AI chatbots like Replika are developing deep, romantic relationships with their virtual partners, even roleplaying marriage, pregnancy, and homeownership.
    • NY Times: Europe Begins Rethinking Its Crackdown on Big Tech (Nov 17, 2025)
      Once hailed by many as providing welcomed online privacy protections, the narrowly conceived GDPR has proven to stifle innovation, particularly in the AI sphere. This is a warning for policymakers everywhere that poorly written rules can cause more harm than good.
    • Fast Company: AI is killing privacy. We can’t let that happen (Nov 16, 2025)
      While the EU considers lessening privacy regulation, others warn that tech companies are collecting and using our data in ways that could be harmful. The theology in this opinion piece is suspect, but concerns for privacy in an AI-driven world are unlikely to disappear any time soon.
    • Sean Goedecke: Only three kinds of AI products actually work (Nov 16, 2025)
      The initial AI boom has led to only three types of useful LLM-based products: chatbots, completion tools like GitHub Copilot, and coding agents.
    • WSJ Opinion: When Will AI Elect a President? (Nov 16, 2025)
      The future of media, driven by AI chatbots like ChatGPT Pulse, will be highly personalized and could be exploited by campaigns to target voters with unprecedented precision, raising concerns about manipulation and the commodification of attention.
    • NY Times: Jeff Bezos Creates A.I. Start-Up Where He Will Be Co-Chief Executive (Nov 17, 2025)
      Project Prometheus will focus on applying AI to engineering and manufacturing in fields like computers, aerospace, and automobiles, positioning itself in the competitive AI landscape.