Issue #4

AnalysisOpenAI used Monday's livestream to ship ChatGPT Images 2.0, an image-generation model that can render multilingual text (Japanese, Korean, Chinese, Hindi, Bengali), lay out full magazine spreads, and, per the company's press materials, "think" before producing output. Gizmodo called it the opening shot of an "AI slop renaissance." The real news sits further down the page. fal, a hosted-inference startup, is already selling an enterprise API for the model on day one, and Codex, OpenAI's coding product, got a concurrent upgrade called Codex Labs. The pattern is familiar. A single keynote bundles a consumer showcase with an infrastructure play so the numbers flatter both sides. Rendering legible Hindi on an infographic genuinely moves the good-enough bar for commercial design work. What it means for anyone paid to arrange pixels is less fun.

AnalysisProject Prometheus, Jeff Bezos's new AI lab focused on "physical AI" for engineering, manufacturing, aerospace, robotics, and drug discovery, is close to a $10 billion funding round at a $38 billion valuation, per the Financial Times and Bloomberg. BlackRock is leading. Prometheus has not shipped a product, has barely described a product, and is being priced above most public defense contractors. The context matters. Software AI is now expensive to catch up in, and the next race is models that move atoms (factory lines, warehouse robots, drug candidates). If you are a second-tier cloud provider selling GPUs, your realistic customer list is shrinking to five or six households with balance sheets. Investors are really buying a position in a physical-AI race they expect to narrow to a few giant labs by 2028.

AnalysisCursor, an AI coding tool built by four MIT graduates in 2022, is in talks to raise $2 billion at a $50 billion pre-money valuation, according to TechCrunch and CNBC. Annualized revenue is reported at $2 billion, with penetration across 67% of the Fortune 500 and roughly 150 million lines of enterprise code generated per day. Thrive Capital and Andreessen Horowitz lead the round, with Nvidia participating. The headline number reads like a bubble marker. The underlying shift is that developer tooling, a category long won by distribution (GitHub) and language (JetBrains), is now won by model quality routed into the editor. Cursor's moat is the routing layer and the habits it builds in its users, because the frontier models come from someone else. That makes it a thinner company than the price suggests, and an unignorable one anyway.

AnalysisKroll, the corporate investigations firm, published a cybersecurity report on Monday: 48% of surveyed organizations have little or no internal governance on AI tool adoption. Companies at low cyber maturity report an 89% incidence rate of AI-related security incidents, versus 54% at the highest maturity bucket. Average spend on testing AI systems with AI runs at 13% of AI initiative budgets. The numbers describe a predictable pattern. Procurement moved faster than security review, agentic tools (AI that can take actions on its own, like sending emails or running scripts) are doing most of the damage, and insurance carriers are several quarters behind. A mid-market CISO (the executive responsible for information security) who green-lit a Copilot rollout last summer is now the one writing this quarter's breach disclosure.

AnalysisMississippi College School of Law announced on Monday that all first-year students must now complete a mandatory course on AI, making it one of the first law schools in the country to do so. The move follows two years of court sanctions against lawyers who filed briefs with AI-hallucinated case citations, and a growing pile of judicial orders requiring disclosure of AI use in filings. The structure is a two-day intensive plus modules through the year. Dean John Anderson framed it as preparing students for jurisprudence that already includes AI, which is a polite way of saying the bar exam will catch up within five years. Other schools have run electives. What shifts here is that the professional pipeline now treats AI as core infrastructure, at roughly the speed word processing arrived in the 1980s.

AnalysisElon Musk confirmed this week that xAI's "Grok Computer," a desktop AI agent that reads the screen, moves the mouse, clicks buttons, types into fields, and navigates between applications, is "coming out soon." A hidden feature toggle in the Grok web app tipped the timing. This puts xAI into a category OpenAI (Operator) and Anthropic (Computer Use) entered last year, and none of the three have made the tool reliably useful yet. The premise is identical across all of them: give the model the same keyboard and mouse a human has, and let it drive. It works on narrow tasks (filling forms, routine admin, scraping data) and fails in weird ways on everything else. The market is large enough that shipping even a mediocre version matters. The security implications, less discussed, are the same reason the Kroll report above looks ugly.

AnalysisEducation Week reported this week that US schools are scrambling to teach media literacy and AI literacy at a pace visibly behind student use. Surveys cited in the piece show AI chat tools are daily-use for a majority of high schoolers, while formal curriculum updates run a school year or two behind. Only a handful of states (California, New Jersey, Delaware) have passed mandates specifically requiring AI literacy. The pattern is the usual one. Kids adopt the tool in months, districts draft policy in years. What is different this time is the shape of the skill gap. A student who uses Claude to write an essay is practicing a different skill (prompting, accepting, editing), and getting worse at the one the assignment was designed to measure. Standardized test scores will tell the story by 2028.

AnalysisUS electric utilities plan to spend $1.4 trillion by 2030 on transmission, new generation, and grid upgrades to power AI data centers, a 27% increase over prior capex (capex is spending on long-lived things like substations, turbines, and transmission lines). Regulators in at least a dozen states are reviewing who pays. The default in most rate structures is "all ratepayers," which is why household bills in Virginia, Ohio, and Georgia have climbed in each of the last three years. The AI infrastructure conversation tends to focus on chip spend (Big Tech's 2026 capex is $650 billion) and quietly forgets the wire. The wire is the slower, more politically exposed number. Maine's data-center moratorium this week is the first legislative backlash. Ohio and Virginia are next. If you pay a power bill, you are already subsidizing the compute race.

AnalysisMIT Technology Review documented this month how Micro1, Scale AI, and DoorDash have hired thousands of gig workers across 50-plus countries, including India, Nigeria, and Argentina, to strap iPhones onto their heads and film themselves doing household chores: loading dishwashers, folding towels, stacking plates, sorting socks. Pay is roughly $15 an hour. The footage trains humanoid robots on the kind of first-person motion data that is hard to collect in labs. The usual label for this work is "data collection." A more honest label is "training your own replacement." The math is brutal. A $40,000 humanoid robot amortized over five years replaces a low-wage worker in a warehouse at well under the cost of employment in most developed markets. The people filming the training data will be first to see the consequence.

AnalysisLaw.com published an analysis this week tracking how recent US court decisions are reshaping copyright liability for AI platforms and their users. The Supreme Court's Cox Communications ruling narrows the grounds on which copyright holders can pursue AI companies absent proof of direct inducement or facilitation. Lower courts continue to classify model training on copyrighted works, including pirated sources, as fair use in most cases tested so far. The Authors Guild settlement in Bartz v. Anthropic produced a payout but no legal barrier. The cumulative picture is clear enough. Training is largely permitted in US courts today, and the live fights are moving to output: style mimicry, voice cloning, generated likenesses of real people. If you are a novelist or illustrator waiting for the courts to save your business model, the calendar is running against you. The fights with teeth are about what comes out, not what goes in.