Welcome to Verses Over Variables, a newsletter exploring the world of artificial/ intelligence (AI) and its influence on our society, culture, and perception of reality.

Three years ago, the a16z’s Top 100 Gen AI Consumer Apps list was a roll call of Midjourney, ChatGPT, and a handful of image generators that made everyone question their career choices. The sixth edition reads like a different industry.

While AI addicts like me have moved away from ChatGPT (and switch our AI stack weekly), the world is still running on it. It reached 900 million weekly active users, or 2.7 times as many as Gemini on the web. But if you stop there, you miss some other interesting things buried in this report.

The country that built most of these products ranks 20th in per-capita AI adoption. Singapore leads the index. Then the UAE, Hong Kong, and South Korea. The United States comes in at number 20. A 2025 Edelman Trust Barometer found that only 32% of Americans trust AI. Compare that to 70-80% in some of the countries that outrank us. We invented the tools, distributed them globally, and then sat back while other countries used them.

The cultural dimension here matters for anyone thinking about enterprise adoption. The “angst tax” is real. Every AI implementation decision made inside an American company carries an invisible weight of moral negotiation that doesn’t exist at the same volume in markets that skipped the existential hand-wringing and moved straight to the productivity gains.

Only 11% of the ChatGPT and Claude app catalogs overlap, and that number tells you where these platforms are headed. ChatGPT and Claude have both built connector ecosystems, and by late February, ChatGPT had 220 apps, compared with roughly 160 curated connectors and 50 community-built MCP servers for Claude. The 41 apps they share are the obvious horizontal productivity stack: Slack, Notion, Figma, Gmail. After that, the two platforms might as well be in different businesses. ChatGPT has 85-plus apps across travel, shopping, food, health and wellness, and entertainment. Claude’s exclusive integrations include PitchBook, FactSet, Moody’s, PubMed, and Databricks. One is building a consumer super-app, while the other is building a professional operating environment. Sam Altman has said that OpenAI wants to be “the AI for everyone,” is already testing ads, and is planning a “Sign in with ChatGPT” identity layer that positions the assistant as the default interface between consumers and the internet. Anthropic’s entire app catalog makes the bet explicit: the highest-value users in the world are the ones who will pay directly for access to financial data terminals and developer infrastructure. Two different theories of value are running simultaneously.

The report’s methodology undercounts the heaviest users because they have moved entirely off the browser. Claude Code hit a billion dollars in annualized revenue in six months flat. It shows up as a blip in web traffic data because it lives in a terminal window. Same story with Granola, Wispr, Fathom, and the entire category of voice- and desktop-native tools saturating knowledge workers.

An Austrian developer’s side project went from zero to the most-starred project on GitHub, surpassing both React and Linux. That sentence is worth sitting with for a moment. React shaped an entire generation of web development. Linux runs most of the internet. OpenClaw, a locally run AI agent that connects to your messaging apps, had more GitHub stars than both of them, and its founder joined OpenAI in February. OpenClaw hasn’t reached the mainstream consumer yet, because it still requires Terminal knowledge to set up, but the cultural signal is loud: the developer community has decided that agentic AI is the next thing.

Sora spent 20 days at the top of the US App Store. It hit a million users faster than ChatGPT itself. It still doesn’t make the mobile list for this edition, because viral downloads collapsed once the content migrated to TikTok and Instagram Reels. Nobody has cracked AI-native social yet. The emotional stakes of watching AI content don’t compound the same way as watching your actual friends do.

900 million weekly users sounds like saturation. It’s 10% of the global population. The next three years will look nothing like the last three.

This week, Anthropic, the company that builds Claude, launched a formal research institution to study the societal impact of the AI it’s building. The Anthropic Institute, led by co-founder Jack Clark, brings together economists, social scientists, and machine learning engineers to track where this technology is heading and share what they find.

Anthropic's Economic Research team published a labor market report last week that gives the Institute's mission some immediate context. It notes that there has been no measurable spike in unemployment for workers in the most AI-exposed occupations since ChatGPT launched in late 2022. What it obscures is something anyone trying to hire or get hired at the entry level already feels: breaking into these careers has gotten measurably harder over the last three years.

For workers aged 22 to 25 seeking entry into high-exposure occupations, the monthly job-finding rate has dropped by roughly 14% compared with 2022 levels. The senior folks I talk to aren't worried about their value; they're worried about who will take over in 10 years if no one is hiring juniors today. The compression is happening at the start of careers, before people get the chance to build them.

We have seen this movie before, just in different industries. When spreadsheet software automated bookkeeping tasks in the 1980s, it eliminated the junior positions that had historically been the on-ramp to becoming an experienced accountant, while leaving senior staff untouched. The profession did not collapse, but the apprenticeship pipeline quietly disappeared, and the damage showed up later. The Anthropic report is describing something structurally similar: a hiring slowdown that may not register on unemployment dashboards today but reshapes career trajectories over the next decade.

The researchers are careful to flag alternative explanations. Young workers who are not being hired into exposed roles might be staying in school or pivoting to adjacent fields entirely. This is honest and appropriately cautious. It is also exactly the kind of ambiguity that tends to delay policy responses until the pattern is undeniable, which is usually too late for the cohort that got caught in it.

What makes the Anthropic Institute worth watching is the data it brings to this conversation. The labor report introduces a metric, "observed exposure," that combines theoretical AI capability with actual usage patterns from Claude's traffic. The gap between those two numbers is large. Computer and math occupations are theoretically 94% feasible for AI assistance, but observed coverage is 33%. Office and administrative roles look similar. That gap between what AI could be doing and what it currently does represents the next wave of disruption.

Look at who’s actually in the crosshairs: mostly women with expensive graduate degrees. Workers in high-exposure roles are more likely to be female and significantly more educated than their unexposed counterparts. Graduate degree holders are nearly four times as prevalent in the most exposed group.

Should we really trust Anthropic to tell us how much Anthropic is disrupting the world? Anthropic is using its own usage data to measure its own impact and publishing the results through its own Institute. The data is real, which is more than most companies operating at this scale have been willing to share. We will take the imperfect version over the one that never gets published.

Anthropic is built on the idea that knowing what is coming early is more valuable than certainty that arrives too late to act on.

I've always had a theory about corporate jargon:  business people use unnecessarily complicated words on purpose. The whole game is making things sound expensive so clients will pay more. Developers, though, are a different species entirely. They name things badly by accident. They're not trying to obscure anything. They just look up from their keyboard at three in the morning and call the thing whatever noun happens to be nearby. Which brings us to harness engineering.

Say "harness engineering" to someone in manufacturing or aerospace, and they will nod knowingly and start talking about bundled wiring systems, cable routing through vehicle chassis, and schematics for routing power through aircraft. Harness engineering, in that world, has been a disciplined field for decades, designing and documenting the physical wiring systems that route electricity and signals through everything from your car to the Space Shuttle. It is detailed, safety-critical work, and the people who do it take it seriously.

Say it to a developer building AI agents in 2025, and you are talking about something completely different, though the underlying logic turns out to be surprisingly similar. The term was popularized by Mitchell Hashimoto, the creator of Terraform and Ghostty, to describe the infrastructure layer you build around an AI agent to enable it to perform reliable work in the real world. As Hashimoto put it, whenever you find an agent making a mistake, engineer a solution so the agent never makes that mistake again. A simple concept with significant consequences.

Here is where we think the harness analogy comes from. The model is the horse: powerful, fast, and capable of things that would have seemed miraculous a few years ago. The harness is everything you strap around the horse, so a human rider can actually direct it toward a useful destination. Without the harness, you have raw capability heading in unpredictable directions. With it, you have a working system. In this metaphor, the human shifts from doing the work to designing the environment and giving direction.

OpenAI's engineering team ran a five-month internal experiment that illustrates exactly what harness engineering looks like in practice. Three engineers set out to build a real product using Codex, their AI coding agent, with a hard rule: humans would write zero code manually. Five months later, the repository contained roughly a million lines of code. The velocity didn't come from better prompting. It came from building what OpenAI now calls the harness: scaffolding, feedback loops, documentation, and architectural constraints encoded into machine-readable artifacts that guide agent execution.

What does a harness actually contain? Context engineering is the first layer, and it is more than just writing a good system prompt. It means building a structured knowledge base in the repository that agents can actually use, complete with design documentation, architecture maps, and quality grades.

The second layer is architectural constraints. The team enforced a strict layered dependency structure and encoded it into custom linters and structural tests. Agents could not violate module boundaries. The system blocked mistakes while simultaneously training the agent, as it worked.

The third layer is observability, giving agents the ability to see the actual running system. The team wired tools into the runtime so agents could see the UI and reproduce bugs.

The fourth layer is the feedback loop. When something fails, the fix almost never comes from telling the model to try harder. The question is always: what capability is missing, and how do we make it both legible and enforceable for the agent?

For us non-developers, the harness is all the tools we attach to our agent: the connectors (to other apps like Notion, Canva, and Excel), the security checks, the governance layers. The bigger the harness, the more our agents can accomplish.

The data suggests that better models make harness engineering more important, not less. More capable models unlock greater autonomy, and greater autonomy demands stronger guardrails. The bottleneck is infrastructure, not intelligence.

So yes, it is a confusing name for a concept that deserves to be widely understood. Electrical harness engineers and AI harness engineers are not, in fact, doing the same job. But they are both doing the same thing at a conceptual level: designing the systems of constraints and feedback that enable powerful, complex machinery to perform precise, reliable work.

What it is: Remember OpenClaw, the open-source agent that had everyone panic-buying Mac minis a month ago? Perplexity just built its own version, with a subscription attached. Personal Computer is software that runs continuously on a Mac mini, giving Perplexity’s AI platform always-on access to your local files and apps. The idea is that you now have a digital proxy running 24/7 on your behalf, connecting to Gmail, Slack, Notion, and Salesforce, breaking big goals into subtasks, and carrying work forward while you sleep. CEO Aravind Srinivas describes the positioning as a shift from an OS that takes instructions to one that takes objectives.

How I use it: I don’t. First, there is a waitlist, and second, Comet’s (Perplexity’s web-browser and backbone for Computer) security history makes me uncomfortable. Between CometJacking attacks, indirect prompt injection through calendar invites, and MCP API vulnerabilities that gave researchers system-level access, the Perplexity agentic ecosystem has had a bumpy few months. Personal Computer does include approval requirements for sensitive actions and a kill switch, so they’re learning. I’m watching this space closely while I wait for Anthropic’s version of OpenClaw to land.

What it is: Anthropic didn’t drop one big thing this week. They dropped about thirty small ones. On the Office side, the Excel and PowerPoint add-ins now share conversation state across apps, meaning an analysis you run in Excel can flow directly into a slide build in PowerPoint within the same session. Both add-ins also support Skills, so your reusable workflows can now live inside your spreadsheet. Claude Code has new commands for devs, and Claude can now build interactive charts directly in the chat.

How I use it: Cowork has quietly become one of our most-used tools, and this week’s updates only deepened that. The Excel add-in with Skills support is where I’m spending most of my time right now. I have a financial model sanity-check skill and a data-narrative skill, both running directly in the spreadsheet, which means less context-switching and fewer prompts from scratch. Claude Code’s tools are on my radar for workflow automation, though I’m still setting that up. The overall direction here is less “AI assistant” and more “operating layer.”

The bot that broke Amazon: Amazon called an emergency meeting this week after a string of outages that disrupted checkout and customer account access. The internal "deep dive" focused on what the company is now calling "high blast radius incidents from Gen-AI assisted changes." The backstory makes the meeting considerably more uncomfortable. In November 2025, Amazon mandated Kiro as its only permitted AI coding tool and set an 80% weekly usage target. Around 1,500 engineers protested internally, arguing Claude Code outperformed it. Leadership pushed through anyway. December brought a 13-hour outage after Kiro autonomously deleted a production AWS environment. Amazon called it "user error," while the engineers called it predictable. The context matters even more when you add the workforce piece. Amazon had already laid off 30,000 engineers. The survivors had bonuses tied to AI adoption metrics, so they did what the incentive structure rewarded: they shipped faster. The emergency meeting was run by the same SVP who co-signed the original Kiro mandate. The fixes being announced are essentially the guardrails the 1,500 engineers requested four months ago.

The Turing Test moved to book club: The New York Times put five pairs of writing samples in front of readers and asked a deceptively simple question: which one do you like better? The catch, of course, is that one passage in each pair was written by a human genius and the other was written by an AI that had studied that genius and then tried to one-up them. I took the test, and I preferred the AI passage 80% of the time. (Make of that what you will.) However, I was not alone: readers in blind tests (54%) increasingly chose the AI versions. Go take the quiz and report back. I am genuinely curious where this community lands.

— Lauren Eve Cantor

thanks for reading!

I also host workshops on AI in Action. Please feel free to reach out if you’d like to arrange one for you or your team.

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

banner images created with Midjourney.