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.

I spent part of last week watching Dan Shipper and Claire Vo talk about how AI has changed the way they work. Different podcasts with different contexts, but both had the same conclusion: they were both doing more work than before. This is not what most people expect to hear from two AI-forward operators. The story circulating in executive circles is that AI means fewer people, less work, and leaner teams. Shipper's company doubled in size this year, and he is still hiring.

His explanation: automation is a lie. Every time you automate something, you create a new place where a human has to show up. The machines did not eliminate the need for judgment. They moved it upstream.

The reason has to do with how models actually work. They make yesterday's competence cheap. Whatever skill used to be hard and expensive gets absorbed into the model and stops being a differentiator. The humans who were doing that work do not disappear. They move on to the next thing the model cannot do yet, which always requires genuine judgment about what the work should be.

Shipper told a story about this. He vibe-coded an app called Proof almost entirely in Claude Code, launched it, and watched the servers crash every ten minutes the day after launch. He fed the problem back to the models. They reported fixes, and the fixes broke other things. When he eventually brought in two senior engineers, they looked at the same list of issues and said: we need to rethink this whole thing. Patches won't hold.

What Shipper is building toward is not humans monitoring agents from a distance. He described his current setup as having a parallel work buddy: CodeX can see what he is doing, he can see what CodeX is doing, both present in the same work at the same time. The quality comes from the back-and-forth, not the handoff.

Which is why what Claire Vo and Thariq Shihipar demoed clicked into place. Thariq switched from markdown to HTML for implementation plans. Simple reason: markdown plans got so long he stopped reading them. When he stopped reading, he stopped editing. When he stopped editing, he started asking Claude to edit the plans instead. He had handed off the judgment without meaning to.

HTML fixed it because it is visual, scrollable, and can hold actual diagrams. He reads it now. He marks it up, pushes back, and reshapes it before the agent runs. A plan he will actually engage with is a different input than one he rubber-stamps. Thariq's estimate: 1% of the tokens he generates are used in production code. The other 99% go into plans, dashboards, throwaway interfaces, and status updates. That coordination work is where the human contribution lives.

Claire Vo calls the role "compute allocator." (She offered it as a joke.) When you tell an agent to run for eight hours, you are authorizing a spend. That decision requires you to have actually read what you are authorizing. An agent running on a brief you approved without reading is not collaboration. It is just a faster way to end up with the problem of you having to take the reins and start from scratch.

The anxiety I hear about AI and jobs is real and pointed at the wrong thing. The question is not whether agents will do more work. They will. The question is what humans do alongside them. Shipper's answer, backed by his own headcount, is that the human role expands into everything that requires genuine presence in the work: deciding what to build and catching the moment the agent is heading somewhere wrong.

The models make yesterday's competence cheap. The humans figure out what comes next.

The first sign that computer use is becoming a real part of how someone works usually has nothing to do with AI. It’s a power setting. They’ve turned off sleep mode, adjusted the display timeout, and done whatever it takes to keep the machine awake overnight. Last year, I would have guessed that was about surveillance. Now I know it’s the opposite: they’re keeping the screen on because something is running, and if the machine sleeps, the work stops.

I’ve also started noticing people walking to meetings with laptops cracked open, riding the subway with the lid up. Often, it’s Claude Code or another agentic coding tool running a session that cannot afford to be interrupted. The lid-closed problem became common enough to create its own small ecosystem: caffeinate scripts, and even a new dongle designed to keep agents alive while humans moved around them. Then the workaround became a product feature.

What looks like a strange human habit is really an artifact of a new machine behavior. The computer is no longer just displaying the work. It is becoming the surface through which the agent works.

What we’re actually witnessing is 'computer use', which is just the industry shorthand for a model physically driving software through the UI. It's taking screenshots and clicking buttons just like you or me. No API required, no developer integration, no assumption that the software was built with automation in mind. That is what makes it different from every other way AI has been connected to software to date.

The underlying loop is simple, which is why the shift is so large. The agent takes a screenshot, reads it, decides what to do, acts, and then looks again. Any interface a human can use by looking at it, the agent can use in the same way. This is the part that changes the map of what can be automated. For the past several years, connecting AI to your software meant either building an API integration, which requires developer time and assumes the software even has one, or doing things the slow way, copying outputs from one window and pasting them into another. Computer use removes that bottleneck. Any tool in your stack, regardless of whether it was designed for automation, is now a candidate.

Computer use rewards workflow design. Before I put an agent inside any real workflow, creative or operational, I ask three questions. Can a human quickly verify the output? If something goes wrong, is the damage recoverable? Can I see, step by step, exactly what the agent did and why? If any of those are no, I add a checkpoint before deployment and keep it there until the answer changes.

The product roadmap is starting to reflect the behavior. The major systems are moving in the same direction: agents with longer sessions, more durable access, and more permission to act while the human is elsewhere. The important change is that they are being given more room to operate. The machine staying on overnight is what that looks like in practice. A workflow, not a workaround.

Google I/O: Twenty Announcements, Zero Sparks

Google spent most of its keynote drawing a deliberate line between AI that assists and AI that acts, betting the future is ambient. You won’t go to Gemini; Gemini will just be running. Whether that sounds like leverage or surveillance probably depends on how much you trust Google with your inbox.

Google’s benchmarks at I/O showed Gemini 3.5 Flash outperforming Gemini 3.1 Pro. That’s a real improvement. It’s also comparing itself to itself. The company spent considerable keynote time on “4x faster output tokens per second” without naming who it’s faster than.

The honest read on I/O 2026 is that Google is closing a gap. Gemini Omni, its new video-editing model, is getting mixed reviews (not as good as the Chinese models, but okay for editing). Google also announced an update to the Search AI Mode, which we aren’t sure the public asked for. Nothing at I/O 2026 felt mindblowing; it felt like a very large, very well-funded company working hard not to be left behind.

Google announced somewhere between 20 and 100 new things at I/O 2026, depending on whether you count the developer stuff. The announcement included: Gemini 3.5 Flash, Gemini 3.5 Pro, Gemini Omni, Gemini Spark, Daily Brief, Android Halo, Gmail Live, Docs Live, Google Keep AI, Google Pics, Universal Cart, Neural Expressive, information agents, mini apps, Ask YouTube, Flow, Flow Music, AI Mode upgrades, SynthID expansion, C2PA credentials, and Android XR glasses. Each one lives in a different app. Each requires a different subscription tier, and several are not available until “this summer.”

Gemini Spark, the feature that generated the most genuine excitement, the personal agent that actually does work on your behalf, isn’t available yet. But it will be next week for Ultra subscribers at $100 a month.

I genuinely cannot tell you which Gemini product to use for which task. I suspect most people at Google cannot either. This is the I/O tradition: a tidal wave of names, a handful of things you can actually use today, and the best part is arriving soon.

Of Popes and Pretraining: This week was quite newsworthy for Anthropic, and for the first time not due to feature releases. Somewhere between a $1.25 billion-a-month SpaceX compute contract and a papal encyclical, Anthropic stopped looking like a challenger lab.

First: Andrej Karpathy (co-founder of OpenAI, architect of Tesla Autopilot, and the person most responsible for the current generation of AI-literate non-researchers) joined Anthropic to rebuild its pretraining research. He cited "the next few years at the LLM frontier as especially formative" and left it there. The man who has spent more time than almost anyone making these models legible to a general audience looked at the field and chose Anthropic.

Second: Pope Leo XIV presented his debut papal encyclical, Magnifica Humanitas, on the protection of the human person in the age of artificial intelligence. Standing next to him was Christopher Olah, Anthropic's co-founder and the world's leading researcher in mechanistic interpretability (the science of understanding what is actually happening inside AI models). The institution with the longest moral time horizon in Western civilization shared a stage with Anthropic.

This was also the week Anthropic posted its first quarterly operating profit and closed in on a $900 billion valuation. Anthropic is apparently winning the talent war, the ethics debate, and the revenue race in the same quarter. Karpathy's move signals that serious researchers believe the frontier is still built at the pretraining layer and that Anthropic is where that work happens. The Vatican's move signals that Anthropic's safety positioning has accumulated enough real credibility to carry weight outside Silicon Valley.

The Baristas were right: Starbucks rolled out an AI inventory counting system promising 99% accuracy and faster counts. Nine months later, after miscounting oat milk as dairy milk, and requiring baristas to wave iPads at refrigerators in a workflow slower than the color-coded labels it was replacing, the company retired it. The core problem wasn't that AI doesn't work. It was a list of conditions familiar to any enterprise software rollout: a supply chain fragmented across regional vendors with non-standard packaging, seasonal product changes that compounded miscounting, no error-handling when counts went wrong, and a legacy IBM inventory platform that couldn't absorb even the accurate data the system occasionally produced. The company hasn't abandoned AI altogether, but it remains in the background.

The Flat Rate was the product: Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable. This is the same company that invested $13 billion in OpenAI and operates the Azure infrastructure running most of the AI industry’s compute. The announcement arrived alongside a memo from Uber’s CTO warning that the company had burned through its entire 2026 AI budget in four months. American AI software pricing has jumped 20 to 37 percent across the board over the same period. GitHub, which Microsoft owns, announced it is dropping flat-rate plans and moving to usage-based billing across its products. Anthropic, OpenAI, and Google all raised effective prices over the last six months. Enterprises built workflows assuming AI costs would continue to fall. They are now watching annual budgets collapse in a single quarter.

— 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.