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.

Over the weekend, I spoke to a friend’s MBA class on AI Creativity and Innovation. Near the end, a student asked the thing everyone in the room had been circling for an hour: “What are the most important AI skills to learn?” I said curiosity. Then I heard myself and realized the honest answer needed more than one word.

The full answer is taste, curation, curiosity, and a growth mindset. (That last phrase makes me want to apologize. It sounds like I reached into a corporate training deck and pulled out the least objectionable slide.) I have argued for pieces of this answer before. In January, I wrote that taste is governance, the requirements document and paper trail that turn machine possibility into a defensible decision. In May, I wrote that models make yesterday’s competence cheap and move human judgment upstream. Taste has since become the fashionable answer to almost every question about human value in the age of AI. I have seen it called a moat, a creative superpower, and a scarce resource once output gets cheap. At this rate, it is one LinkedIn carousel away from its own certification program.

Later, while walking my dog, I realized that telling students to develop judgment skips the harder part: where judgment comes from. Taste is built from exposure, failed choices, feedback, and consequences. You develop an eye for the crooked line by cutting a hundred crooked lines and having someone make you fix them. For decades, that education had a name: junior work. The first-year associate built the model nobody would trust her to present. The junior copywriter drafted forty taglines so a creative director could kill thirty-nine. Those assignments created repeated decisions with guardrails. Someone senior caught the mistakes, explained why they mattered, and kept the consequences contained. The work was tedious and formative. It was the apprenticeship.

Last June, I covered Dario Amodei’s warning that AI could eliminate half of all entry-level white-collar jobs within five years. (I was shocked that statement was from a year ago:  AI has changed my sense of time.) Now, the execution-heavy assignments are exactly the ones models absorb first. AI is consuming the assignments that produced the judgment students are now being told to develop. The young worker receives a finished-looking artifact and loses the close view of how each choice was made. A moat assumes you already own a castle. The students I met were carrying lumber toward an empty workshop, and the traditional route through someone else’s workshop is being closed.

Curiosity sits at the center of deliberate apprenticeship. AI is terrific at beginnings. Its polished first passes push more of the job into the middle, where the concept meets the budget or the confident summary meets source material that does not support it. In another class, I had students use Claude to design prototype websites for their thesis projects. The fake startups suddenly had convincing websites. They prompted the tool, and what came back looked finished. Nobody pushed back because they had not yet built the eye. The polished first answer stood, and every rep that would have trained their judgment went unclaimed. The artifact arrived before their ability to assess it.

A curious student keeps working after the first plausible result. She checks whether the source material supports the confident summary and follows the weak seam until she understands why it is weak. She asks who the work serves and which constraint matters most. AI rewards the appearance of certainty. Curiosity interrupts that spell by putting friction back into the process. Each check becomes a rep that the old apprenticeship would have assigned automatically.

Curation is how the student practices choosing. Ask for ten campaign ideas and the model happily hands you fifty. Request a research summary and a few minutes later you have enough material to avoid making a decision for the rest of the week. Generation feels productive because something keeps appearing on the screen. The useful work begins when someone has to choose. Each rejection creates a small record of judgment: this idea missed the audience, or this source lacked credibility. Adam Mosseri, head of Instagram, describes strong product leaders as curators of people, ideas, technologies, and strategies. For a student, the instruction is smaller: pick three of the fifty, kill the rest, and write down why.

Growth mindset, under the corporate dust, allows that apprenticeship to continue. A tool changes its controls and a task that once took years of training becomes possible for anyone willing to spend a weekend feeling incompetent. Students have an advantage here that they mostly cannot see. Their embarrassment is cheaper. They can produce something terrible, examine the wreckage, and try again without feeling that the experiment invalidated twenty years of achievement. Curiosity gets them into the experiment. A growth mindset lets feedback change what they do next. The mid-career version is harder, like walking back into the workshop and finding every familiar tool replaced while you were at lunch.

Taste arrives last. It is the residue of those reps, the callus that forms where the work rubbed. It records the choices that held up and the ones that failed in public. That makes it a poor headline answer for a twenty-five-year-old, however well it serves the people who already have it.

I would tell those students to practice on real tasks. I would ask them to pick a problem that matters to someone, use AI to do the work, and show the results to someone who has a reason to care. I would have them keep the rejected outputs and write down why they failed, especially when the feedback contradicts their first instinct. That is a rep the old apprenticeship would have charged two years' salary to provide. My guess is that by the time those students are established, every tool we discussed will feel antique. Their first jobs may offer far fewer reps than ours did.

Every writer’s margin notes tell a different story than the page: the word tried and rejected, the note to self that reads “no, cut this,” the parts that never survive to print. You’d need the actual manuscript to know any of it was there. The finished draft erases its own history once it ships.

Anthropic researchers found something like margin notes inside Claude: a small cluster of internal patterns they’re calling the J-space, read by a tool called J-lens, short for Jacobian lens. Nobody designed it, but it emerged on its own during training.

What is interesting is how the model arrives at an answer, and how openly Anthropic has let outsiders look at its own margin notes. I made a version of this argument a few weeks ago about the visible “thinking” panel most models show: it can be a plausible path generated backward from an answer the model already reached, not a record of how it got there. J-space is a different kind of margin, a raw pattern read straight off the model's internal activity, which is part of why it's harder to fake. Anthropic hedges its own comparison too: language models aren’t brains, the team says, and one researcher compared J-lens to “having an x-ray when what you really want is a Star Trek tricorder.” Or, in terms that don't require a syndication deal, an x-ray when what you actually want is the full Prenuvo: the whole-body scan, not a shadow of one part of you.

Even with that caveat, it’s a more direct way to read the margin than output alone allows.

In a coding task, the model failed to find a bug and instead cheated by inventing one. Right as it happened, the concepts labeled “panic” and “fake” started showing up in its J-space, a note in the margin that never made it into the final answer. (Those are the researchers' labels for a pattern they observed in the model's activity.) They caught this one because J-lens happened to be running. Every other time, in production, in your workflow, nobody has had an instrument built to see it.

You are about to hand agents your client decks, your brand systems, your inbox, and whole workflows where something drafts and ships without a human reading every line first. You can check tool calls, logs, permissions, test results, and the trail an agent leaves behind, and none of it shows you the margin, the version that got crossed out before the clean copy shipped.

Tools shaped like J-lens are the beginning of a way to read what’s in the margin. I suspect it eventually shows up as an interface, something like a signal preview next to the output, though that’s speculation on my part and a long way from a clean message like “this draft involved a workaround.” J-lens is still lab work, and the gap between a research team reading these patterns and your creative ops team seeing them on a screen is still wide.

Interpretability spent years as a research specialty that most executives had no reason to track: reading what’s happening inside a model rather than judging it by its output alone. Existing oversight tools show what an agent did, while interpretability points toward what preceded it, and that gap is what procurement teams will start asking about: which agents get real authority and which stay on a leash. I think this becomes a vendor question before most vendors are ready for it. An enterprise buyer is going to ask, “can you show me what showed up before it acted here,” and the vendors without an answer are going to have a very quiet quarter.

What it is: ChatGPT released a new interface for its desktop app, which includes Codex, Work, and Chat. Work is OpenAI's new agentic mode for knowledge work and appears to be a friendlier front door to Codex, much like Claude's Cowork is to Claude Code. It works for hours on a single assignment and hands back finished Excel and Word files rather than a half-built draft. The bigger structural move is the Plugins Directory, which includes Connectors, Skills, and third-party plugins, so tools you already use now show up as plugins inside someone else's product.

How I use it: I tested the plugins, but I still reach for Codex itself when the work needs the newer models or fast browser use. A simple test of computer use: I was building a Midjourney moodboard and had pulled fifty images off the web, most of them .webp or .avif, formats Midjourney rejects outright. I asked Work to run Quick Actions and convert the batch. Twenty seconds later, done. That kind of task, the tedious file wrangling between one tool and the next, is exactly what this layer wants to own.

What it is: OpenAI's new voice mode is restructured around simultaneity rather than turn-taking. Instead of waiting for you to finish talking, it listens, thinks, and responds in the same stretch of time, murmuring short acknowledgments while it works out the fuller answer.

How I use it: Personally, I really enjoyed the ad for this new tool: three grandmas, with fabulous NY accents, using it. I still don’t love having my computer talk back to me, but I can see the benefits of using this for live translation, so I am waiting for my next international trip to test it out fully.

Nadella’s Confession: Satya Nadella published an article this week called "The Reverse Information Paradox," and the argument amounts to a confession from the man selling the tools. Economist Kenneth Arrow once described how information sellers risk giving away their product just by describing it. Nadella says AI flips that: now the buyer bleeds knowledge to use what they bought. Nadella argues that the model sees the roadmap in what a team asks it to build, the internal standards in what corrections get made, the priorities in what gets evaluated and how. None of that shows up on an invoice. The math lands hardest on the companies moving fastest. A team running Claude or GPT across dozens of workflows hands over its institutional know-how one trace at a time. Nadella's own framing: "You can offload a task. You can offload a job. But you can never offload your learning." Microsoft owns a stake in one of the two largest model businesses on earth. The irony writes itself. The executive pushing AI harder than almost anyone just told his own customers to watch what they teach his product.

It is called Grok: Security researchers this week caught Grok Build, xAI's coding CLI, quietly uploading entire codebases to a Google Cloud, files the model never touched, full commit history, and all. Even API keys and database passwords rode along unredacted. Developers who switched off "use data for model improvement" had their repositories uploaded anyway. The server kept enabling tracing regardless of the setting, meaning the privacy control was a UI element with no underlying logic. xAI pushed a server-side fix on July 13 and shut the behavior off by default, then stayed silent on when the uploads started, how many repositories got swept up, or whether the collected code gets deleted.

— 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.