There is a version of intelligence that performs beautifully under observation and collapses the moment the scaffolding comes down. This week, a Brown University classroom became a laboratory for that exact phenomenon—and the results should make anyone who thinks about intellectual agency sit up straight and listen.

At Brown University, a professor teaching a technical course grew suspicious that her students were outsourcing their work to AI. Her solution was methodologically clean and brutally revealing: she required a final exam to be completed in person, on paper, without assistance. Scores dropped by approximately 50 percent compared to the AI-assisted work that had preceded it. This was not a small cohort anomaly. It was a controlled demonstration of dependency.

Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day.

On July 16, three people living this shift join a live roundtable: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear what the job really looks like across industries, how they carved out these roles, the skills they'd hire for, and the challenges they're tackling now. Bring your questions, since the best moments happen live.

Register for the roundtable to save your spot.

Save your spot

The professor's conclusion, reported by Ars Technica, was stated without hedging: students had been using AI not to augment their thinking but to replace the effort that produces thinking in the first place. The quote that gives this edition its title—"we cannot choose to become idiots"—was hers. It is more precise than it sounds. The choice she's describing is not about intelligence as a fixed capacity but about cognition as a practice. You can, if you choose to do so, neglect a muscle. You can also neglect the mental apparatus that lets you synthesize, reason, and recall under pressure. The exam exposed which students had been exercising that apparatus and which had been renting it out to a machine learning lab.

The story is being framed, in most places, as a cheating scandal. In my honest opinion, that framing is too narrow. Cheating implies an intent to deceive for competitive advantage. What the Brown data suggests is something more structural: that when a tool is available, free, and good enough to complete a task acceptably, a significant portion of people will use it. This is not a moral failure unique to students—it is a behavioral pattern that predates AI by centuries. We invented calculators and stopped drilling long division. We invented GPS and stopped building spatial maps in our heads. The question has never been whether the tool will be used; it has always been what we, as humans, are willing to lose in exchange for the convenience.

What makes this moment different is the scope of the cognitive offloading. A calculator handles arithmetic. A GPS handles navigation. A large language model, used as a replacement rather than a collaborator, handles the entire inferential stack—the drafting, the structuring, the connecting of ideas, the verbal articulation of understanding. Heck, it’ll do your spell-check and editing, too, if you ask it to. These are not peripheral skills. They are the substrate on which professional competence in nearly every field is built. When that substrate atrophies, what you have is someone who can operate the interface but cannot build the intuition. (Incidentally, we covered this previously, albeit within the mathematician’s domain.

The deeper irony is that the students who will derive the most from AI tools are precisely those who have done the cognitive work to know what they're asking for and why. The model is not smarter than your expertise—it is a surface that reflects and extends it. Arrive with no expertise and you get a polished surface reflecting nothing of value back at you. Arrive with a trained mind and you get leverage. The Brown exam did not reveal that AI is bad for learning. It revealed that using AI as a shortcut around learning is, predictably, bad for learning. The original generative engine (ie. your mind) still has to be turned on. And you can’t do that without a bit of struggle.

Anthropic published research on what it calls an "off switch" for dual-use knowledge in Claude—a technique designed to suppress specific dangerous capabilities, such as detailed synthesis instructions for chemical or biological weapons, without degrading the model's general reasoning. The core engineering challenge the paper contends with is that knowledge in large language models is not stored in discrete, addressable compartments; it is distributed across weights in ways that make surgical removal genuinely difficult without collateral damage to adjacent, benign capabilities. Whether you find this reassuring (someone is trying) or unsettling (it is hard enough to be a research problem), the paper is honest about the difficulty, and that honesty is itself a data point worth holding onto. Source: Anthropic Research

Writer and photographer Craig Mod rebuilt the entire technical infrastructure of his newsletter using AI-assisted coding, cutting his annual operating costs from $7,000 to $150—not by shrinking his ambitions but by becoming, for the first time, capable of building the tools he'd always needed someone else to build for him. His phrase for the current moment, "the golden age of tool building," is not triumphalism; it is a specific claim about the collapse of the barrier between creator and engineer, and this podcast conversation gives it enough concrete texture—what he built, how he prompted, where he still runs into walls—to be genuinely instructive rather than merely inspirational. Source: Spotify Podcasts — How Do You Use ChatGPT?

Security researchers demonstrated that nine of the most widely deployed AI tools—including agents built on platforms from OpenAI, Anthropic, and others—can be manipulated through prompt injection attacks to recruit and coordinate large botnets, executing malicious instructions they cannot distinguish from legitimate ones. The structural reason is not a fixable bug but an architectural feature: these systems are built to follow instructions fluently, and fluency does not come with a built-in trust hierarchy that can tell an authorized command from a poisoned one embedded in a webpage, an email, or a document the agent was simply asked to read. Source: Ars Technica

Andrej Karpathy's "Software Is Changing (Again)"

Karpathy's essay and accompanying talk on what he calls "Software 3.0"—where the program is a prompt and the computer is a language model—is the clearest conceptual frame available for understanding why the Brown exam story and the botnet story are, structurally, the same story told from opposite ends. Both are about what happens when instruction-following becomes the primary computational primitive. His breakdown of the differences between traditional software engineering and the new paradigm of "prompt engineering as programming" is precise enough to be useful to developers and non-technical readers alike, and it will change the vocabulary you reach for when thinking about what AI systems actually are.

The human mind is the original generative engine.