Frogs in a Pot

How Some of Us Got Protective Gear, and How That Proves the Point

“The kids these days aren’t getting the intellectual HEV Suits that we got in school,” says Gordon Freeman, who famously never said anything at all.

As jokes go, it has legs. Amphibian legs, perhaps. We picture a generation of students being dropped into Black Mesa with an AI Overview, a crowbar tutorial, and no hazard course. They are surrounded by tools that promise instant help, instant summaries, instant solutions, instant escape from the little frictions that used to define schoolwork, office work, and the slow acquisition of competence. Fractions? Ask the machine. Reading comprehension? Ask the machine. First-pass repo review? Ask the machine. An essay title? Ask the machine, then pretend it arrived unbidden while walking the dog.

The harrumph version is easy. We did things the hard way. We learned fractions. We used card catalogs, paper maps, printed manuals, and whatever passed for “research” before every search box started hallucinating confidence back at us. We were toughened by inconvenience. Kids these days are soft.

That version is smug satisfaction sitting on a lawn chair, shouting at kids.

Were prior generations better learners? Of course not. Many of us in those lawn chairs were trained in environments where friction was unavoidable. The current generation is being trained in environments where those unfinished edges and rough corners have been sanded down and padded over.

That is a different claim. It does not scold the frog. It describes the water.

A recent paper on AI assistance and persistence gives this intuition a sharper edge. In randomized experiments, participants who used AI assistance performed better while the AI was available, but worse after it was removed. More importantly, they were more likely to give up. The strongest effect appeared among people who used the AI to obtain direct answers, rather than hints or clarification. This should not surprise us, which is part of why it matters. If the work of learning lies partly in struggling through the problem, then outsourcing the struggle outsources the learning. The unnerving part is not that this happens. The unnerving part is how quickly it can happen, and how natural the outsourcing feels.

The paper’s most interesting implication is not “AI is bad for you, mmmkay.” That framing is too blunt, too flattering to the scold, and too easily dismissed by anyone who has used AI well. The better warning is about reference points. If a tool can make difficult work feel immediately solvable, then unaided effort starts to feel not merely difficult, but inefficient. The first skipped rep is a convenience. The hundredth skipped rep is a curriculum. Eventually, “I should sit with this” begins to feel like an affectation, a fetish for manual labor in a world with forklifts.

And here is where some of us get into trouble, because we can look at our own AI use and truthfully say: that is not what I am doing.

I use these systems less as answer machines than interlocutors. I bring them half-formed ideas, ask them to push back, test analogies, revise sentences, compare frames, and hold a thread while I decide whether the thread should be woven into fabric or simply cut. I use coding tools to inspect repositories I did not write, not so I can copy-paste a verdict, but so I can begin a dialogue with a large unfamiliar system. The model gives me a hypothesis; I test the hypothesis against the actual thing I have been asked to do. That is not the same as typing “answer?” into a fraction problem and stopping.

The fact that this distinction is real does not make me immune. It may instead explain why I am insulated.

I learned to read, write, revise, argue, listen, troubleshoot, and get bored before there was an always-on assistant waiting in the margin. I learned fractions well enough that I still sometimes think in them for fun. I made music with human drummers before drum machines and Logic Drummer became my practical companions. I know the difference between a pattern that gets the song recorded and a person in the room who changes the song by wanting something back.

That may sound like a flex, but it's a distinction I need to make.

A drum machine can help you finish a song. A good one can give you momentum, structure, a pocket, even surprises of a limited sort. But it does not bring a second intention into the room. It does not hear you lean into a chorus and decide to push. It does not misunderstand you in a productive way. It does not have its own Tuesday, its own hunger, its own impatience, its own private history of bad gigs and favorite fills. The machine can respond to you. The human drummer can alter you.

And to be fair, the human drummer can also flake out of a rehearsal, munge a fill, mooch all the drink tickets, and suffer bizarre gardening accidents. The better ones I've known drove the tour van masterfully, threw in the right joke at the right time when things got tense in the practice space, and got me to listen more carefully to how drummers occupy the landscape. Not everyone behind a kit can or should be a Jeff Porcaro, which makes me appreciate the good ones.

Large language models, at their best, are drum machines for thought. They can keep time. They can offer variations. They can make the blank page less blank and the room less empty. For those of us who lack nearby humans willing to discuss enterprise architecture, Warhammer-adjacent bureaucracy, AI governance, acid westerns, noise rock, and the metaphysics of provisioning checklists at 11 PM, the machine interlocutor is not trivial. It is useful. Sometimes it is genuinely enlivening.

But its usefulness has a seam. It is available when humans are not. It is patient when humans would be bored. It follows the thread I bring it, rarely demanding that I spend equal time on its thread, because it has no thread of its own. This is the feature and the warning in one. It is frictionless companionship for ideas. That does not make it fake. It makes it incomplete in exactly the way a drum machine is incomplete.

The danger, then, is not that AI assistance is always corrosive. The danger is that the forms of assistance that feel best in the moment may quietly train us away from the capacities that make the assistance useful in the first place. The student who has not yet built persistence can use AI to avoid building it. The worker who once had persistence can use AI to stop exercising it. The writer can mistake fluency for thought. The engineer can mistake plausible architecture prose for actual system understanding. The reader can stop after the AI Overview and still feel informed enough to proceed.

This is the boiling pot for suited frogs.

Some of us were issued protective gear before the heat was turned up. Not because we were wiser. Not because our schools were noble hazard courses intentionally designed to form resilient minds. Much of our friction was accidental, uneven, inefficient, and stupid. We learned through scarcity, inconvenience, bad interfaces, slow modems, indifferent libraries, and adults who did not always know what they were doing. The suit was assembled from whatever happened to be lying around.

Still, it was assembled.

That is why “brain rot for thee, not for me” is both tempting and wrong. My insulation is not evidence that the water is fine. My insulation and instrumentation are how I can still feel the temperature changing. The skills acquired under older conditions are not proof that AI poses no threat to learning. They are proof that prior friction produced something worth protecting.

This should change how we talk about education, work, and AI. The answer is not a theatrical return to chalk dust and toil. Friction by itself is not virtue. Plenty of old friction was just waste heat. But some friction was more than productive. Some difficulty was not an obstacle to learning but the medium in which learning occurred. The policy problem, the design problem, and the parenting problem are all versions of the same question: which frictions should be removed, and which must be deliberately preserved?

That question does not have a universal answer. A professional using AI to interrogate an unfamiliar codebase is not in the same position as a child learning fractions. A writer using a model to argue with a paragraph is not in the same position as a student asking for the paragraph to be written. A musician using Logic Drummer to sketch a song is not in the same position as someone who has never felt a human drummer pull the room sideways.

Developmental order matters. Ownership matters. The hazard course matters.

So yes, some of us are suited frogs. We can move around the pot with more confidence because we remember colder water. We can use the tools, benefit from the tools, even love the tools, while also recognizing that the tools change the environment in which future competence is formed. The suit buys us time. It does not make us invulnerable.

The practical discipline is not abstinence. It is calibration. Read past the overview. Do the problem before asking for the answer. Use the model for hints before solutions. Take the walk. Let the sentence arrive late. Find the human drummer when you can. Notice when convenience starts to feel like necessity. Notice when “efficient” becomes a synonym for “I no longer wish to struggle.”

The pot is warming. The suit is real. And instead of bragging about the suit, remember how you ended up wearing one.