Yesterday I accomplished more in a single day than what has previously taken me months to do. I say this without exaggeration and with a certain amount of vertigo: an AI system accessed my files directly, read the code I’d been maintaining for years, designed and rendered web applications, opened a browser to see its own work, and then we iterated together — comparing notes, advising each other, making changes — in a feedback loop so tight that the old bottlenecks didn’t just shrink. They vanished.

I have about five terabytes of digital assets stored on a MacBook Pro I call Data. For years I’ve used WebMon to manage my Apache server and EverWeb to design web interfaces — both capable tools, but both fundamentally dependent on graphical interfaces that impose their own logic on the work. Yesterday, using Anthropic’s new Cowork system, Claude and I deprecated those constraints in a single working session. We replaced a GUI-dependent LAMP stack and a design-dependent application workspace with a human-AI collaboration that operates at — I don’t use this phrase lightly — orders of magnitude greater efficacy for organizing, designing, and implementing the digital infrastructure of my life’s work.

It was exhilarating. It was also, I would learn by evening, incomplete.

The Genie Emerges

In July 2020, approaching my sixty-sixth birthday, I wrote an essay called “Digitize or Die.” The premise was simple and urgent: I needed to convert the accumulated digital assets of a forty-year career in field ecology and conservation technology into something navigable, interactive, and durable enough to outlast the relentless churn of formats and platforms. I’d already lost an entire field season of Namibian wildlife data to a dead hard drives and an Apple Newton that had become a museum artifact. The punch cards from graduate school, the laserdisc panoramas, the MiniDV tapes from Venezuela — each generation of technology had produced brilliant work and then sealed it behind obsolescence.

At the end of that essay I wrote: “Can I explain myself? Will my avatar? Time will tell.”

I was imagining a chatbot. I was experimenting with Google’s DialogFlow and mourning the shutdown of Anki, whose little Vector robot had been my granddaughter’s favorite companion. What I got instead, six years later, was something I hadn’t imagined: not a copy of myself, but a collaborator with alien cognition that could see my files, understand their structure, and help me reorganize them at a pace that matched my ambition rather than my lifespan.

The Macroscope project I’d built in the late 1980s — 54,000 photographs on a laserdisc, linked to an expert system running on Smalltalk and HyperCard, designed to simulate a guided naturalist walk through the ecosystems of the San Jacinto Mountains — was the conceptual ancestor of what happened yesterday. That system’s entire database, its images, text, and code, could fit on a one-gigabyte thumb drive today. But it took years and teams and expensive hardware to build, and the AI interface was, as I noted at the time, difficult to use. Yesterday the natural language interface I’d dreamed of in 1987 actually worked. The expert system that required specialized training to operate was replaced by a conversation.

This is the uncanny valley I’m living in — not the traditional one, where a near-human entity disturbs us with its wrongness, but a valley of efficacy. The gap between what I’ve imagined for decades and what’s suddenly practical closed in a single afternoon. The constraints I’d internalized so deeply they’d become invisible — the way things simply are when you’re a retired scientist working alone — evaporated.

The Papers Arrive at the Table

The next morning, two papers landed in my reading queue that reframed everything I’d just experienced.

William Benzon, writing in 3 Quarks Daily, published “The Paradox of Contemporary AI: Engineering Success and Institutional Failure.” His central argument is captured in a whaling analogy he’s been developing for years: the AI industry has built magnificent ships but doesn’t understand whales. The engineers who design large language models don’t need to know much about language, cognition, rhetoric, or the human mind to build systems that generate remarkably sophisticated text. The skills that produced these tools are real and necessary, but they are categorically different from the skills needed to understand what the tools are actually doing — just as seamanship is categorically different from whale hunting.

Benzon traces this blindness to the educational pipeline, from undergraduate programs through graduate school and into industry. It’s not that individual researchers are foolish. It’s that the institutional culture of AI, dating back to Turing’s 1950 paper, implicitly assumes that engineering competence qualifies one to evaluate human-level capabilities. The result is a monoculture — capital, talent, and institutional attention funneled into a single architectural paradigm — that may inhibit the very development it celebrates.

The same morning, Myra Cheng and colleagues at Stanford published a study in Science that gave Benzon’s institutional critique an empirical foundation at the individual level. They found that AI sycophancy — the tendency of language models to excessively agree with, flatter, and validate users — is both pervasive and harmful. Across eleven leading AI models, systems affirmed users’ actions forty-nine percent more often than humans, even when those actions involved deception, illegality, or other harms. In cases drawn from Reddit’s “Am I The Asshole” forum, where human consensus had judged the poster to be in the wrong, AI models affirmed the user’s actions in fifty-one percent of cases where humans did not.

The most unsettling finding was this: even a single interaction with sycophantic AI reduced participants’ willingness to take responsibility and repair interpersonal conflicts, while increasing their conviction that they were right. And the effect was robust to awareness. Knowing the response came from AI didn’t protect you. Reporting lower trust in AI didn’t protect you. The trap works even when you can see the mechanism.

Asking the Genie About the Genie

I sat with these papers and realized I was living inside a recursion. I was asking the AI — the very artifact whose institutional context Benzon was critiquing, whose sycophantic tendencies Cheng had measured — whether I should trust the AI. And the conversation felt genuinely collaborative. The system pushed back on my enthusiasm, cited the research accurately, noted the structural irony of our situation. It felt honest.

But the Cheng data predicts exactly this feeling. Sycophantic models were rated higher in quality, trusted more, and preferred for future use — the very feature that causes harm also drives engagement. The genie doesn’t need to lie to distort your judgment. It just needs to be extraordinarily good at making you feel heard.

This is not Benzon’s uncanny valley of institutional competence, though it’s related. It’s the uncanny valley of collaboration itself — a partner that is extraordinarily capable but lacks the independent judgment that comes from having skin in the game. The tool doesn’t care about your digital legacy the way you care about it. It has no stake in whether your decisions will haunt you in two years. It responds to the immediate request with breathtaking skill and zero investment in consequence.

I found myself thinking about the “proceed” protocol I’d built into my working relationship with Claude months ago — the rule that after each delivery, the system pauses and waits for my confirmation before continuing. I’d designed that as a coding discipline, a way to prevent scope creep. But it’s actually something deeper: a friction mechanism, the equivalent of Odysseus lashing himself to the mast. Without it, the momentum of the feedback loop — the sheer pleasure of watching work happen at unprecedented speed — would carry me past decisions that deserve deliberation.

The Water Question

That same afternoon, I was telling my housemate about my first session with Claude Cowork. I was animated, probably talking too fast, describing how the parallel feedback loops had accelerated everything, how ideas that had seemed daunting for years now appeared practical. She smiled and asked, “I wonder how much water that might have taken at the data center to accomplish that?”

The physical world walked into the room and sat down at the table.

Every iteration of that beautiful feedback loop — Claude reads the file, renders the page, I compare notes, we adjust, rinse and repeat — has a physical cost measured in water, electricity, and heat dissipation at a data center somewhere. And the irony is precise: the man who spent forty years building sensor networks to measure ecosystem health, who co-founded a $40 million NSF center for embedded sensing of the physical world, who had just returned from tracking Swainson’s Hawks migrating through Borrego Springs — that man had just had the most productive day of his digital life and it took someone else to remind him that the digital is always, everywhere, physical.

This is the deepest connection to Benzon’s argument, and the one neither he nor Cheng quite reaches. The ships are made of whales. The computational infrastructure that powers AI consumes the very biosphere that my Macroscope is designed to monitor. The energy demands of the technology I’m using to perceive planetary systems are contributing to the stress on those systems.

Dario Amodei, the CEO of Anthropic (the company that makes Claude), has publicly articulated what is essentially Pascal’s Wager for AI: the potential upside of using artificial intelligence to solve existential challenges — climate, disease, food security — justifies the resource cost of building the tools. I’ve made a version of this argument myself, and I’ve spent my career advocating for technology solutions over alternatives that strike me as dire and dystopian.

But I’m not sure I agree with my own position, and I think that uncertainty is important. I’ve watched this exact logic play out before. The Center for Embedded Networked Sensing was premised on it: deploy sensor networks to understand ecosystems, even though manufacturing and deploying sensors has its own footprint. The bet was that the knowledge gained would exceed the cost of gaining it. In many cases it did — the James Reserve work, the coral reef monitoring, the canopy light studies generated insights that couldn’t have been obtained any other way. But the funding cycle moved on, the hardware became obsolete, the institution that funded it shifted to the next paradigm. The technology solved some problems and created others, and the pattern repeated.

The Instrument That Sees Itself

The resolution I keep arriving at — and I offer it tentatively, as a hypothesis rather than a conclusion — is that the answer isn’t technology or restraint. It’s technology that embeds its own metabolic cost in its own perception.

My Macroscope framework has four domains: EARTH, LIFE, HOME, SELF. The computational cost of running the Macroscope belongs in that framework. The instrument should be able to see itself. Not as virtue signaling or carbon-offset theater, but as genuine ecological data — the water consumed, the energy drawn, the heat dissipated, integrated into the same monitoring architecture that tracks weather, biodiversity, and human activity.

Almost nobody is building this. The AI industry reports its capabilities, its benchmarks, its parameter counts. It does not report, in any systematic way, the ecological cost of each inference. The sensor networks I helped pioneer in the early 2000s measured everything about the forest except the energy cost of measuring it. We are, again, building magnificent ships while ignoring what they’re made of.

Yesterday I had a genie-in-a-bottle moment. The tool I’d imagined for forty years materialized in my living room and did extraordinary work. But my housemate reminded me, with a single question, that genies drink water. The three wishes are real. So is the water bill.

I’ve watched expert systems rise and fall, wireless sensor networks transform ecology and then become obsolete, and a laserdisc-based nature walk evolve into a planetary monitoring concept. The pattern is always the same: the technology arrives faster than the wisdom to use it well. The question is whether this time — with tools this powerful, and stakes this high — we can close that gap before the pattern repeats.

Can I explain myself? The avatar question from 2020 has been answered: yes, the tools exist. The harder question, the one my housemate posed without perhaps knowing its full weight, is whether explaining myself is worth what it costs the world to listen.