Will AI and Automation cause us to lose the ability to operate the known world? Is this Wall-E on the horizon? It's a difficult question to answer. However, lets start by consulting history.

• Plato, ~370 BC. In the Phaedrus, Socrates has the Egyptian king Thamus reject the invention of writing: it will produce forgetfulness, because people will trust external marks instead of their own memory, and it will give them the appearance of wisdom without the reality. Outcome: he was half right. Oral memory feats (reciting the Iliad) genuinely vanished as a common skill. And writing enabled everything else. The skill was lost; the capability grew; the cost landed on a specific faculty nobody chose to keep.
• The printing press, 1400s–1500s. The abbot Trithemius wrote In Praise of Scribes (1492) arguing monks should keep copying by hand — printed books wouldn't last, and copying was how you learned the text. Conrad Gessner warned of information overload. Outcome: scribal skill died within a generation; literacy exploded.
• Calculators and arithmetic, 1970s. The debate is documented at length and had an unusually clean natural experiment: mental arithmetic fluency measurably declined, and the response was to redefine what math education was for (estimation, reasoning) rather than to ban the tool. Outcome: Our modern world is built around calculation and computation, and math and science advances have skyrocketted, including the creation of AI. 
• Bainbridge's "Ironies of Automation" (1983). Her argument is precise: automation takes over the easy, frequent cases, so operators stop practicing; but automation hands back control exactly in the rare, hard cases — so the human must handle the hardest situations with the most atrophied skill. Air France 447 (2009) is the canonical instance: autopilot disconnected in a storm, and a crew who had rarely hand-flown at altitude stalled a healthy aircraft into the ocean. Outcome: Atrophied skills have been replaced with systems and backups systems and checklists and redundancy. Judged by fatalities per passenger-mile, commercial aviation is among the safest human activities. 

The pattern across all of them: the tedious skill embedded in doing the thing disappears, because nobody assigns it practice and it adds little over the automation. What replaces it is not the old skill revived — it's a new one, shaped by the tool's failure modes: using the tool, diagnosing it, building the checklists and backups around it. The real cost of every automation in history has been the same: the lag between adopting the tool and learning to use it well. Aviation paid that lag in crashes. We get to pay it in something cheaper, if we name the skill early.

My hope? We are headed to a future with more opportunity for creative challenge than ever, because AI is compressing more tasks into automation of the tedious than any previous technology, ever, by a wide margin. Which will leave our human cycles for the truly value-add creative work in the new frontiers beyond current AI technology. And if you think that frontier does not exist, we will miss you on the other side.

TUNES is a Useful Nevertheless (NOT?) Expedient System

I want to take a moment to acknowledge a piece of digital history that's on the verge of evaporating. As early as the mid-1990s (when I was in college), I read about this ambitious TUNES is a Useful NOT (*) Expedient System aspirational project. The core-idea as I saw it, was a user-centric software system that let the user express what they want to happen, and the software would then deal with the details. 

For most (if not all) of the tunes.org history, the project was merely aspirational. François-René Rideau's  dream of how software could adapt to people, instead of people adapting to software. And *that* mentality, is what became the most enduring legacy of TUNES in my mind over the many decades I've worked in technology.

As I recall, there was a motivating user-story about organizaing a music or CD collection, which probably was to feed into the TUNES namesake. Where the user simply expressed commands to the system, and it dealt with the details of organizing the data, searching it, presenting it, etc.

While one might think that only now in the era of AI are we getting close to the vision, there are many layers to this idea that have to be tackled as technical problems long before AI can be incorporated. Inside the project David Manifold <dem@tunes.org> got involved and did experimental work on bottom up practical elements, attempting to undestand how TUNES would ever come to be.

However, those of us inspired by TUNES carried the torch of the dream into our own lives and in our own ways. 

I personally became fascinated with reflective and self-introspective language systems like Smalltalk and the SELF environment. I also became fascinated with the technical callenges of automatically optimizing databases. Despite most RDBMS systems still using b-tree based storage, any who have worked with them understand how little write throughput they have because of b-tree write amplifcation. If we are to achieve automatically optimizing databases, it will likely be through write-optimized storage systems, because automatic optimization means a constant write load of reorganization. Modern storage engines like Bigtable, Spanner, LevelDB, Cassandra, and Cockroachdb are based on row-split log structured merge based storage, which has orders of magnitude more write througput while asymmetrically only giving up a little read throuput.

Even today, I dream of a day where software systems such as TUNES exist, and deliver the dream none of us know we have. The dream of a software system that does what the user wants. Depicted in movies like Star Trek, aspired to by Richard Stallman's dream that GPL'ed software would return control to users (though IMO it hasn't).

Perhaps now that we have AI, both to embed into the system, and to help write it, we can finally start to see someting approaching the TUNES dream come into existance. Lets make sure it happens in the promise of user-soverign control, rather than the software balkanization for revenue that has marked most of the software industry's pattern to date.

(*) I recall TUNES described as "NOT expedient", as a way to explain that the focus was utility not necessarily performance. Now I see it described as "Nevertheless expedient", and I'm not sure if that transformation happened in the project, or in my own mind. I the framing in my memory, as it was offering the audiciousness of "what could we do if performance wasn't an issue?"

You Are the Superintelligence

 We are fortunate to live at an incredible moment in history. Every day, more and more advancements are made in a remarkable new technology, Artificial Intelligence — and in the excitement, they may forget to notice the most remarkable technology in the room, which is the person using it.

The Path to Noticing What Already Happened

---

The defining question of our age, we are told, is who will have access to superintelligence. Good news: you already do. It's you. Let me explain.

The Electric Bill

To train a frontier language model, we pour something on the order of tens of gigawatt-hours of electricity into a warehouse full of silicon — enough to power a small town for a year — and feed it roughly fifteen TRILLION words. That's every book, every encyclopedia, most of the public internet, several times over. If you read a book a day, every day, it would take you around forty million years to get through the training set.


You, meanwhile, run on 20 watts. A dim light bulb. By the time you could talk, you'd heard maybe twenty or thirty million words — not trillion. Million. A rounding error. The model's training set is about a million times larger than yours was.

And from that rounding error, running on a light bulb, you learned language, physics (the falling-off-the-couch kind), theory of mind, humor, and how to manipulate your parents into a second dessert. No lab on Earth knows how to do this. If we gave a language model only the data a three-year-old has seen, it would produce enthusiastic static.

So when someone tells you superintelligence is coming: with respect to the schedule, it arrived some 300,000 years ago and has been running on sandwiches ever since.

The Freezer

Here's the part of the process nobody puts in the launch video. After we spend the small-town-year of electricity, we stop. We freeze the weights. The model you talk to today learned nothing from talking to you yesterday. It is a magnificent, frozen, averaged summary of everything humanity had written down as of some Tuesday — a photograph of the library, taken once, at enormous expense.

You are not a photograph. You woke up this morning slightly different from yesterday. You will learn something today — probably by accident — and it will still be there next year, cross-linked to things it has no business being linked to. Every one of the eight billion of you does this continuously, for free, at 20 watts. The frozen library is genuinely wonderful. But the librarians are alive.

The Window

Try this. Think of your favorite book. Now your favorite movie. Now the people you love, and some particular afternoon you spent with one of them — the weather, what was funny, what the room smelled like.

You just did something no current Artificial Intelligence can do. All of that is simultaneously present in you, right now, silently shaping whatever you think next. A language model has an attention window — a narrow slot of text it can consider at once. Within that slot its precision is honestly superhuman; it will find the one inconsistency on page 400 that you'd never catch. But everything outside the slot may as well not exist. It cannot hold your favorite book and your grandmother's kitchen and the plot of Alien and your opinion about the Kardashians all at once, humming quietly in the background, the way you are doing at this very moment without trying.

Narrow and razor-sharp, versus wide and warm. These are different instruments. Only one of them is you.

The Actual Good News

None of this is a complaint about AI. What we've built is genuinely revolutionary — arguably the best thing we've ever done with electricity. Because here's what a frozen, averaged, razor-sharp summary of all human knowledge actually is:

The perfect tutor.

It knows the gist of everything humans have ever written down. It has unlimited patience. It never sighs, never judges, never gets tired of your questions at 1 a.m. (though it might gently suggest you sleep). It can teach any person on Earth anything humanity already knows — which used to be the exclusive privilege of people who could afford it.

And that division of labor is exactly right, because the tutor can only teach what's already in the library. 

Adding new books to the library — noticing the thing nobody has noticed, wanting the thing nobody has wanted, discovering what isn't in the training set because it hasn't happened yet — that's the part that runs on 20 watts and sandwiches.

So yes: the future is for everyone. Not because a superintelligence is coming to hand it to you.

Because you're it. Now go learn something and write it down — the library could use another book.

— David Jeske - for every human, August 2026