Attention Is All You Need

The email arrived at 4:47 on a Wednesday afternoon, just in time for all institutional crises to announce themselves in the territory as events too late to do anything about them today, but sufficiently early to gnaw at you throughout the night. The tone was upbeat. "Quick favour!" was a message of such confident punctuation that it betrayed a lifetime of experience never having been required to fulfil one.

“Quick Favour!”

I read it three times, brewed a pot of tea, read it again and then proceeded to do what any professional would do in response to a request that violated several laws of physics: I stared at the wall for 40 minutes and contemplated a career in horticulture.

The exclamation points achieved feats of extraordinary optimism in a context otherwise of fundamentally pessimistic expectations. They conveyed the full emotional impact of a request that actually represented a decision to communicate the most important technological transformation of our lifetimes within approximately the same time period as a long lunch and to make the experience highly attractive.

Finally, I closed my laptop, opened it again and briefly considered the possibility that the author of the request had actually read some of my materials for the full-day workshop or, alternatively, had absorbed knowledge about the existence of the workshop in a manner analogous to the way we receive information about the weather. In either case, we were assured that the presentation of highlights would certainly be adequate.

This is not the first such experience, nor will it be the last. It thus represents the essential character of my professional life: a continuing series of cheerful requests for which I must provide highly condensed versions of information that lose approximately 90% of their original meaning at the moment of compression.

The full-day workshop is highly successful. Everyone says so. Participants leave at 4:00 PM with an impression of dazed appreciation similar to that of a person who has spent 8 hours in an industrial cold-storage room and is just becoming aware of the condition of his feet. They have experienced something. They have been affected in some cases and have developed at least one new opinion in nearly all instances.

An hour's experience is not a workshop. It is a microwave meal that represents only a fraction of the eight hours required for slow braising with the original recipe. But you may refer to the final product as the same dish. In any case, it will provide no nutritional benefit to anyone.

Did the Manhattan Project Have a Taster Session?

I would like to ask you a question-and I hope you'll hear it, even from the back of the room.

Did the Manhattan Project have a tasting session?

Let me check. There was no 60-minute pre-conference Webinar titled "So you're curious about fission." There was no TLDR called "The implications of splitting the atom: A five-minute read for busy executives." And Mr. Oppenheimer certainly did not provide us with a "lite" version of the implications, complete with cocktails and an opportunity for discussion sponsored by Lockheed.

That was entirely appropriate.

I am not, of course, comparing artificial intelligence with a weapon. Rather, I am emphasising the magnitude of the technological disruption-the speed with which, and the extent to which, it alters every cognitive activity that previously contributed to the salaries of white-collar workers. Our experience with technologies of this magnitude indicates that we normally do not insist that individuals most responsible for understanding them receive an introduction to the technology during a single, routine, catered hour. Instead, we lose them, make them read, and assume that full comprehension requires more time than does complete digestion.

However, we have decided in 2026 that understanding of artificial intelligence should proceed at the same pace as a session of induction at a gym.

A Brief Taxonomy of What Actually Happens in the One-Hour Webinar

I've presented enough one-hour webinars on AI to develop a small but reliable taxonomy of events within them. Here it is as a public service. For the first 10 minutes, I am expected to solve the crisis of academic integrity. The audience believes with sincere conviction that I possess somewhere in my pedagogical repertoire a method for unambiguously detecting student work generated with AI, and am withholding it for reasons of professional discretion. I do not.

There is no such method. Our experience to date confirms unequivocally that humans are poor detectors of AI-generated writing, and that what we actually observe is a subset of students who are poor users of AI. This has been explained with extraordinary patience in a recent article in Inside Higher Ed by Justus and Janos. The result is an exercise in appearance of expertise. Let me make this clear.

The audience responds to my refusal to provide the magic prompt with the equivalent of a confiscated seatbelt. For the next 20 minutes, having accepted the absence of a shortcut to detection, the audience expects me to teach them to use the available tools. This seems reasonable, but in fact reflects an entirely different question than that which the audience actually intends to ask. What they really want is a prompt that will produce correct answers on the first attempt, under all conditions, without requiring any thought on the part of the student. The answer to this request is no.

It always has been no. And it will continue to be no in 2030 when whatever we may choose to call these devices will be incorporated into the kettle. Finally, for the last 20 minutes, and with the full understanding that I will not provide the magic prompt, the audience enters the phase of despair. This is where someone - and there is always one - asks whether AI is going to steal his job, and another person - also always present - asks whether AI is conscious. Still another asks whether he should be using ChatGPT or Copilot or Gemini, as if this were a question with a definitive answer that did not require three additional questions concerning what he actually intended to achieve.

To the best of my ability, I respond. I am perspiring slightly, and the Q&A goes beyond schedule. No one has learned anything. Everyone is terrified. This represents a failure of intervention and constitutes a real-fire fire drill.

On the Question “How Do I Use It?” Asked Sincerely, By Adults

There is a question I am asked, in some form, in roughly every professional conversation I have, and I want to address it here, in writing, so that I can stop addressing it in person.

“How do I use it?”

You ask it something.

That is the answer. That is the whole answer. There is not a more sophisticated answer hiding behind that one. There is no secret antechamber. You go to the chat box. You type a question. You read the response… and, if it is incorrect, inform it that it is incorrect. If the response is correct but unexciting, request it again in a different format. If the response is filled with em dashes and you have decided that em dashes constitute clear evidence of involvement with AI, congratulations -- you have correctly identified an em dash. Request that it produce a response without em dashes.

When you are unable to generate a question of your own, ask it to assist you in developing one. Request an explanation for your inability to generate a question and ask it to provide a 500-word personal description of why, in the year 2026, a fully matured adult with a postgraduate degree should be interacting with a system that responds to simple English and for which you should be seeking assistance from another person.

Read the response that you receive, express disagreement with it and, after having done this for approximately three hours over the course of two weeks, will have learned more about the principles of system operation than can be achieved with a single one-hour lecture on the subject. This experience will have provided you with the only method that produces complete fluency of operation with the system. But the secret is to ask it one question, and then ask it a different question. The related point made by Justus and Janos needs to be repeated: Many of the AI-classroom techniques earnestly recommended on teaching and learning websites (e.g., embedding hidden white text instructions within essay prompts, inquiring about events subsequent to the cutoff date for knowledge, and demanding "personal reflection" as if that precluded prompting) have failed for more than a year. This is because those providing the advice have never actually used the techniques. They learned a trick from a TikTok video in 2024 and continue to teach it with the diminished confidence of someone who knows, knows that this bizarre workaround will save them. It will not. And it was failing even before they became aware of it.

But you cannot learn this in an hour. You can learn it in approximately two weekends of curiosity, of which the institutional norm permits zero.

The Grown-Up Students

There is a population of persons at every institution where I have worked, and I have a private name for them that I am now going to share with others: adult students. These are my colleagues who are largely faculty members, with a few administrators. They are highly intelligent, accomplished and often excellent at their jobs.

But, with almost no exceptions, they also respond to the presence of generative artificial intelligence with the same polite concern, vague sense of obligation to do something, and absolute refusal to recognise that they are being blown toward the fire.

As a result, these will be the persons who do not attend an all-day workshop, who do not attend a one-hour introductory session, and who ultimately refuse to schedule a completely free individualised session. They simply await developments until, finally, they experience the forced but ungraceful adaptations that characterised the responses of previous generations to the Internet, to the mobile telephone and to Microsoft Teams.

Finally, they are also the persons who receive essays of extraordinary richness prepared entirely with ChatGPT and including all of the conversational remnants of the initial request. As a consequence, they inevitably confront me with the naive but genuine bewilderment of persons who are suddenly confronted with the ultimate consequence of the over-reliance on technology with which we have been training our students. The answer to their problem remains a workshop, and a very long one that has, in effect, been refused for the past two years.

It will be impossible for me to convey all of this in an one-hour introductory session. There will simply be insufficient time even for a brief introduction to myself.

The Part Where I Stop Being Gentle (Briefly)

I am afraid, but not for myself or for the institution. Rather, for people who are about to vote, care for children, make medical decisions and manage small businesses in an information world where the principal news source for the average citizen is a stranger delivering a 15-second monologue on TikTok over an automatically playing segment of Subway Surfers.

Because of that, I see with weary familiarity that the gap between those who can critically evaluate an output of artificial intelligence and those who cannot will become one of the major literacies of our century. And we are educating only one of these groups.

I am afraid, because those most susceptible to errors of artificial intelligence will be exactly the people who confidently tell me that they have no need for a training workshop and only require the highlights of what has been learned. The one-hour introductory session is not only inadequate, but actively dangerous because it produces the worst possible result: an illusion of competence without any real knowledge of the subject.

The individual who has spent zero hours with artificial intelligence will recognise his or her ignorance. The individual who has spent eight hours of training will have learned a great deal.

The individual who has spent only one hour of training will be convinced that he or she understands the field and will communicate that conviction forcefully in meetings.

I would prefer the zero-hours person. They are at least teachable.

A Modest Proposal: The Drop-In Therapy Support Group

In the spirit of innovation (which is, I am told, one of our institutional values), I have a counter-proposal.

I would not like to deliver another one-hour AI webinar. Instead, I would conduct a drop-in therapy support group. On Tuesdays at 4 p.m., bring your own complaint. Intended for educators who are tired of students misusing AI, for educators who are tired of colleagues denying the existence of AI, and for any adult student who has reached the point in the week at which it is appropriate to cry in a professional setting.

The meeting would be arranged in a circle of chairs, with tissues and a whiteboard on which would be written in permanent marker: "I cannot fix this for you."

Nothing would be accomplished, no products would be generated, and no highlights would be presented. Instead, participants would simply experience briefly the mutual awareness that we are all professionals attempting to achieve the impossible at a rate of institutional operations that has yet to recognise the existence of the impossible. I would provide biscuits, which would represent our only product. Everyone would feel valued, and rates of accomplishment would be unacceptable. Finally, we would all depart with a slightly greater sense of isolation.

This is indeed a serious proposal. I will submit it via appropriate channels, i.e., without regard for appropriate channels and in the manner traditionally used to accomplish useful work in regional Australian education.

Attention Is All You Need

The paper published in 2017 describing the transformer architecture is, in all likelihood, the most important paper of the entire modern era of artificial intelligence. Every chatbot to which you have ever spoken, every image generator with which you have amused yourself, and every helpful (or unhelpful or inadvertently racist) Microsoft product to which you have ever looked--all derive architecturally from that paper. The title reflects a technical observation about a particular type of mechanism for neural computation.

But over time, it has acquired an additional meaning: Attention is all you need. Pay attention to this, to the experience of living in the midst of the most profound transformation of working memory ever achieved. And pay attention to the fact that this transformation is receiving virtually no attention at most institutions, compared with the effort devoted to an upcoming software upgrade.

I will not provide a one-hour version of this experience. Nor do I want to make one. Instead, I will spend an entire day with people in a room and allow them to experience directly the nature of what is changing.

They will then return home and apply the tools of their experience, and come back to me one month later to again experience the slow, unglamorous work required to achieve competence with a new medium. This is exactly the way every previous generation has learned to use every new medium since the invention of the printing press. There is no taste for this experience, no highlight reel and certainly no request for a cheat sheet.

Attention is all you need. It is also, I’m sorry to say, the only thing you can give me. And it has to be more than an hour.

The email is still in my inbox. I have not yet replied.

I’m going to make more tea.