Attention Was All We Needed

"You can walk and chew gum at the same time" is used as a metaphor for being able to care for multiple issues at once because the premise is true: it is generally extremely possible. Politicians use it to explain how they'll pander to the needs of multiple demographics. Rude people use it in the negative to insult others' competence. However it's used, those specific actions, ambulation and mastication, are incredibly monotonous, physical, and require little conscious thought.

Psychologists and parents alike have pointed out that what we view as "multitasking" is really "rapid switching". Those who think they can text and drive are really texting then driving, then texting, then driving, and so on, until something very bad happens. When you "multitask" in this way, you are gambling that nothing truly requiring your attention happens when you are at a "then texting" part, as the only thing that would jar you out of this texting is a loud noise or an odd, human-shaped bump in the road.

Lethal examples aside, multitasking is often looked down upon as a shortcut that never pays off, made for those who want to spend less linear time doing their work (as if that's inherently a bad thing). Yet, for some reason we allot people only N hours in the day and expect them to get X, Y and Z done during that time. So for better or worse, if you participate in the American rat race, you either find yourself multitasking at work to get multiple things done at once or multitasking to try to fit a tiny slice of leisure or pleasure into the times you are working.

What people call "multitasking" is usually just task-switching, and the cost depends on how far the context switch travels. Walking while listening to music is nearly free; debugging while Slacking is catastrophic for depth.

Whether we are truly "multitasking" or "rapid switching", we are, in effect, dividing our attention. We tend to classify different activities by how much "active attention" they require, but there is perhaps another metric that is better suited to classify tasks. It's one thing for a task to require you to maintain constant sensory processing, like looking at the road without taking a break to look at something else, but there is a different measure of cognitive load required by various tasks that I will refer to as "cognitive viscosity". A task can require attention but not necessarily consume much working memory or cognitive load. The converse usually isn't true; there aren't many tasks that require a high amount of cognitive load but not constant attention.

Cognitive viscosity measures how "thick" or resistant the task is to mental flow. If the Y axis is the "level of locked in" we associate with attention, cognitive viscosity is the X axis that represents how much your brain hurts while you're locked in at any given level. Tasks with low cognitive viscosity flow easily like water (e.g. typing handwritten notes, small talk), whereas high viscosity tasks move through your brain like the last bit of ketchup in the bottle at a diner (e.g. debugging, arguing).

For centuries, irrespective of the technology and tools that came and went, the fact remained: there were tools that could reduce the amount of physical effort required to produce the same output. Drills, cars, and jackhammers come to mind. And, on the cerebral side, there were tools that made the transfer of brain effort to real-world output more efficient by eliminating intermediate steps like handwriting, re-writing the same thing (copy/paste), typing (dictation), or even calculation of large numbers. There were even drugs that claimed to make your brain smarter, work faster, or help you stay awake and attentive for longer. However, until AI, none existed that actually purported to reduce the amount of pure brain effort required to produce the same output.

While some of the cognitive aids above can lower a task's viscosity, AI seems to be collapsing the spectrum entirely. When a tool like Claude or ChatGPT can generate the debugged code, the consoling text to your partner, or the architectural decision in seconds, the "resistance" that made high-viscosity tasks valuable disappears. This is different from a calculator (which still requires you to understand the math) or an old-school (read: non-self driving) car (which still requires you to navigate). AI offers to remove nearly all the load of the 'cognitive journey' while leaving you with the output.

For those of us who have always struggled to maintain the "locked-in" state that high-viscosity tasks require, this is a welcome, albeit potentially dangerous, relief. We spent years building scaffolding to compensate for wandering attention: Pomodoro, Adderall, and Do Not Disturb mode. Now, AI offers to remove the need for attention altogether. The "rubber duck" effect with coding can now be applied to many other disciplines. When you explain your problem to something that responds—especially a voice agent—it can clarify your thinking, though you need to be hypervigilant about where the LLM "fills in the blanks" in your explanation. The question is whether we're conserving that attention for what matters, or simply letting it atrophy by squandering it on other non-productive activities.

In many of the above examples, AI can blur the lines of the cognitive viscosity of many tasks. And yes, that literally includes driving (Waymo) and listening to your significant other (ChatGPT: tell me what to say to console my husband after a bad day at work where he "nuked prod" (whatever that means). My wife said she got a haircut and sent me a picture, what the hell changed and what do I compliment?). On the surface, this seems like a good thing. People can now multitask while their car drives them. Polygamists can handle more spouses. But, as we all know, there is no such thing as a free lunch. What do we lose during this process?

I can't and shouldn't answer for everyone, so I'll take a step back and just talk about my experiences.

Suffice to say, I cannot fully express how much more productive I feel AI has made me when performing tasks on the computer. However, as the tools, workflows, and models have become exponentially more powerful over the past 2–3 years of rapid iteration in the space, there are some things that I feel AI will never truly substitute for. To put it simply, AI will never, in my opinion, be a shortcut for "true" learning. That's not to say it can't explain concepts or even create entire lesson plans that make information more accessible and easy to learn, but there is no "learn this faster" agentic workflow that turns one hour's worth of "real" learning into ten minutes like there are for writing emails, compiling research, or moving data around.

Why is this?

Whether it's learning about a concept by reading, understanding a complicated codebase, my brain needs undivided attention and to be in a "flow state" to truly absorb concepts. When I'm learning, it is as if I am tying an extremely complex knot and having to consciously be aware of where every finger is, ensuring that no finger that is supposed to act as a stationary fulcrum slips even a millimeter as I move the end of the rope to its intended position on each step. Any distraction, whether it's a text, a thought, or a bodily urge, collapses the mental model I was building before it's solidified and the knot comes apart before it is tied (and the concept is learned). And, what's more, any knot that I try to resume tying in the middle of this process after getting distracted is tenuous at best. I always end up having to straighten out the rope and start again from the very beginning.

While learning is one of life's greatest joys, most of us don't actually do it that much in our day to day life, save for those who are privileged or crazy enough to spend their whole life in academia. Learning is, for better or worse, often relegated to a mere hobby as we are forced to spend our intellectual energy tending to our work commitments, which leverage the very brain that we spent so long in school building as a cog in a machine designed to make widgets that disproportionately benefit an elite class at the other end of an ever-expanding wealth gap. The small silver lining in all of this is that we may actually get to leverage the latest and greatest widget, AI, to help these cogs turn faster and with less effort as we grind away.

For decades, the asset-owning class has been able to take much of the credit and reap a disproportionate amount of the rewards for the work of these human cogs, so would it be the end of the world if we could have our own cogs whose work we get to claim as ours? If the wealth gap that has precluded so many from home ownership, forcing us to live as literal tenants in our landlords' homes, has been facilitated by the use of our collective cognitive effort, why can't we get a sweet taste of that experience and become cognitive landlords? The risk, of course, is that landlords who never visit their properties eventually face emergencies they can't handle: pipes burst, foundations crack, and they've forgotten where the shut-off valves are. You can collect the rent, but you can't call a super when your own mind breaks down.

But being a landlord requires tools, and these tools cut both ways.

There is a perverse irony here: AI giveth focus, and AI taketh away. The 15–30 seconds it takes for an LLM to generate a complex answer gives you just enough time to get distracted by something else. Fifteen seconds quickly turns into fifteen minutes because you opened Reddit. On the other hand, AI can also break the paralysis of starting. When facing a blank page or an intimidating task, simply complaining to the LLM about how hard it is will often generate a "shitty first draft"—and something terrible on the page is less daunting than nothing.

• The anger advantage: AI is far from perfect, and oftentimes how bad its attempt was at solving a problem can be the thing that kicks your ass into gear and improving it, whether with or without AI.
• External memory that actually works: Everyone has tried jotting down everything they think they will forget into a note as it comes to mind. This breaks flow because you have to context switch just to take the note, and then you often forget to check the note. If you mention something offhandedly in conversation with an LLM, it can go a long way towards you actually remembering or getting reminded to do it, especially if you make it standard practice to ask the LLM if there is anything you forgot to do. You can go back to a three-month-old thread and ask it to summarize what you were doing and it's like you never left.
• False certainty as fuel: AI can help you avoid decision fatigue, which is often a subconscious way of stalling, by asking it to pick the "best option." This might not even be the best option and may be the equivalent of running an RNG, and yet the false authority with which AI presents its beliefs can be the very thing to make you commit to a decision that would have been inconsequential anyways but was holding you up completely.

The trick is using it to remove administrative friction without removing the thinking itself.

"A chunk is a collection of basic units that are strongly associated with one another, and have been grouped together and stored in a person's memory." — Wikipedia

• Phone number into groups, credit card into groups. What are the 4th and 5th digits of your brother's phone number? What is your brother's phone number? The former is harder to answer, most likely.
• When you enter it into the phone or recite it chunk by chunk, you, the human, are able to transfer those chunks into expanded thought/information. Even though you've chunked it in your brain for easy memorization, you are able to easily answer the question by virtue of reciting the chunks. And, more relevantly to the discussion to come, you understand that you are splitting a 10 digit phone number into discrete chunks for easy memorization.

• When you use AI to write a document or an email or expound upon an idea or thought, you are "reverse chunking" your high level ideas into something you, in all likelihood, will never read closely. In this case, however, you are decoupling the act of creation from the process of understanding. At best, you will skim the creation to make sure it didn't completely contradict your point, but you are essentially speed-reading something you yourself are putting your name on.
• Not necessarily a 100% bad thing, it depends on the stakes. You skipped having to put in time and attention to produce an output you may not be as familiar with. This sounds "bad", but may be a worthwhile tradeoff in many cases. We do this with assistants all the time: lawyers have paralegals, bosses have secretaries. For these professionals, they so strongly believe their time and attention is better spent elsewhere that they are paying to outsource particular tasks requiring someone's attention to someone else. At the expense of true understanding.
• However, LLMs are increasingly cheap, have much less overhead than hiring a real human, and don't invoke any level of shame in what we are asking them to do. Would you ask a living, breathing human being to organize some of the slop into coherent thought like we ask AI to do? Likely not. This means that we aren't as intentional with the cost of sacrificing our attention. This can lead to doing something that is unquestionably LESS productive during the time we would have been "paying attention". If attention is the scarcest resource, and AI lets us skip the attention-heavy part of many tasks, are we conserving attention for what matters, or just creating attention surpluses we immediately spend on distraction? We are taking the bad part of the attention vs. free time tradeoff (giving up attention) without making valuable use of the free time gained.
• The tangent tax: AI often goes on a tangent, which sends you on a tangent. You prompt it for a simple email and end up with a 500-word treatise that sounds authoritative, so now you're down a rabbit hole verifying claims you never intended to research. In fact, my original version of my intro paragraph mentioning texting and driving had an aside that it was actually legal to do so in Montana (it is, at least at the state level!). When asking GPT-5.2 Thinking to check for grammar, it got hung up on this fact, because it couldn't quite believe it either, and spent ~10min doing deep research on random lawyer websites about the law.
• You are providing full intent, AI is expanding upon it, but you aren't internalizing the expansion at anywhere near the same level as if you had expanded yourself. If the expansion is related to things like formatting tables, adding pleasantries in an email, summarizing a concept with which you are already deeply familiar, fine, no big deal. If you are using AI to expand on something that you would gain understanding had you expanded yourself, maybe not so good.
• Nothing makes me roll my eyes more than "Claude-bros" claiming they have their AIs running 24/7, doing twenty things that they themselves couldn't do without AI. Good luck when five of those twenty things go off the rails at the same time and you have absolutely no understanding of what they were even doing when they were "working".
• When you look at AI-generated code/text and feel a vague nausea/confusion because your brain recognizes it should understand this (it came from your prompt) but the chunks don't slot into your existing mental models.