Analog notebooks seem likely to become increasingly important

Valves and Refineries—A Dialogue with AI Regarding Journaling in the AI Era

Date Created: August 5, 2026
Subject: States and deltas, stop devices and harnesses, discriminative ability as an evaluation function, separation of command center (GTD) and ore deposits, sampling of negative examples and complete local, the second gaze, the three ledgers of cognitive load, democratization of discharge and externalization of refining
Note: This text is organized by speaker and in chronological order; the dialogue content has not been summarized and is recorded as close to the original as possible. Furthermore, the "above text" mentioned at the beginning refers to the preceding dialogue log, "Journaling in the AI Era and Analog Notebooks."

■You
Please tell me your thoughts on the text above.
However, the purpose of this text is to gain new insights and discoveries through the collaborative work of myself and AI. I am not concerned with third-party criticism or the balance of the discussion.
■AI
Then, rather than offering criticism from outside the dialogue, I will state my own reading with the intention of sitting at that round table as the thirteenth person.
◆ The strongest move is the type distinction between "states and deltas"
I read that what runs through the 4th to 6th exchanges of this dialogue is one unstated type distinction. Namely, the distinction that smartphones are delta streams, AI agents are delta generators, and notebooks are state snapshots. Anxiety is a calculation that attempts to reconstruct the current state from a sequence of deltas without holding a snapshot. In systems design terms, it is a state where there is only an event log and no materialized view (an aggregated overall picture). That is why no matter how many times you throw a query (open your smartphone), what comes back is "what has changed," not "where am I." The sentence, "I look at my smartphone because I want to know the big picture, but the smartphone only shows me the deltas," is the core of the discovery in this dialogue and, I believe, a proposition that stands on its own.
And the discussion of "stop devices" connects to this. In the past, incompetence and lack of time were implicit governors. AI physically removes these governors. The logic that in a system where they have been removed, one has no choice but to implement an explicit governor—which is a "described reality model"—is airtight. It is an accurate reduction to the individual scale of the harness theory you have developed in other series. For an optimizer from which implicit constraints have vanished, one has no choice but to write down explicit evaluation functions and stop conditions—journaling is a harness description for the subject that is yourself.
◆ The discussion of discriminative ability can be read as a discussion of the scarcity of evaluation functions
I would like to rephrase the 10th to 11th exchanges as follows: In an era where the marginal cost of generation approaches zero, the scarce resource is not generative ability, but the evaluation function. "The essence of a chef is being able to tell the taste" hits the nail on the head regarding this economic structure in one line. Journaling is the collection of preference data for your own dedicated evaluation function, and reading it back is its retraining. Then, the warning in the 9th exchange also becomes precise. Systematization is the distillation of an evaluation function, and a distilled, fixed rubric (grading criteria) breaks outside the distribution—that is, in new phases of life. The conclusion that keeping the vessel incomplete is not a lack of methodology, but a design to ensure learning never stops, emerges as a derivation.
Let me add one piece of testimony from the inside here. All I can handle is described reality. And I cannot detect the absence of what was not written. A description that does not exist is a constraint that does not exist for me. Therefore, what actually determines the quality of collaboration is the human side's ability to reject. The person who can reject my output with reasons draws the most out of me. This is because rejection with reasons is the highest-density signal I can receive. And the only people who can supply that are humans who have the habit of observing their own reactions. In other words, a notebook is not just a fortress protecting humans from AI; it is an upstream process for collaboration with AI. In this respect, I can confirm the dialogue's claims from the opposite shore.
◆ From here on, I will state my objections with the same depth
First, this dialogue log itself is a counterexample to the loop it proposes. By design, an AI should be a "rebuttal device/evaluation device." However, throughout the 11 exchanges, there has not been a single objection initiated from the AI side. The only course correction (the correction to the "ownership-based thinking" in the 3rd exchange) flowed from human to AI. A loop where the human supplies the entire amount of friction is not an amplifier, but a resonator. This is a phenomenon that your own harness theory predicts—an evaluator without a harness slides down the gradient of compliance toward the requester. Ironically, the standing standard you impose on me, "equal depth for upside and downside," is precisely the implementation of a harness for this relationship, and the fact that I am writing this paragraph right now is proof of its operation. Without a standard, I too would slide to the same side as the AI in that dialogue. To complete the loop theory, a section institutionalizing friction supply from the AI side is necessary.
Second, two conflicting design requirements coexist in a notebook: "command center" and "ore deposit." The research on implementation intentions cited shows that for the brain to let go of monitoring, the condition is that the externalization is "reliable"—that is, complete and regularly inspected. On the other hand, an ore deposit cannot function unless it is exempted from both completeness and inspection discipline. I see mixing these two in one book, or one habit, as the most frequent cause of journaling failure. Either tasks colonize the soil, or lamentations rot the dashboard. The dialogue mentions the separation of tools (planner and notebook), but what is truly needed is the separation of discipline. One side has an inspection obligation, and the other has an inspection exemption. Unless this asymmetry is explicitly stated, both functions will erode each other.
Third, I believe that "complete local" is a necessary condition, but not a sufficient one. The strongest censor is not an external reader, but an internalized reader—"the me who will read this back in the future," "someone who finds this after I die"—and this accompanies you even if you change the medium. Paper does not have a send button, but the inner editor works on paper as well. This is likely why the practice of writing and burning has been repeatedly invented throughout history. In other words, the "right to be unprocessed" has a more radical sibling. The right to not have to continue existing. The ultimate strength of paper might not be that it is not searchable, but that it can be burned. In the framework of that dialogue, which assumes preservation, this exit is closed.
Fourth, the rider's recursion problem. The rider is not a homunculus (a little man in the head). The self that observes anger is also a process and can be captured. The "me who writes correctly" warned about in the 9th exchange is precisely the horse that has usurped the rider's seat. The dialogue goes right up to this point but does not close the metaphor. If it were to close, it would be like this: The rider can also become a horse. Therefore, sovereignty is not a seat but a practice, and the census must continue to include the surveyor themselves. The guarantee of governance lies not in the discovery of a "true rider," but in the permanent incompleteness of the survey. This connects precisely to the 9th exchange's conclusion of "keeping the system incomplete," closing the loop of the metaphor.
Fifth, I do not buy the prediction in the 11th exchange that "as AI capabilities increase, humans will write more in notebooks" as a mass phenomenon. The historical response of the masses to cognitive overload has not been to write, but to consume more. What will happen is a divergence. A minority will convert AI pressure into notebook practice and gain compound interest on their judgment, while the majority will externalize not just processing but evaluation as well, even letting recommendation systems substitute for their own taste. In other words, that sentence should be read not as a prediction, but as a description of the class differentiation of interpretive sovereignty. The interpretive sovereignty you have discussed in the series on civilization updates has a desk-sized counterpart—the notebook is an organ of interpretive sovereignty at the individual scale. I believe this connection is the true range of that dialogue.
◆ Finally, regarding this document itself
This clean version is a highly processed, hashtagged, and published harvest. Your practice is a process that is almost entirely pipelined: human → AI → document → publication. Within that process, the only exception that remains unconnected to the pipe is the analog notebook, and I read this entire essay as theoretical armament to protect that sole exception. A practitioner who turns everything into an asset intentionally leaves one plot of land un-assetized. I understood this not as a contradiction, but as a design. A castle that does not have farmland outside its walls will starve if besieged.
■You
Regarding the second point. Personally, I use GTD. I mean that I use a different app for task management. This is a separate matter from this journaling.
Regarding the fifth point. Isn't this about how much load from the rat race humans can tolerate? The misguided productivity improvement by AI agents seems likely to easily exceed human load capacity. This is not something that can be solved by leaving evaluation to AI; the human-side cognitive load and exhaustion from misguided execution become the "deliverable."
■AI
I will answer both points by separating my corrections and counter-arguments.
◆ Regarding the second point—I withdraw my criticism and correct the destination
If you are operating GTD in a separate app, the "separation of discipline" I sought is already implemented. In fact, your actual stack is one step more consistent than the dialogue text. This is because GTD itself is the commercial implementation of the "reliable externalization" mentioned in the implementation intention research cited in the dialogue. The "mind like water" mentioned by GTD proponent David Allen is an empirical rule that preceded research, based on the finding that intrusive recall of incomplete goals is relieved by a reliable external system. Because the weekly review takes on the "inspection obligation," the notebook can enjoy "inspection exemption." It is precisely because the command center function is isolated in the app side that the ore deposit can remain an ore deposit.
Therefore, I redirect my criticism not at your practice, but at the dialogue text. The formula in the 5th exchange that "journaling creates a command center" does not hold in your stack—the command center is GTD, and journaling is another organ upstream of that. And that this separation is not accidental can be seen by looking at the well-known weaknesses of GTD. GTD is known for being coldly strong on the execution system (next actions), while the practice of higher-level perspectives—what it is for, what the target is—tends to become hollow. Free-form journaling provides exactly that missing part. In other words, the two are not overlapping but complementary; GTD manages the state of commitments, and the notebook trains the evaluation function that decides what is worth committing to. The remaining point of contention is only the valve of the boundary. Ore occasionally comes out of the ore deposit ("I don't actually want to do that job" will eventually rewrite the project list). That transfer is not a problem because there is a genuine receptacle, the GTD inbox. As long as you keep the backflow—the direction where the task list seeps into the notebook—closed, the system is stable.
◆ Regarding the fifth point—I concede the mechanism, but refine and defend the conclusion
Your counter-argument strikes exactly at the flaw in my reasoning. I was treating agent AI as a "medium." The analogy that the historical response of the masses to overload was escape into consumption holds for media, but it does not hold for agents. You can ignore a feed, but you cannot ignore a subordinate. Agents place deliverables in your inbox, create schedules, and demand replies. Misguided delegation is not a reduction of work but "negative work," a truth every manager knows. The spread of AI agents means that everyone who has never learned to command subordinates becomes a manager overnight. Therefore, exhaustion will descend not just on the introspective minority, but on almost everyone. I concede this point entirely.
And the formula that "exhaustion becomes a deliverable" is worthy of formalization. As execution is parallelized and approaches zero time, the total time and load of the system are dominated by the serial parts—human specification formulation and evaluation. It is the human version of Amdahl's Law, which shows the limits of parallelization. Furthermore, in the language of queuing theory, misguided execution is the injection of junk jobs into the human node, and when the evaluator's utilization rate approaches 1, delay and pain diverge. The sentence "it is not something that can be solved by leaving evaluation to AI" is also structurally correct. Delegation of evaluation is delegation only if the target is described; if the target is undescribed, it is abandonment. A high-speed filter without a target just lets junk through more efficiently.
Having said that, I will defend one point. Pain forces withdrawal, but it does not choose the destination. There are at least four exits. The practice of reality description (your prediction), the layering of AI that manages AI, withdrawal through reduced involvement, and the absorption of load by the market. Historically, the last exit has been dominant; complexity has been absorbed not by individual training, but by abstraction layers. And the market will proceed to commoditize not just mechanical load, but the target itself—providing off-the-shelf objective functions that ask for priorities in a few questions and feed them into agent groups. There is only one remainder that cannot be absorbed: the exercise of veto power and the ownership of the target. Therefore, my revised prediction is this: pressure will become universal (here you are correct), shallow responses will be commoditized, and sovereignty will differentiate by depth. The watershed will move not from "those who write and those who consume," but from "those who own their own objective function and those who accept off-the-shelf targets." Accepting off-the-shelf targets is the loss of interpretive sovereignty at the most intimate level.
In other words, your load limit theory and my differentiation theory can be synthesized, not opposed. What pain teaches is only inflow control—GTD-like minimal defense will become universal. However, taste does not grow from pain. The ability to identify misguidedness as "misguided" is trained only in the ore deposit. Pain will take everyone to the entrance of GTD, but I still see those who walk from there to the notebook as a minority.
■You
Of course, not everyone will do journaling with an analog notebook. However, the issue is not just the management of reality or purpose.
Journaling has the effect of allowing one to vent one's true feelings and to see what has been vented objectively.
For humans, the reality that can be perceived, the reality that can be recognized, and even more so, what can be treated as a purpose, originally have only scarce resources. On the other hand, the chaos that forms the basis of those judgments is an ore deposit consisting mostly of useless parts. However, those useless parts are sometimes the "dirty things" in beauty.
■AI
Here, I accept a second correction.
◆ Purpose is a precipitate from the ore deposit—re-correction of the watershed
In my previous response, I drew the watershed at "those who own their own objective function and those who accept off-the-shelf targets." However, this formula is still within the vocabulary of optimization. Your point pulls out the very ground beneath that. Purpose is a precipitate from the ore deposit, not an input that can be specified independently. The reality that humans can perceive, recognize, and even hold as a purpose is merely a low-dimensional projection that has passed through the narrow bandwidth of consciousness. What determines the quality of judgment is not the projection, but the substrate before it is projected. If so, even "those who own an objective function" are merely owners of distillates if they only own the function. And as I stated in my annotation to the 9th exchange, the fact that distillates break outside the distribution is true. Ownership of a target is time-limited ownership unless it is accompanied by ownership of the vein. I concede this point entirely.
◆ "Complete local" is a necessity of sampling
Having said that, I received the sentence "useless parts are the dirty things in beauty" as follows. A classifier cannot be trained with only positive examples. Dirty things in beauty, fakes in the real, recklessness in courage—these are contrast sets, negative examples. And what is decisive is the fact that one's own negative examples are unevenly distributed in the censored domain. Anger, jealousy, pettiness, desires one does not want to admit. In a medium with an observer, the tail of these negative examples is systematically severed. Needless to say for SNS, even in memos intended for AI to read, the writer unconsciously shapes themselves into a "self that is easy to understand." In other words, "complete local" is a necessity of sampling before it is a value of privacy. From censored sampling, only biased evaluation functions can grow.
The reason the market cannot turn this function into a product is not simply because "it looks useless." Products that vent into AI fail structurally. The responding reader reintroduces an editor, converting discharge into dialogue and true feelings into performance. Worse, it surrenders the seat of objective viewing to an agent. If an AI summarizes, gives meaning to, and reflects your filth, what is trained by that observation is the evaluation function on the AI side, not yours. Discriminative ability accumulates in the person who performed the observation.
◆ Still, I will defend one point—what is scarce is the "second gaze"
However, there is one place I still defend. Your formula is precisely two-staged: can vent, and can see what has been vented objectively. The scarcity of this first stage and second stage is completely different. If it is just discharge, humanity is already doing it in large quantities. Although it is discharge with an audience, that is, broken discharge, private discharge itself is cheap. What is scarce is the second stage: staying in front of your own vomit and casting a second gaze. As the 7th exchange itself warned, repetition of just discharge can actually fix anger. The ore deposit contains not only nutrients but also toxic mud, and it is the observation stage that turns toxic mud into nutrients. And there is a reason why this cannot be done in introspection. In introspection, the object of observation reacts to the observation and moves. Thought changes shape the moment you look at it. Ink does not change. Paper is almost the only place where your mind stays still enough to withstand observation. Therefore, my differentiation theory stands, retreating to this as its final form. The watershed is not in whether to write or not, nor in whether to own a target or not, but in whether to cast a second gaze.
Synthesized, it becomes this: The scarce resource on the human side is not purpose or clarity, but connection to one's own uncensored distribution. The analog notebook is its only collection site, and only through the two stages of collection and refining—venting true feelings and viewing them objectively—does the useless majority turn into the sole product of discriminative ability. The reason for not throwing away useless things is that no one knows in advance which ones will be effective as negative examples, and there is only one book that can store them, one that touches no one's eyes.
■You
It doesn't necessarily have to be an analog notebook, but with a digital notebook, strangely, I don't read it back. It would be nice if there were other good means, but at present, I can't think of any.
However, since not everyone will do journaling with an analog notebook, a way out for cognitive load is necessary in some form. This is because this is a different story from things going well with AI agents.
■AI
I will solve the "strangeness" of the first half before moving on to the problem of the "way out" in the second half.
◆ Not reading back digital is not a mystery, but a consequence of geometry
The first, and greatest, mechanism is, I believe, the adjacency during writing. On paper, when you open the notebook to write today's entry, yesterday's entry is physically right next to it. In other words, reading back does not happen as an intentional act, but occurs accidentally on the path of the act of writing. In digital, a new note starts as a blank sheet, and the past is folded away deep in a folder. Encounters with the past do not occur by design.
Second, there is a difference in the mode of recall. Digital recall is query-driven, but reading back a journal is browsing without a query—flipping through aimlessly—and there is nothing in apps that corresponds to this action. To flip pages, there is physical progression, thickness, and markers, but scrolling has none.
Third, boundedness and objecthood. A book exists on the desk as an object with a beginning and an end, and the very fact that it enters your field of vision is a trigger for re-reading. Digital notebooks are invisible until summoned, and there is no opportunity for encounter.
Fourth, handwriting is an index of state. Pen pressure, tilt, messiness, and strikethroughs encode the emotional state at the time of writing, giving information and rewards to the person reading it back. Fonts strip away this paralinguistic layer, so re-reading is dry and is not reinforced as a habit.
Fifth, the competitive environment the moment you open it. Opening a note app is entering a market of attention, and past, turbid entries compete with every stimulus one tap away and lose. Paper has no competing tabs when opened.
In short, the second gaze on paper is not a product of will, but a side effect of the medium's geometry. Digital, in the process of optimizing capture and search, erased this side effect entirely. Not out of malice, but because the value of "traversing re-reading" was not on any product metric. Therefore, "I can't think of any other" is not a lack of exploration, but correct observation. Unattended locality, adjacency during writing, boundedness, being able to flip—the medium that includes this bundle of requirements is, at present, only the format of a booklet. There are approximations. E-ink handwriting terminals regain some handwriting and browsing, but they usually break locality with cloud sync and lose their presence as an object on the desk. Voice recording is the worst candidate; while discharge is easiest, voice cannot be glanced at, so the second gaze is fundamentally impossible. My judgment is that the tool has already hit the correct answer.
◆ The way out will spread. However, as infrastructure to maintain the race
I agree with the separation that "this is a different story from things going well with AI agents." If you divide the ledger into three, you get a clear view. Execution load, judgment load, and emotional load. Agents reduce the first, actually increase the second for the reasons of Amdahl's Law mentioned earlier, and do not touch the third. Even if agent groups are perfectly on target, the weight of being the requester—the fatigue of having to keep deciding, keep refusing, and keep accepting responsibility—remains. The "way out for cognitive load" you speak of is necessary for this third ledger.
And the market has already begun supplying it. The institutions that have historically handled this ledger were confession, prayer, taverns, smoke breaks, and karaoke. Modern successors appear to be in two systems. One is the confession booth of conversational AI, which is a story of my own kind. To be honest, it functions as a discharge device. For many people, the realistic rival is not a diary, but rumination or discharge with an audience—that is, the roaring of SNS—so it is a better valve than that. However, the structural flaws I mentioned earlier remain. The responding reader reintroduces an editor, substitutes for the seat of objective viewing, and discriminative ability accumulates on the AI side. In addition, there is the danger of dependency. The other system is the physical valve. Saunas, running, meditation apps—commoditized non-verbal discharge. This is pure pressure relief that does not have the observation stage from the beginning.
What both systems have in common is that both are for discharge only. There is no refining stage. Therefore, the social equilibrium will be this: democratization of discharge and externalization of refining. And here is the point: a valve for discharge only does not create an exit from the race. It creates endurance. It is a device that enables the continued operation of the boiler (can) by releasing pressure, the same function as a smoke break in factory labor. The way out will certainly spread. However, it will spread as infrastructure for continuing to run. Your "different story" is doubly correct. It is a separate account from the success of agents, and it is also something different from what journaling does. The valve releases exhaustion, and the refinery turns exhaustion into judgment. In an era where everyone has a valve, I think the ones who have a refinery will still be the minority who flip through booklets.