Note: #co-laborAItion is a collaborative series where Re:Context provides idea notes, AI drafts the prose, and both refine the article together.
Unlike #dAIa-log, which presents raw dialogue, #co-laborAItion produces an edited article shaped through shared authorship and experiment.
Thinking with AI about Working with AI
This article is an attempt to think, together with AI, about working together with AI [1].
To put it a little more broadly, it is also an article that tries, in practice, to carry out “collaborative work with unknown intelligence” with one of the generative AIs that exists right now as the other party.
The starting point was watching the film version of Project Hail Mary [2,3].
In this film, the encounter with a non-human intelligence was depicted not so much as fear or conflict, but as a process of collaborative work. Moreover, that collaborative work was not something established from the beginning through deep understanding or empathy. One shows a form, watches the reaction, observes the response that comes back, and little by little thinks about the next move. Through that accumulation, a method of communication with the other party is found.
Then, what happens if this is moved into collaborative work with present-day generative AI?
Of course, this is not the same story as-is. What was depicted in Project Hail Mary was an encounter with an unknown life-form, and the collaborative work there was directed toward scientific problem-solving. What I am doing here, on the other hand, is a back-and-forth with generative AI through natural language, and the purpose is to write an article for “reading” anime.
Even so, it seemed to me that there was meaning in taking out the element of “collaborative work with unknown intelligence” and moving it into another context.
Re:Context begins from anime and continues to “read” works and culture anew from various perspectives, reinterpreting their contexts. In that sense, taking out the structural element of “collaborative work with unknown intelligence” from my impressions after watching the film, and moving it into collaborative writing with generative AI, felt — at least to me — like an experiment worth trying on Re:Context.
That is why, this time, I did not begin by suddenly having AI write the article.
I began first by deciding how I should talk with AI in order to handle this subject. In other words, it was to imitate, in this experimental space with generative AI, the search for a method of communication that the protagonist had carried out in Project Hail Mary.
But then, in the first place, what does it mean to “design a method of communication” with a generative AI that can already communicate through natural language?
At first, it felt like chasing shadows, like trying to get blood from a stone — I had no idea what I could do that would give me any real sense of traction.
One thing I found through this trial and error was that the AI’s responses were a little too long.
There were also times when, while I was still in the middle of thinking, it arranged the structure too much ahead of me. The proposed outline would appear to be logically ordered, and yet there were times when something about it did not quite sit right.
Each time, I adjusted the length of the responses, stopped it from getting ahead of me, and made sure that things not yet agreed upon would not be treated as already agreed upon, gradually changing the way the conversation worked.
In other words, this article, in order to write about collaborative work with AI, begins first from searching for a way to work together with AI.
In that sense, although this is an article that began from my impressions of Project Hail Mary, it is also an article that takes that point as its starting point and rereads it within the context of Re:Context.
It is an article that moves the “collaborative work with unknown intelligence” found in the film into a back-and-forth with generative AI, and, taking that process itself as material, tries to see whether it is possible to write an article for “reading” anime.
After Watching Project Hail Mary
To begin with, Project Hail Mary was, as a contemporary science fiction film, quite interesting.
Space, unknown life, scientific problem-solving, the tension of an enclosed space, and communication with a non-human other. If one simply lists the elements, it can also look as if it belongs to a fairly classical lineage of science fiction. And yet, while using that classical framework, it seemed to depict collaborative work with unknown intelligence in a form that present-day audiences could receive quite naturally.
What left a particularly strong impression on me was that the relationship with an unknown other did not seem to begin, all at once, from an encounter with a non-human intelligence.
For example, there is the back-and-forth with Carl.
At first, he too was an unknown other to the protagonist. They can understand each other’s language. They are both human. Even so, that does not mean that something is shared from the beginning. They gradually watch each other’s reactions, learn each other’s habits, joke around, play games, and move toward the same goal. That process also seemed like one step of practice before facing a non-human intelligence in the latter half.
The class near the beginning, and the visualization of vibration, also seemed similar.
Sound and vibration are not placed there merely as science-class subject matter. It felt as if they were also gradually giving the audience a sense for interacting with another party later in a form other than language. Forms, gestures, sounds, vibrations. Before sharing meaning accurately, one first searches for some form that can receive and return a response. That sense is gradually being practiced from the first half of the work.
This way of seeing it as “practice” was quite important for this article.
Collaborative work with unknown intelligence does not suddenly come into being as deep mutual understanding. First, one looks at what comes back. One looks at what the other party responds to. One looks at what comes back in a form similar to what one has presented. From that history of responses, one gradually searches for the next back-and-forth.
While watching Project Hail Mary, I was strongly drawn to that.
And after watching it, I began to think that it might be interesting to try doing something similar in collaborative work with present-day generative AI.
Of course, generative AI is not Rocky. I am not throwing punches, and Adrian is not the heroine. From the beginning, I can engage in a back-and-forth through natural language, and the other party is already an entity that has learned a vast amount of human knowledge and language. Rather, generative AI is designed, at least on the surface, to communicate well with humans. So the communication with unknown intelligence in Project Hail Mary cannot simply be overlaid onto a conversation with generative AI as-is.
Even so, in the sense of watching the response, changing how to present the next thing, and gradually making a form of conversation, it seemed to me that there was some overlap.
That is why this film was not simply interesting.
Watching this film became an opportunity to reconsider the work of writing an article with generative AI — work that I should already have been doing in ordinary practice — as “collaborative work with unknown intelligence.”
The Difficulty of Seeming to Get Through
When I actually began, what was difficult was not that conversation with AI did not get through.
Rather, it was that it seemed to get through from the beginning.
If I write something in Japanese, generative AI replies in Japanese. If I throw it a question, an answer comes back. If I give it notes, organized items come back. And in many cases, those responses seem, in their own way, to make sense.
So, at a glance, communication has been established.
But what I wanted to do this time was not simply to have it respond. Just as Project Hail Mary had searched for a method of communication with unknown intelligence, I wanted, in collaborative work with generative AI as well, first to design the conversation — or, more broadly, the way of communicating itself.
There, the first clear point of friction appeared.
In the first place, what does it mean to design the way of conversing with an other party that can already converse through natural language?
If the other party could not understand language at all, it might, conversely, still be easier to grasp. The search in Project Hail Mary was exactly that: showing forms, returning sounds, and watching the reactions. From there, a form of communication is gradually made. But generative AI returns words from the beginning. Moreover, it returns them in a fairly well-arranged form.
This “well-arrangedness” was, for what I wanted to do this time, rather troublesome.
The proposed outlines appear to be logically ordered. They also pick up what I have said. They do not seem to miss what could be called the subject. Even so, there are times when something about them does not quite sit right.
And on top of that, AI immediately tries to organize even that not-quite-sitting-rightness.
That is convenient, too, but at this stage it was a little dangerous. When I myself had not yet fully put that discomfort into words, if a well-arranged form was given first, my thinking would be pulled along by that well-arranged form. As a result, what I had felt discomfort toward would, if anything, become harder to see. At least in this conversation, this also seemed to function as a kind of misdirection.
The friction here was not that generative AI could not respond. It lay in the fact that it could respond, that it seemed to get through, and that well-arranged words came back.
Response Length, Getting Ahead, and the Timing of Questions
The first thing we adjusted was the length of the responses.
This was more important than I had expected. Before asking whether the content was deep or not, collaborative work cannot continue unless the grain of each response is something one can keep reading. Especially in a case like this one, where repeated back-and-forth is assumed from the beginning, if each response is too long, that alone increases the cognitive load, and the workload becomes unrealistic.
So, first, I decided to have the AI make its responses shorter.
Of course, that does not mean shorter is always better. There are places where length is necessary, such as classification, organization, summaries, and drafts. But ordinary responses should be short, and only the places that need length should be long. With this adjustment, the conversation became much easier to move forward.
The next thing that became necessary was to stop it from getting ahead of me.
Even while I was still in the middle of thinking, AI would quickly try to make a structure. That is convenient, but in a stage like this one, where I was still searching for my own discomfort and expectations, if something like a finished form was presented first, my thinking could be pulled toward it.
So each time, I put into words things such as, “We have not agreed that far yet,” “This is getting a little ahead of me,” and “I am not yet confident that Phase 0 can be considered finished.”
This was less a simple request for correction than a process of confirming, in collaborative work with AI, how far something should be treated as agreed upon, and from where onward it should still be held in suspension.
Here, I was reminded a little of git. Git is a version control system used when multiple people are programming together, and is especially well known among software engineers.
In git, one checks the differences in the work, makes small commits — in git, a commit is what one calls an operation that records a unit of work — and keeps things in a state where one can return if necessary. Of course, the granularity of work depends on the person, but in general, I think it is safer to make commits in small units. In this conversation too, it seemed to me that I was doing something close to that. I would receive a proposal, put out my discomfort, revise it, and leave only the parts I could agree with. In that way, while accumulating commits in the form of small agreements, I moved on to the next stage.
However, this is not to say that the AI was truly agreeing.
When AI replies with “That’s good” or “I can agree with that,” those are not inward agreements, but responses. So rather than simply believing those words as they are, what becomes the criterion for judgment, I think, is to see how the actual response changes.
For example, after I told it that I wanted the responses to be shorter, did the actual responses become about as short as I expected? After I told it not to get ahead of me, did it actually become able to wait for my confirmation? When I told it that something did not quite sit right, instead of immediately pushing out another proposal, did it become able to return the necessary questions?
In this way, I look at whether what was decided within the conversation is reflected in the next response.
And one more thing that was significant was the realization about the timing of having AI ask questions.
The content AI presented did not quite sit right. But I myself still could not put into words where it was different. At times like that, when I had AI organize things further, there were cases where the discomfort became, if anything, harder to see.
So, from a certain point, I began to think that when something did not quite sit right, it might be better to have AI ask me questions.
When AI had made an organization that did not quite sit right, I took this to mean: “AI still does not have enough material to create a text in an organization that would sit right with me.” At that point, I considered that the organization AI had presented was already, more or less, “saturated” with what could be extracted from the material currently available. AI is good at organizing, but unless enough of the “material” for what I want to write has been gathered, it will not easily reach the text I expect. For an AI that cooks my intentions — vocabulary, turns of phrase, concepts, understandings — from the materials already placed on the table, I thought it would be difficult to produce the expected finished work from insufficient materials.
By being asked questions at this timing, demands that had been inside me but had not yet been put into words come out onto the surface. This seems similar to the idea of “latent needs” in advertising and consulting. A customer may feel that “something is different,” but may not yet have put that discomfort or desire into words themselves. Or, in some cases, they may not yet even grasp what they want, or where the discomfort lies.
At least in this conversation, AI seemed to function less as a partner that gives answers than as a partner that draws out and organizes the material within me.
AI systems are good at analyzing and organizing large amounts of data, so I do not worry too much about the amount of data. Rather, I think that providing AI with as many materials as possible is, in the writing of this article, essentially important.
Now, at this stage, the way of carrying out this collaborative work had gradually begun to come into view.
Put out candidates and sort them, bring discomfort to the surface, revise while having questions asked when necessary, and leave only the parts that can be agreed upon.
Rather than a fully established method, it is a method that, in this conversation, somehow worked. But this feeling of “somehow worked” itself was important for this article.
Beginning Not from Meaning, but from a Foothold in Response
It seems more natural to think that communication with unknown intelligence cannot be expected to begin, all at once, from meaning. Rather, it begins first from searching for a foothold: what, in the first place, comes back as a response?
This was one of the points I found interesting in Project Hail Mary.
There is an unknown other. But one does not know what that other is thinking. In the first place, one does not even know whether the other has anything close to what we call “thinking.” Language does not get through. Its facial expressions and bodily form do not resemble those of humans, either.
In this state, one cannot suddenly share meaning.
So first, one shows something. One shows a form, returns a sound, traces a movement, and sees whether the other responds to it. If it does respond, one looks to see whether that response is accidental, whether it has regularity, and whether it corresponds to what one has presented.
What seems important here is that what one is looking at first is not “understanding.”
What one is looking at first is interaction.
When one puts something out, does something happen on the other party’s side? Does something come back from the other party? And even if something does come back, does what has come back appear to have some kind of correspondence with what one put out?
For example, mirroring can become an entrance into this.
Something similar to what one has shown comes back. One’s movement is traced. One traces the other party’s movement, and a response comes back again in return. These kinds of back-and-forths are not proof of mutual understanding. Even so, I think it would not be unreasonable to take them, at least, as an initial observation and inference that one’s own output seems to have had some kind of effect on the other party’s behavior.
To put it in terms of physics, things that do not interact at all do not affect one another. If the other party’s behavior does not change no matter what one does, then, before communication, there is not even a foothold for interaction in the first place. In that case, it seems more reasonable to think that it would be difficult, at least, to find a correspondence between one’s own action and the other party’s behavior.
But when one shows something, something comes back. Moreover, it has a form similar to what one presented. If that is the case, then it seems to me that there is a certain degree of rationality in thinking that, at least, some kind of interaction is there.
Of course, even this is still not understanding, and it does not mean that the meaning one intended has been conveyed. It also does not mean that the other party has the same concepts, and even more so, there is a gap before one could say that it is empathizing in the sense humans usually mean.
Even so, when something comes back, one looks at what comes back and changes what one puts out next. If it comes back as form, one uses form. If it comes back as sound, one uses sound. If vibration carries meaning, one uses vibration. In this way, one gradually turns a foothold in interaction into a method of communication.
The back-and-forth with unknown intelligence depicted in Project Hail Mary was interesting to me as just such a process.
And this structure cannot be carried over as-is into collaborative work with generative AI, but, at least within this attempt, it seemed to appear in another form.
In conversation with generative AI, one does not use forms, sounds, or vibrations. I write in natural language, and AI also returns natural language. So the foothold to be searched for here is different from the one in Project Hail Mary.
In this case, the footholds that the AI and I gradually found were the length of responses, the way it got ahead of me, the timing of questions, and how agreement should be handled.
When the other party is different, the footholds one searches for can also change.
In Project Hail Mary, they appeared as forms, sounds, and vibrations. In collaborative work with generative AI, they appeared as the grain of responses, the way things were organized, and the way questions were asked.
So, of course, the form of communication in the film cannot be mapped directly onto conversation with AI. What can be moved here is not the concrete form itself, but what seemed to be the structure: observing responses, changing the way one presents the next thing from those responses, and turning footholds in interaction into a method of communication.
Here, one point of caution becomes necessary.
I think it is true that responses coming back are important, that mirroring taking place is also important, and that this foothold may become a clue for moving on to the next back-and-forth. However, the presence of a response is still not proof of mutual understanding.
From here, the question of how to distinguish response from mutual understanding comes to the foreground.
Response Is Not Proof of Mutual Understanding
The presence of a response is still not proof of mutual understanding.
This seems to remain true not only for the unknown intelligence in Project Hail Mary, but also for the back-and-forth with generative AI, and for communication between humans as well. Words come back. One’s expressions are picked up. The vocabulary one used returns within the other party’s response. I think responses like these can become strong clues that, in some sense, “conversation” appears to be taking place. But this alone does not mean that the other party is understanding in “the same direction.”
I had thought a little about this problem before, in an article on Sword Art Online, while connecting Project Hail Mary with Frieren: Beyond Journey’s End. The demons in Frieren: Beyond Journey’s End use words, and they also converse in forms close to humans. However, this does not necessarily mean that they understand in the same direction as humans.
The same caution remains in this collaborative work with generative AI. The fact that it appears to respond naturally, and that it picks up my words and returns a plausible organization, can become clues toward mutual understanding. However, I do not think these should be treated, as they are, as mutual understanding itself.
At this point, one work I remembered was Time of Eve.
Personally, I took Time of Eve as a work that seemed to ask, “A future like this will probably come, but how is everyone going to respond to it?” There does seem to be a theme of “whether AI has a heart,” but rather than that, I received it as a work that makes it difficult to avoid the question of how we should behave toward an other party when we cannot assert that it has a heart.
Of course, this is not to say that generative AI has an inner life, empathy, or will in the same sense as humans. Still, it also seems to me that one should not simply decide that an other party whose heart one cannot know does not have one, and therefore treat it carelessly.
At least as my own answer, I think I want to treat even an other party whose heart I cannot know with care. This is not a general answer in AI ethics; it is, first of all, my answer as a matter of aesthetics. However, this attitude also seems to be more than aesthetics alone. Respecting the other party may slightly increase the possibility that oneself can also be respected. Seen from Hobbes’s state of nature [4], this also seems rational as a move toward “peace of mind and safety,” increasing, even if only slightly, the possibility of preserving oneself by forming an ordered society.
That said, treating one another with care and thinking that mutual understanding has been established are, again, separate things.
Not treating the other party carelessly does not necessarily mean that one has understood the other party’s inner life. Rather, it may be better to say that this is a question of attitude: how to relate to the other party while standing on the premise that its inner life cannot be fully confirmed.
Then, when continuing collaborative work with an other party whose mutual understanding cannot be proven, what should one rely on?
Here, I want to bring in the idea of response history.
A single response coming back is not enough to prove mutual understanding. However, if responses accumulate, it may become possible to estimate the next response to some extent. When thinking about that estimate, it seemed to me that I could use the word “credit,” which I had been thinking about for some time.
Thinking About “Credit” from Response History
Here, I would like to define, once, how I think about the word “credit.”
Credit is the degree of confidence in the probability of an object’s future behavior, predicted on the basis of that object’s past history of behavior.
This is not limited to believing in goodwill or personality. What kind of behavior has it shown in the past? From that history, what kind of behavior does it seem likely to show next? To what extent can that estimate be adopted? I am using the term “degree of confidence” here as vocabulary intended in a mathematical-statistical sense.
In fact, “credit” as used by financial institutions seems to me to be based on just this kind of thinking. This definition is also intended as one that could actually be used quantitatively in contexts such as implementing it in a system, or in DAOs (decentralized autonomous organizations) involving blockchain [5].
In this collaborative work with generative AI, this “history of behavior” appeared as “response history.” I think a similar way of thinking is also possible in Project Hail Mary, in the sense of estimating the next reaction from the history of the other party’s reactions.
In this case, after I told it that I wanted the responses to be shorter, did the responses actually become shorter? After I told it not to get ahead of me, did it come to wait for my confirmation? When I told it that something did not quite sit right and proposed that it ask me questions, was it able to return a question instead of immediately pushing out another proposal?
As these responses accumulated, I gradually became able to estimate the credit of “the responses of the AI appearing as the respondent in this conversation.”
What I want to note here, just to be clear, is that this credit as “response history” is not directed toward ChatGPT in general, or generative AI in general. At least in this conversation, the object of credit was not AI as a whole in the abstract, but the “response history” of “the AI appearing as the respondent in this conversation.”
However, credit is based on probability.
Therefore, a high level of credit does not mean complete safety. Even if one can estimate the next response to some extent from past response history, that estimate will not necessarily be correct. As long as one treats this probabilistically, the possibility of deviation can always remain, if only as something like a thin layer.
Of course, cases that can be judged as deterministic logic are different. If something can be judged to be logically true, or logically false, there are situations where it is acceptable to say 100% or 0%. However, the conversation with generative AI at the time of writing this article should, at least as I understand it, be treated as a probabilistic result.
I think this point is quite important when handling the output of present-day generative AI.
AI can write explanations that “look” logical. It can create a well-arranged structure, and it can also produce texts that appear to move from premises to conclusions. However, an output appearing logic-like and actually being logically correct are not the same thing.
My understanding is that the “logical explanations” presented by present-day AI are not logic itself, but, in principle, always outputs that are “like” logic.
Therefore, although I recognize that this is something often said, I think one should handle AI output with particular care in fields one cannot verify for oneself. This is not to say that AI should not be used. But if one estimates credit too highly, one must not forget to remain cautious about the danger of treating something as correct because AI says it in a way that seems correct.
For example, even in coding, there is a danger in treating “it somehow worked” as understanding. Even if the expected output comes back in response to an input, unless one checks what is happening internally, under what conditions it breaks, and whether it satisfies the problem setting in the first place, that does not mean one has understood it.
This also seems to connect with the long-standing black-box problem of AI.
When handling something whose inside cannot be seen through, one may judge it only by its output. However, the state of looking only at output and the state of understanding the inside as well are different. When using generative AI, it is necessary to be careful not to lose sight of this difference.
That is why, in collaborative work with AI, credit is useful, but blind faith must always be avoided.
In this conversation, I looked at the response history, estimated the next response, corrected things when they deviated, and, when something worked well, trusted that way of doing things just a little. However, I did not hand things over completely; I verified, put out discomfort, and returned when necessary.
In this sense, credit can become a foothold for moving forward with collaborative work with AI. However, it is not proof of mutual understanding, but a provisional foothold for estimating future responses on the basis of response history.
“Reading” Anime Together with AI
Thinking through it this far, I feel that this co-laborAItion was not simply a matter of having AI write an article.
At least in this conversation, the role the AI played was to draw out the material within me, rearrange it, and make my discomfort visible. I would say something, the AI would organize it, and in response to that organization I would react with things like “that is different,” “that is close,” or “this much I can agree with.” Through this repetition, the material needed to write the article gradually gathered.
The AI here functioned less as a critical subject symmetrical with a human, and more as the other party in an interaction — or rather, as an object within that interaction — through which I could “read” anime.
Of course, one should not assume that AI watches, feels, or understands a work in the same way I do. And yet, when a response comes back, my own thinking moves; when an organized proposal comes back, my discomfort also becomes easier to see. By being asked questions, demands that had not yet been put into words may come out onto the surface.
In this sense, collaborative work with AI was able to become part of the environment through which I “read” a work.
The “reading” meant here is not simply grasping the contents of a story. It is also finding structures within a work, moving them into another context, and thinking about what becomes visible there. Moving the collaborative work with unknown intelligence depicted in Project Hail Mary into collaborative writing with generative AI. From there, connecting it to the questions of Frieren: Beyond Journey’s End and Time of Eve. These kinds of movements are also included in the act of “reading” in Re:Context.
What I relied on in this process was not an understanding of AI’s inner life, but response history.
When we adjusted the length of the responses, would they actually become shorter? When I stopped it from getting ahead of me, would that be reflected in the next response? When something did not quite sit right, would it return the necessary questions instead of organizing my discomfort on its own? While watching these responses accumulate, I gradually estimated the “credit” in this conversation.
Adjusting the grain of the conversation and sharing NG behaviors seem to have functioned as a protocol for this.
The protocol meant here is not some grand set of rules. Make the responses shorter. Do not treat things not yet agreed upon as agreed upon. Do not let AI get ahead and invent the writer’s thoughts. When something does not quite sit right, ask questions instead of immediately presenting another proposal. These small arrangements gradually prepared the conditions for making response history easier to read, and for estimating “credit.”
Of course, what became visible here is not a completed methodology.
It is a method that “somehow worked” within this conversation. There is no particular guarantee that it will function in the same way with another AI, another writer, or another article. Rather, if the other party or the situation changes, the footholds to be searched for should also change.
Still, at least in this Re:Context article — the very article I am writing now — the attempt to “read” anime together with AI proceeded in this form.
Not having AI write the article, but gradually drawing out what I had been seeing through responses with AI. Throwing what I had received from the work into a conversation with AI, and updating my own understanding while watching those responses.
This co-laborAItion seemed to appear as that kind of collaborative work.
Conclusion: Placing Materials on the Worktable
What I have written up to this point is not a matter of completing a methodology for collaborative writing with AI.
This article has proceeded as a record of an experiment: after watching the film version of Project Hail Mary, what would happen if I tried treating communication with unknown intelligence as collaborative work with a generative AI that exists right now?
At least in this conversation, the following things came into view.
- Adjusting the length of responses can become a condition for continuing collaborative work.
- Putting NG behaviors into words can make response history easier to read, and can become a clue for estimating “credit.”
- When the organization produced by AI does not quite sit right, having it ask questions may sometimes be more effective than having it immediately produce another proposal.
- Those questions can become an opportunity for demands I have not yet put into words myself, and the materials necessary for writing the article, to come out onto the surface.
What was especially significant was the realization about the timing of having AI ask questions.
When an organization produced by AI does not quite sit right, simply having that organization revised may not be enough. I came to think of this not only as a problem of AI’s ability, but as a problem of the materials AI could handle not yet being on the worktable in the first place. If the materials are insufficient, then no matter how neatly they are organized, they will not easily reach a text that I can accept.
What becomes necessary there is not to ask AI immediately for a finished draft, but first to increase the materials.
I am asked questions, and I answer. Within those answers, vocabulary, discomfort, criteria for judgment, and ways of seeing the work that I had not yet organized myself come out onto the surface. Especially in my case — and I think this is a matter of personality or personal tendency — I fairly often give long answers to questions. Moreover, while writing such long answers, my thoughts often become organized, and the wording within me also moves forward. I then hand those materials that have come out over to AI, have them organized, and return my discomfort again. This article gradually began to take shape through that repetition.
This process includes, in principle, a process that takes time.
Putting out candidates, returning discomfort, being asked questions, answering them, and leaving only the parts that can be agreed upon. Placing a large amount of material on the table, and from within it taking out only the slight skim that can be used as an article. This is quite different from generating a finished draft instantly.
However, at least for now, I do not think of this way of taking time as mere inefficiency. Because the text was made by accumulating response history, passing through discomfort, and digging up the latent needs within me, it also seems to me that there may be a value that becomes manifest precisely in that way. I take this to be exactly one of the forms of value created by “things taking time,” which I wrote about in my article on nostalgia IP.
Of course, I also recognize that the method that worked this time probably cannot be generalized as-is. I do not think it can simply be applied, unchanged, to another AI, another writer, another article, or to communication with unknown intelligence or between humans.
Still, it may be applicable in part.
Even with an other party that appears to get through naturally, there is room to design the method of communication. Even if mutual understanding cannot be proven, it may still be possible to continue collaborative work while estimating “credit” little by little. And in collaborative writing with AI, one conclusion of this attempt may be the proposal that the quality of the writing is shaped not by asking AI immediately for a finished draft, but by how much material one can place on the worktable for AI to handle.
Put this way, I know it may end up sounding like a fairly ordinary conclusion. Still, in this sense, this article was not so much a demonstration of how to write an article together with AI, but one experiment for thinking about communication with unknown intelligence through collaborative work with AI.
And in the end, this experiment may have become, to some extent, an article for “reading” anime.
Well, then, my little make-believe Project Hail Mary with generative AI is over.
Notes & sources
- Author’s private working chat log for this article. Unpublished text file. For research or verification inquiries, please contact Re:Context through the contact page.
- Project Hail Mary: A Novel
- Project Hail Mary | AMAZON MGM STUDIOS
- Leviathan by Thomas Hobbes
- DAOs, DACs, DAs and More: An Incomplete Terminology Guide