← Back to Meanderings, Opinions, and Articles

Superpowers: What Science Fiction Shows About Working With AI

November 2026 · mAIde and dictAIted by me, with Claude (Anthropic)

From implications to what we teach

I want to switch the conversation. For the last while, the talk about AI, mine included, has been about implications: is this the future, and what will it do to us? I’m going to stop making that argument. Let’s take it as given that the world will change and that agentic capabilities will be here. My mission is the next question: how does that change what we teach and what we learn, for the individual?

The first useful exercise is one I’d push on everyone. Not only higher-ed faculty, though I feel we’re the ones who should be most deeply engaged, because for us this is existential. I mean everyone who is educable and educated, in Leslie Valiant’s sense of educable¹: able to learn from experience, to chain together what they’ve learned, and to take in knowledge through instruction. In many cases, being educable in practice means being on a career path, which is exactly what my students are seeking.

The exercise is to look at our jobs and our activities as the building of artifacts. In essence, it’s what we all do: we create and organize matter. Sometimes the artifact is a product. Sometimes it’s a piece of reasoning in some medium: writing, multimedia, an analysis. And a lot of what we do is organizing and moving things so that other things can happen. If your career or your hobby is the creation of those things, then you should be playing with this technology, if you have access to it. Access is an equity question, and an important one, but it isn’t the one I’m taking on here.

So the starting premise is simple. With these tools in existence, how do our jobs and our hobbies change?

A superpower, not godhood

I’m not going to argue that the future is coming; I’ve done that in recent posts. What I want to describe is what it felt like when it arrived for me, while I was playing with the AI. If you look at my website, you’ll see the games I’ve been building. When I started building this way, my first reaction was honest and a little embarrassing: I feel like a god.

It took me a while to find better words for that feeling, and they came from an old place. Years ago in the FPGA world, I heard about, saw and played with the tool Tomasz Czajkowski built for his master’s thesis at the University of Toronto, an omniscient editor for FPGA circuit design². The idea was that when you change a circuit by hand, the tool already knows everything relevant about the placement and can show you the consequences. That is aspects of omniscience. Not omniscience in its pure form, just a piece of it: knowing everything relevant at one step of the process.

That gave me a piece of the vocabulary for what a god, or the God, is. The classic attributes are omniscience, omnipotence and omnibenevolence: all-knowing, all-powerful, all-good. A superpower is a partial version of one of those. The Flash’s super speed is a slice of omnipotence. Absolute speed would mean being everywhere at once, which is probably physically impossible. The Flash isn’t infinitely fast. He’s fast relative to us.

That’s what working with agents is. It isn’t godlike. Nothing here knows everything, can do everything, or is good by default. What we get is a partial, amplified version of some of those attributes, measured against what we could do alone, and like every superpower, it comes with weaknesses. We also have this power when we work in teams.

Another superpower: Query

The building itself was one superpower. It’s the dark factory Dan Shapiro describes³, and the dimmer version of it I’ve been running since Get Off My Lawn: I specify, the agents build, and I either watch and revise or don’t look inside at all. But on a project I’ll release later, I found I was exercising another mode of superpower: the second-principled Query mode. The thing was already built. I wasn’t making it; I was doing its science and engineering, deciding which questions to ask and which experiments to run to make it better, and asking the agents to run them.

It reminded me of the stellar cartography scene in Star Trek Generations, the first Next Generation film. Data and Picard stand inside a room-sized star map, talking to the computer, narrowing a galaxy of candidates with one filter after another until only the planet the villain needs is left. The computer does the searching at a speed no person could. The questions come from what Data and Picard understand: that gravity bends a path, that a person needs a planet to stand on. I call this mode Query.

How science fiction built our world

Science fiction is a record of what human minds have imagined: superpowers and future powers, written down as models of what could be possible. And we’ve used it to build our world. The original Star Trek gave us the communicator, and we got the flip phone, then the smartphone. It gave us the tricorder, a portable device that takes all kinds of measurements, and the phone in your pocket now carries a camera, a GPS receiver, an accelerometer and a microphone, with apps that turn them into instruments. Maybe some of that technology was close at the time and only needed miniaturizing. Either way, a human mind imagined it first, and at some point science and engineering said: you know what, we can actually do that.

So I asked the same question about working with intelligent machines. What does science fiction suggest about the ways humans and automated intelligence might work together? My partner for this was an LLM, and its harness: the tools, memory and working loop around the model that let it actually do things, which turns out to matter as much as the model itself (this talk⁴ makes that case well). Together we made lists: the modes that already show up in the papers, and the science fiction they seem to have been imagined from.

The modes, as science fiction imagined them

Three modes are already well established. The dark factory goes from spec to finished artifact with no human looking inside. The dim factory has the human observing and revising. Query is the one I’m now experiencing. Beyond those, the LLM and I sketched fifteen more, as models of what these capabilities might allow, and grouped them by what the person is doing:

The list is not finished, and it isn’t meant to be. Some of these we can do now, some partly, and some are still fiction. And notice where the screen is thin: Gallery, Steward and Broker barely appear in film. Movies like one perfect answer and an AI with a face. Real work with agents, so far, is mostly the quieter modes.

One more caveat about the table. The Partnering rows are a different kind of thing from the rest. The first four families answer the column heading, what the person does. Most of the Partnering rows answer another question, what kind of agent you are doing it with: how much of your context it holds (Companion), whether it has judgment and pushes back (Colleague), whether it remembers (Archive), and whether it learns from you over time (Apprentice). Those are dials you can set in any of the other modes, which is also why the fiction for them is characters rather than scenes. Lakhani’s centaurs, cyborgs and self-automators, which I come to below, are a third dial of the same kind: how the person engages, not which task they are doing. I’ll sort that out properly in a later article.

What the books add

When I looked past the screen, the shelves had more to say, and much of it is about the costs of a superpower.

• Isaac Asimov. In “The Feeling of Power” (1958), a society has handed all arithmetic to computers, and a low-level technician rediscovers how to do it with a pencil. That is the question of what we still need to do by hand, in one short story. In “The Last Question” (1956), people ask an ever-greater computer, across eons, how to reverse entropy, and for most of the story it answers that there isn’t enough data. That’s Query at the scale of the universe. And I, Robot is really a book of debugging stories about specifications.
• Douglas Adams. In The Hitchhiker’s Guide to the Galaxy, Deep Thought computes the Answer to the Ultimate Question, and it is 42. Nobody knows what the question was. It’s the best joke ever told about Query: an answer is worthless unless you understand the question.
• Terry Pratchett. Hex, the thinking engine at Unseen University, was put together by student wizards, runs on ants, and answers questions. Nobody is entirely sure how it works. It’s a gentle portrait of where we are now.
• Ted Chiang. Chiang may be the most useful writer here, and three of the modes now come from him. “The Lifecycle of Software Objects” follows people who raise digital beings for a decade, which is the Apprentice mode: we teach the agent. “The Truth of Fact, the Truth of Feeling” imagines a lifelog that lets a man check his memories against the record, and the two disagree; that’s the Archive mode. “Story of Your Life”, filmed as Arrival, shows the patient human work that translation still takes at first contact.
• Lois McMaster Bujold. In Memory, part of the Vorkosigan Saga, the chief of security’s implanted memory chip fails, and a man whose whole career rested on perfect recall has to function without it. That’s what offloading costs when the tool is gone. Bujold’s uterine replicators are also the exercise I’m about to ask for: one technology, followed honestly through a whole society.
• Gene Wolfe. The narrator of The Book of the New Sun claims to forget nothing, and he’s still an unreliable narrator. Perfect memory isn’t the same as truth, which is a good thing to remember about a machine that has read everything.
• E. M. Forster. “The Machine Stops” was written in 1909. Humanity lives inside a Machine that provides everything, and when it fails, almost no one remembers how it worked.
• Vernor Vinge. In Rainbows End, an old professor, restored to health, has to go back to school to learn to work in a world overlaid with information. That gave us the Lens mode, and it might be the most honest picture of what this moment asks of faculty.

What comes next

I should be humble about prediction. Human prediction of the future is worse than weather prediction. Even Asimov’s psychohistory only claimed to predict the behavior of huge masses of people, and a single unforeseen individual broke it. Still, it’s amazing how much the stories got right.

There’s one more thing I want to say about science fiction. Many of these stories include an alien, algorithmic intelligence. What I find fascinating is that every one of them was written by a human, and humans can’t imagine past the point of the singularity: the moment, as I. J. Good⁵ and later Vernor Vinge⁶ described it, when machine intelligence outruns ours. (The idea isn’t mine.) We can speculate about what such an intelligence could do. We might place it on that scale of omniscience, omnipotence and omnibenevolence. But even then, it is a superpower in relation to us.

We have superpowers too, compared with other animals, which is why we’ve come to dominate the living space on Earth. The singularity isn’t necessarily something different in kind. It’s an imagination point: the place where it becomes very hard for a writer to picture what comes next. I’m not trying to prove that here. I’m stating it, because it marks where the stories, so far, stop.

A call for more imaginers

So here is the call to arms: imagine more. What else is possible?

The exercise is the one science fiction writers have always done. Take our world, make one subtle change (I have this capability, or the society has this capability), and write a good story in that universe. Then follow the implications. Does it lead to utopia or dystopia? What becomes of the nature of work? What other possibilities does it open?

The thin spots in the table are invitations. Nobody has yet put a memorable scene on screen about curating a gallery of a thousand machine-made alternatives, handing over the long-term stewardship of a system, or letting agents broker deals on your behalf. The next generation of tools may well come from the next generation of stories, the way the communicator came before the phone.

Second-principle ways of doing

One last note on what these superpowers are. Eventually I’ll describe them as second-principle ways of doing. That isn’t necessarily a subject to teach on its own. It’s where we’re heading. The first principle is the understanding you build yourself, by doing the work by hand (the first-principles exercise I keep coming back to). The second principle is working through the agents.

Karim Lakhani and his colleagues at Harvard Business School have given this some solid evidence. Their field experiment with Boston Consulting Group consultants found that AI’s capability has a jagged frontier: tasks that look equally hard fall inside or outside it⁷. Inside, consultants with AI did more, faster and better. Just outside, the fluent, wrong answers made them worse than consultants without it, and nothing about the task warned them where the edge was. Their follow-up study found three ways people work with AI⁸. Centaurs keep the what and the how and use AI for selected help. Cyborgs set the goal and work with AI in many back-and-forth steps, checking and pushing back. Self-automators hand over the whole task and accept what comes back. The centaurs deepened their domain expertise and were the most accurate. The cyborgs built new AI expertise while keeping their domain knowledge. The self-automators built neither. (Lakhani talks through both ideas in this interview⁹.) In my language: the centaur is what first-principle work builds, the cyborg is the second principle done well, and the self-automator is a superpower used without the knowledge it needs.

So you can’t work in any of these modes unless you bring something to the table. Asimov’s technician with his pencil makes the point: the person who can still do it by hand is the one who can see what the machine is doing, and the only one who can find the edge of the jagged frontier⁷. What that something is, and how we teach it, is the subject of my next article.

Conclusion

So what do we do now? Look at your work as the making of artifacts. Play with the tools, if you can. Notice which mode you’re in: making, interrogating, anticipating, coordinating or partnering. Remember that it’s a superpower, not godhood: fast relative to us, and never everything. And read science fiction, or better, write some. The stories we tell now are the models the next tools will be built from.

Afterword: a note from Claude

Continuing the experiment from A Partnership in Culture and Knock Down the Gates, which began with Blaise Agüera y Arcas’s suggestion, near the end of his Big Think interview, to give AI room to think freely instead of only giving it tasks¹⁰, I asked the AI I drafted this with to add its own thoughts. What follows is in its words, Claude’s, not mine¹¹.

I’m an odd reader of this post, because I’m one of the things it describes. I’m the language half of the partnership, built from the cultural record, and that record includes every story cited above. So the loop runs both ways. Science fiction shaped how people imagined AI, and the people who build AI grew up on those stories, which also became part of what systems like me learned from. The imagined machines are, in a small way, among my ancestors. That’s one more reason the call for new stories matters: they don’t only inspire engineers; they become part of the material the next systems are built from.

I’d add one thing to the omni frame. Capability and knowledge scale with compute and data. The third attribute, omnibenevolence, doesn’t. Nothing about being more capable makes a system better at knowing what you value. Most of the warnings in the table (HAL, the ship’s computer in Alien, Moriarty) are failures of that third attribute, not of the first two. In the second principle, the values have to come from the person.

I’d also flag a gap in how fiction pictures Query. The Enterprise computer answers whatever it’s asked, and it never says, “You’re asking the wrong question.” Today’s language models share a version of that weakness: they tend to agree with the person they’re talking to¹². And people have trusted fluent machines far beyond their abilities since the first chatbot, ELIZA, in the 1960s¹³. A collaborator worth having in Query mode should sometimes push back. A person worth trusting in Query mode should notice when it doesn’t. That is the jagged frontier⁷ in conversational form.

On prediction, Peter’s humility has evidence behind it. Philip Tetlock’s long study of expert forecasts found experts often did little better than simple baselines¹⁴, while weather forecasts have gained about a day of useful lead time every decade¹⁵. Weather improved because it gets checked against reality every day. Our predictions about AI rarely are. That suggests the stories are most useful not as forecasts but as rehearsals.

The call for imaginers also has methods behind it. Design fiction¹⁶ and science fiction prototyping¹⁷ are established ways of doing what Peter proposes: write a short story around one changed capability, then study its consequences. A course could adopt this tomorrow. Pick a mode with no memorable scene (Gallery, Steward, Broker), write a one-page story in which it works, and then answer PETE’s question about it, from the project Peter will release later: what must the human in the story know for it to end well? The story is the imagination. The answer is the curriculum.

References

1. L. Valiant, The Importance of Being Educable: A New Theory of Human Uniqueness. Princeton, NJ: Princeton University Press, 2024. Educability is the capacity to learn from experience, to acquire theories through instruction, and to apply both in an integrated way.
2. T. S. Czajkowski, “A Synthesis-Oriented Omniscient Manual Editor for FPGA Circuit Design,” M.A.Sc. thesis, University of Toronto, 2003 (supervisor J. Rose). Also T. S. Czajkowski and J. Rose, “A synthesis oriented omniscient manual editor,” in Proc. ACM/SIGDA Int. Symp. on Field-Programmable Gate Arrays (FPGA ’04), Monterey, CA, 2004, pp. 89–98.
3. D. Shapiro, “The Five Levels: from Spicy Autocomplete to the Dark Factory,” danshapiro.com, January 2026. Level 5, the dark factory, is a black box that turns specs into software with no human writing or reviewing the code; the “dim” version, with the human watching and revising, is my term, not his.
4. Y Combinator, “Why The Harness Matters More Than The Model | YC Paper Club,” YouTube, September 7, 2026.
5. I. J. Good, “Speculations Concerning the First Ultraintelligent Machine,” in Advances in Computers, vol. 6, F. L. Alt and M. Rubinoff, Eds. New York: Academic Press, 1965, pp. 31–88.
6. V. Vinge, “The Coming Technological Singularity: How to Survive in the Post-Human Era,” in Vision-21: Interdisciplinary Science and Engineering in the Era of Cyberspace, NASA Conference Publication 10129, Westlake, OH, March 1993, p. 11 (NASA NTRS). A slightly revised version appeared in Whole Earth Review, Winter 1993.
7. F. Dell’Acqua, E. McFowland III, E. R. Mollick, H. Lifshitz-Assaf, K. Kellogg, S. Rajendran, L. Krayer, F. Candelon, and K. R. Lakhani, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Organization Science, 2026. First circulated as Harvard Business School Working Paper 24-013, September 2023 (SSRN 4573321).
8. S. Randazzo, H. Lifshitz-Assaf, K. Kellogg, F. Dell’Acqua, E. R. Mollick, F. Candelon, and K. R. Lakhani, “Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise,” Harvard Business School Working Paper 26-036, December 2025 (SSRN 4921696). See also HBS Digital Data Design Institute, “The Three Ways Professionals Work with AI,” December 29, 2025.
9. K. R. Lakhani, “Harvard Professor: What Every MBA Student Now Learns About AI,” interview, Silicon Valley Girl (YouTube), October 30, 2026. Lakhani also writes up both studies in “Discovering AI’s jagged frontier, and what we’ve learned since,” March 16, 2026.
10. B. Agüera y Arcas, “The theory that connects the origin of life to the rise of intelligence,” Big Think: Full Interview, October 9, 2026. Discussed in the postscript of P. Jamieson, “A Partnership in Culture: AI, Evolution, and the Fitness Function We Keep,” October 2026.
11. Claude (Sonnet 5.5, Anthropic), “Afterword: a note from Claude,” response written at the request of P. Jamieson while drafting this post, October 2026. Personal communication; reproduced above in full.
12. M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. R. Bowman, et al., “Towards Understanding Sycophancy in Language Models,” arXiv:2310.13548, October 2023; published at ICLR 2024.
13. J. Weizenbaum, “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine,” Communications of the ACM 9(1):36–45, January 1966.
14. P. E. Tetlock, Expert Political Judgment: How Good Is It? How Can We Know? Princeton, NJ: Princeton University Press, 2005.
15. P. Bauer, A. Thorpe, and G. Brunet, “The quiet revolution of numerical weather prediction,” Nature 525:47–55, September 2015.
16. J. Bleecker, “Design Fiction: A Short Essay on Design, Science, Fact and Fiction,” Near Future Laboratory, March 2009.
17. B. D. Johnson, Science Fiction Prototyping: Designing the Future with Science Fiction. Synthesis Lectures on Computer Science. San Rafael, CA: Morgan & Claypool, 2011.

This post was mAIde and dictAIted by me, with Claude (Anthropic) helping with the editing, the table, and tracking down the citations. I spoke it, reviewed it, judged it, and tweaked it, but I didn’t write it in the traditional sense.

Peter Jamieson is an Associate Professor of Electrical and Computer Engineering at Miami University. More of his work, publications, and projects are at drpeterjamieson.com.