"Wow…" I stare at him. "Humans spent thousands of years looking up at the stars and wondering what was out there. You guys never saw stars at all but you still worked space travel. What an amazing people you Eridians must be. Scientific geniuses."
Andy Weir, Project Hail Mary
Last week I finished reading the book, and then I went with my son to watch the movie Project Hail Mary. The computers in the story are not particularly intelligent. At times they come across as basic, even a bit dim compared to the generative AI tools we are using today. What I found more interesting was the alien, Rocky, and the way the human character collaborates with it. There is something in that collaboration that made me think about how we work with AI agents and tools. I highly recommend the book and the movie, and I will try not to spoil too much below.
The human and the alien intelligence start off without a shared language. They don't really understand each other. Gradually, they find common terms grounded in physical and mathematical laws, and they build a shared vocabulary on top of that base. Dr. Ryland Grace writes a small program that translates the sounds Rocky makes into English and back. If he had had ChatGPT, Gemini, or Claude on board, I suspect he could have used them instead of the custom code. Over time, the two of them pick up enough of each other's language that the translator becomes less necessary. As we work more with AI tools, I think we develop a similar kind of fluency. We learn why the model phrases something a particular way, where it tends to be confident, where it is bluffing, and how to ask for what we actually want.
Their bodies are not built the same way. Rocky has no eyes; it perceives the world through sound, closer to a bat than to us. It has a much better working memory than Grace does, and it does arithmetic faster. Their work together only starts to hit off when they each understand what the other can and cannot do. Grace stops describing things that depend on vision. Rocky stops handing Grace numbers to memorize. They divide the work according to what each is built for.
Most of the time, in 2026, we are still asking AI to do what we would do, and grading it on how closely it matches. Maybe instead we should ask the question the way Project Hail Mary does. What is this thing actually built to do, what are we actually built to do, and given that, how should we divide the work?
Different libraries, not just different senses
There are more interesting anecdotes in the book that made me think of gaps for working with other intelligences. Rocky's species, the Eridians, had been moving through space for far longer than humans, and their materials chemistry produced things Earth had never built. But they had also not worked out the theory of general relativity. The two of them were each, in their own corners, far ahead and far behind at once. A lot of the working tension in the story turns on neither of them assuming the other already knew what they knew, and neither assuming the other did not.
That asymmetry is the one I think we mishandle very often with AI. The models in 2026 have read what amounts to the recorded body of human science. They sound conversant in fields that took centuries to build, but sounding conversant and reasoning correctly are not the same thing. The gap shows up in domains with thin or non-canonical data, in problems that ask you to connect ideas across disciplinary lines humans have not yet connected in print, and in the interpretation of new results that contradict the existing corpus.
The discipline is to ask what the model knows, and then to ask what its version of relativity is. What is the load-bearing concept it does not have, that this particular problem rests on? A research lead who poses that question on purpose is doing different work from a research lead who only asks the model what it knows.
The second direction of this gap is easy to forget. There are concepts the model has assembled across a corpus larger than any single researcher could read, that the human researcher has not. The collaboration can produce 1 + 1 larger than 2 when each side is willing to be educated by the other, and when each takes seriously the possibility of operating with a blind spot it cannot see from the inside.
Imagining the world otherwise
AI is unusually good at running disciplined what if experiments. What if gravity were a third of Earth's. What if the atmosphere were ammonia at 200°C. What if a supply chain did not pass through one country, or one port, or one strait. Imagining a material or a system in conditions far from the ones you grew up assuming is among the older moves in science.
It is also a hard move to do alone, because most of our intuitions are built on Earthly assumptions we have stopped noticing as such. Reaction rates depend on temperature in a way that goes peculiar at very low temperatures, and most of us never reach for that fact in normal practice. Biological molecules fold the way they do partly because water is the solvent. Gravity is constant enough on Earth that we treat it like a number rather than a variable. AI seems well suited to the work of varying these unmarked fixed quantities and walking through what might break.
The point of the exercise is not to publish an answer. The point is to surface what you were assuming. In the book, Rocky's perspective keeps making Grace see things he had stopped seeing, and vice versa. A practical version of that move, for those of us who lead research or strategy, is to use AI deliberately to run scenarios you would normally rule out as irrelevant, and then listen for what your own intuitions object to. The objections are often where the buried assumption lives.
This is also the spirit of the Virtual Lab paper I unpack in this issue's framework section. The biggest payoff there is not the binding nanobody. It is the freedom to ask questions that used to be too expensive to investigate.
A collaboration with higher stakes
One more parallel runs through the book and is relevant to our current situation. The stakes in the story are high for all involved. The future of humanity and the future of Rocky's home world both depend on the collaboration working well. That stakes-bound quality changes how both sides behave. They listen harder. They explain more. They tolerate more confusion before they give up. The posture that the book describes, sustained curiosity about a partner who genuinely does not see what you see, in a setting where it matters whether you get it right, is the posture that might make AI work actually pay off.
There is a moment in the book when Grace is watching Rocky work:
"While we chat, he uses his many hands to assemble some complicated-looking piece of equipment. It's almost as big as he is. I recognize several parts on it as things he's been repairing these past few days. He can hold a conversation and work on delicate machinery at the same time. I think Eridians are much better at doing multiple tasks than humans are."
Rocky has many hands, and he uses them. The parallel to the multi-agent setups now becoming common in research is fairly direct. The collaborator has more hands than we do. What work will we choose to put them to?
Where the metaphor breaks
A note on the limit of the metaphor before we close. Rocky is one alien. He has continuity, intent, memory of what he told you yesterday. AI agents in 2026 are not that. They are many, often stateless, easily replicable, and they do not really want anything.
The fictional human worked with one Rocky over months. We can work with a hundred copies of a similar agent over an hour. What we can do with that, for instance running the same conversation five times in parallel and letting the disagreements between the runs tell us something, is something neither humans nor Rocky have ever been able to do with each other. The closest practical example I have seen this year is in the Stanford Virtual Lab, where five parallel runs of the same research meeting become the engine that makes the system reliable.
Your action step
Pick one decision you are about to delegate to AI this week. Before you ask the model what to do, ask three questions in this order:
- What is this model actually built to do well, that I am not?
- What am I built to do well, that this model is not?
- What is the load-bearing concept this decision rests on, and do I believe the model has it?
If you cannot answer the third one, you are not yet ready to delegate. You are ready to learn what the model's version of relativity looks like for your problem. That is a different kind of session, and it is the one most worth running first.
If you are designing how your teams collaborate with AI agents and want a second pair of eyes on what to delegate and what to keep, that is the conversation I have with leadership teams in AI strategy advisory engagements and keynote and workshop sessions across Europe.