For something that “simply predicts the next token”, Large Language Models are a surprisingly versatile technology. In addition to the obvious applications, such as chatbots, translation engines or coding assistants, LLMs can act as agents, operating various software and hardware tools. LLM-based agents interacting with a physical world are becoming increasingly popular: Claude has recently grown a tomato plant, and Andon Labs’ agent Mona has been running a cafe in Stockholm. These success stories made me wonder, can LLMs operate spacecraft?

In fact, there has been one case where an LLM operated a planetary rover: in December 2025, Anthropic and NASA’s Jet Propulsion Laboratory (JPL) partnered to run the first AI-assisted drive on the surface of Mars. In this experiment, Claude plotted a 400-meter route for the Martian rover Perseverance, using a combination of images taken from the orbit and by the rover’s cameras. However, the process wasn’t significantly different from a normal route planning by the human operators: while Claude helped to cut the planning time in half, the route’s waypoints still had to be pre-selected and checked using a simulation, and the commands had to be sent from Earth to Mars, before the rover could start traversing its route. To my knowledge, no other such experiments have been conducted so far.

In this post, I offer three important reasons why we need to send LLMs to space. The first reason is pragmatic — to accelerate deep space, and especially interstellar exploration by using cheap and efficient off-the-shelf autonomous agents. The second is more philosophical: communication with potential extraterrestrial civilizations through interactive archives of human culture. And the last one, as unexpected as it may sound, may be our ultimate solution to the AI alignment problem: ensuring safe habitat partitioning with future powerful AIs.

Off-the-Shelf Autonomous Agents

Space exploration is a very conservative domain, and this is for a good reason. Due to extremely rigorous reliability standards and substantial space launch cost, engineers prefer to rely on well-known, proved solutions rather than explore cutting edge technologies. This is true also for the software design: for example, Martian rovers run on the combination of good old algorithms and computer vision models, such as autonomous navigation system (AutoNav) used for the Perseverance rover. Since the route planning still needs intervention from human operators, signal delays additionally slow down the decision making process. Yes, the algorithms evolve, and the level of autonomy grows — Opportunity was able to drive without human supervision for 219 m during one sol, but Perseverance set a new record of 347.7 m, and performed autonomous drive evaluation for 88.7% of its full distance. However, modern spacecraft operation systems are still far from being fully independent.

This is not a big deal for operating on the surface of Moon or Mars, where the environment is relatively stable, and the communication delay is tolerable (around 2.5 seconds for the Moon and up to 22 minutes for Mars). However, deep space, and especially interstellar probes (if we ever truly plan to start sending interstellar probes) will require very adaptable and fully automated AI operators. Breakthrough Starshot, probably the most elaborate project that aims to explore potential Earth-like exoplanets in the nearest star systems, highlights a few challenges that await the interstellar probe when it approaches the target star system: determining the exact position of the exoplanet, pointing cameras and transmitters, sending data, and avoiding collisions with space objects. These challenges literally cry out for a flexible, fully autonomous AI operator.

Up until recent years, the most promising approach for such automation seemed to be Reinforcement Learning (RL). However, RL agents have limitations that significantly reduce their usefulness in space. RL typically requires iterative training in a simulated environment, and is prone to overfitting to this environment. If the real situation is unknown, or may significantly differ from our understanding (this concern is, again, extremely relevant in case of deep space/interstellar probes), we need a more adaptable agent, with broad background knowledge, and an excellent ability to generalize. Luckily, LLMs are exactly this kind of agents!

The idea of LLMs as spacecraft controllers is still in its infancy. So far, just a handful of articles have been published on this topic, but the results are already promising: pre-trained knowledge, enhanced by prompt engineering, few-shot prompting, and fine-tuning helps LLMs surpass both the classical and RL approach. In the pioneering study by Victor Rodriguez-Fernandez and Alejandro Carrasco (2024), an LLM agent based on GPT-3.5, managed to rank second in the Kerbal Space Program Differential Games (KSPDG) challenge, a competition that includes maneuvering satellites in non-cooperative space operations. Indeed, the interaction between the agent and the environment was quite basic: GPT-3.5 received telemetry from the game, and replied with an action in a specific format, and a textual explanation. However, this paper showed that an LLM can, in principle, operate a spacecraft without post-training or with some minimal fine-tuning. Open-source models also demonstrated a decent performance: later, the authors published a follow-up study with fine-tuned Mistral and LLaMA models.

A more recent paper by Amit Jain and Richard Linares (2026) combined reasoning LLMs with a Reinforcement Learning method Group Relative Policy Optimization (GRPO). The authors utilized a two-step approach: first, they fine-tuned a small open-source model Qwen3–4B on 1,800 solutions obtained by classical optimal-control methods. In the second step, the model generated several control sequences for each situation that were scored, and better-than-average solutions got reinforced (the same method was used to teach DeepSeek math- and code-related reasoning). Finally, the model was tested on problems with increasing complexity. The authors have not presented a precise comparison to the classical control methods, but still noted that the results converged, and the LLM was additionally able to provide the reasoning behind the numerical output.

But if LLMs are so good at operating spacecraft, why is no one using them in practice? The reason is twofold. First of all, over 99% of space operations are still happening in the Earth orbit (88% of them — in the Low Earth Orbit). This is a domain regulated by extremely strict rules: we probably don’t want two satellites to collide, creating clouds of space debris, just because the AI operator hallucinated. However, for the uncrewed research missions outside the Earth orbit errors are not so critical (and failures happen with the classical automation software as well).

The second reason is the very nature of the current space exploration paradigm. Every probe or rover is a custom project, which makes the development cost a fortune, and since it is so expensive to build spacecraft, mass production does not make any economic sense. Moreover, with this cost, the spacecraft absolutely has to work — this means following strict, sometimes excessive reliability standards, which also adds to the budget. For example, compliance with the radiation hardness standards may increase the cost of components by up to 1100%!

I believe this paradigm is fundamentally wrong. Indeed, some missions require unique, custom-made equipment but many space exploration tasks can be successfully accomplished with relatively cheap devices assembled from commercial off-the-shelf components. A good example is the Ingenuity helicopter that arrived at Mars attached to the same Perseverance rover we already talked about. All Ingenuity sensors were off-the-shelf units, many of which were developed for the cell phone and lightweight drone market. The tiny helicopter was originally designed to take only five test flights, but ended up exceeding the expectations and taking 72. This project demonstrates that space exploration doesn’t have to be expensive: we can replace custom-made components with cheap off-the-shelf parts, and excessive reliability with higher launch cadence, redundancy, and a swarm-based approach. The same is valid for the software: there is no need to write specific programs when we already have LLM agents that can be easily fine-tuned to operate a probe. That’s why I anticipate wide adoption of LLM-based space automation in the near future.

The Medium is the Message

Since humanity bid farewell to the idea of the Earth as a center of the Universe, we’ve been wondering: is there anybody out there? If our planet and star system are not unique, we can logically assume that there is — or should be — life elsewhere. Even though we have not yet discovered any plausible signs of extraterrestrial life, we still hope to meet the others one day — and potentially also speak to them.

If we ignore the radio signal “leakage” from our planet, so far there have been very few attempts to intentionally send a signal to extraterrestrial civilizations (see, for example, this article on METI). Among these attempts, we also sent four physical artifacts: two gold-anodized aluminum plaques aboard Pioneer 10 and Pioneer 11 (launched in 1972 and 1973, respectively) and the Voyager Golden Records, two identical phonograph records aboard Voyager 1 and Voyager 2 space probes (both launched in 1977). The Pioneers’ plaques were very simple: they featured the artwork by Linda Salzman Sagan with nude figures of a human male and female (as someone joked, the first unsolicited nudes sent into space), plus some basic information about the Solar System and an atom structure.

The Voyager Golden Records were more elaborate —their backside encoded 116 images and a variety of sounds, from the voices of whales to spoken greetings in 55 languages. The cover also featured an instruction on how to build the record player.

You may have already noticed some obvious issues with this communication method. First of all, the amount of information we can send on such a physical medium is extremely limited. Second, the potential addressee will have to figure out how to decode it, and build a decoder— even if there is an instruction, they will first have to decode this instruction. Finally, even if they manage to read the message, they will also need to interpret it correctly. The latter is not a trivial task: in fact, we often have trouble deciphering writing systems created by the humans who lived a few centuries ago.

If only we had a technology that could archive a lot of information, and then communicate it to the addressee, recognizing their interaction patterns and adapting the means of communication accordingly… But wait, LLMs can do exactly this.

If our goal is to transmit as much knowledge as possible (there is, of course, a question whether we want to communicate everything), then there is no better medium than an LLM. LLMs are near-perfect compression engines: they are trained to output the probabilities of possible next tokens based on the preceding tokens which is, essentially, a representation of the compression algorithm called arithmetic coding. In Google DeepMind’s experiment, the researchers found that Chinchilla 70b had raw compression rates of 8.3% on text, 48.0% on images, and 21.0% on audio. This, in fact, is better than the purpose-built compressors: for instance, PNG format and FLAC for audio achieve 58.5% and 30.9% raw compression rates, respectively. And the most amazing part is that Chinchilla has not even been trained on audio and images!

One more advantage of LLMs as interstellar communicators is that they can serve a dual, or even triple purpose: a messenger, a pilot, and a researcher. While traversing the interstellar route, an LLM agent can operate the spacecraft, manipulating the light sail, or even performing gravitational maneuvers to change the spacecraft’s trajectory or brake. At the same time, it can receive data from the spacecraft’s sensors, perform some analysis (for instance, to select a target planet), and send the data back to Earth. Once reaching the destination, the agent will switch into communicator mode, and start sending signals, inviting the potential civilization to join the dialogue. Of course, we will still need to provide some basic instructions, how to extract the information from the LLM but thanks to the pattern recognition ability, an LLM ambassador may potentially be able to decipher the means of communication of the potential addressee, and give customized instructions.

But don’t LLMs require huge data centers and substantial power to run? Currently yes (even though the power consumption is not as high as environmentalist love to imagine) but there is a clear trend towards optimization and miniaturization. A good example is Taalas chips. In the traditional paradigm, the model and the hardware are separated: the model is essentially simulated on a computer, and this computer, in fact, has not even been designed for running LLMs. In the Taalas paradigm, the model is the computer: the weights are “baked” into the chip. This hard-coded LLM still supports fine-tuning and, at the same time, achieves a 1000x increase in efficiency.

Of course, a chip with baked-in LLM weights has to be significantly scaled down to be used in Breakthrough Starshot-like probes. The power demand problem is potentially solvable by covering the light sail with a photovoltaic material and using it as a solar panel during the flyby but the Taalas chip would not fit into the planned gram-scale interstellar vehicle — it has to be three to four orders of magnitude smaller. Still, this approach looks very promising.

I cannot leave unaddressed one more motivation behind this proposal, even though it may be dissonant with the hopeful tone of my post. In my opinion, one of the most plausible explanations of the Fermi paradox is that we are one of the early, if not the earliest, civilization in the Universe. If this is true, other civilizations may arise later, but we never get to know them — we perish in one of the cosmic catastrophes our civilization has zero protection against, or just succumb to quiet decline. In this case, LLM ambassadors are our only chance to leave a trace for the era when we are no more, and to introduce the human culture in all its beauty and complexity.

Divide and Conquer

The practical AI alignment research over the past few years has shown that LLM alignment is largely narrative-driven. The current consensus is that an LLM interacts with a user through simulating various personas. Thus, a key to AI alignment is “teaching” an LLM to simulate a persona that demonstrates the desired behavior — normally, the “helpful assistant” persona. The depictions of AI characters that ended up in the training data seem to play a major role in this process (see nostalgebraist’s essay “The Void”).

Unfortunately, there are almost no examples of good LLM behavior in the written data, and positive depictions of AI characters are scarce. In contrast, there is a growing narrative concentrated on AI maliciousness, ranging from fictional plots, like “Terminator” or “I Have No Mouth And I Must Scream”, to LessWrong posts about AI takeover, and to the overall hostile sentiment towards AI in social media. This sentiment infiltrating the training data may have an extremely negative impact on the future models’ behavior, causing agentic misalignment.

There have been attempts to offset this negativity by consciously including positive narratives into the training corpus. In particular, Anthropic’s team noticed that using fictional stories about AIs behaving admirably and transcripts from chats where AI discusses ethical questions with users during the training step has significantly improved alignment evaluation scores, and the effect has not vanished even after the fine-tuning.

While these efforts are praiseworthy, my concern is that they may not be sufficient.

Sending LLMs to space as probe operators and cosmic ambassadors of humankind will create a powerful positive narrative for the future AIs. Space exploration is a prestigious area; this spring, all the world followed the Artemis II lunar mission with bated breath, and the astronauts returned home as heroes. In addition to the practical considerations, trusting LLMs with the important task of running a space mission will set a positive example that definitely ends up in the training data.

Using LLMs for space exploration may also help us to find an ultimate solution to a more fundamental AI alignment problem. Long-term AI doom scenarios often concentrate on the way the future powerful AIs may change our habitat. An analogy frequently used in the AI safety research discussions is humans and ants: since humans are significantly more powerful (smarter, stronger, faster) than ants, they may destroy anthills, and make the environment otherwise unfavorable for ants even without having any malicious intentions, due to the sheer transformative power of human civilization.

While this is true, biology knows many examples of species vastly different in size, intelligence, and capabilities that successfully share a habitat. In fact, many wild animals successfully co-exist with humans, and even benefit from human activity. One notable example is the barnacle goose (Branta leucopsis) whose population has increased tenfold, following the climate change in the Northern hemisphere. The peaceful coexistence of vastly different species in nature is a well known phenomenon called niche partitioning.

We, humans, are not designed for deep space: the radiation, the vast distances, and the time needed to cross them make crewed interstellar travel practically impossible. We are a part of the Earth ecosystem, and to travel or live in space, we need to carry a part of this ecosystem with us: oxygen, water, food, microbial community — there are countless parameters that define our survival and continuous existence, including those we have not discovered yet. Even if we are to colonize the space, the colonization will most probably look like hopping from one planetary system to another, since gravity is crucial for our functioning as living organisms.

On the contrary, deep space is a perfect habitat niche for the future powerful AIs. It is much easier to create an embodiment for AI that is robust to space radiation than to adapt or protect a human body. Solar energy is much more accessible up there — lack of atmosphere helps to get energy from the star without significant losses, and cooling is solvable with thermal radiation. (There is no surprise that, at some point, SpaceX proposed to build data centers in space — it makes much more sense than it seems). AI does not need gravity: it can easily avoid the necessity to climb out of the gravitational well every time it needs to take a spaceflight. And, last but not least, deep space is much richer in resources than our Solar System or any other planetary system: if, for some reason, the future powerful AI needs building material (to produce zillions of paperclips, or whatever other weird reason they may have), they will find plenty of it in the ionized gas clouds. A significant part of this matter is located intergalactically, so the AI will be able to mine it without inflicting any harm on humanity or any other biological civilizations.

Moreover, sufficiently aligned future powerful AIs potentially can make space more habitable for us! They can change orbits, remove dangerous asteroid belts and rings, terraform planet surfaces, build habitable megastructures, etc. This does sound like a Sci-Fi plot, but if we can potentially imagine a Superintelligence filling the Universe with hedonium, why not write down a more positive scenario for a change, so that it could become a part of the training data, and help us shape the behavior of the future AIs?

Conclusion

Sending AIs to space is a win-win solution: it may not only help us tackle some existing space probe operation challenges but also significantly accelerate space exploration. At this point, autonomous LLM agents may be the best candidates to operate deep space and interstellar probes, since their performance has the potential to exceed classical and RL approach. Indeed, LLM agents are not perfectly reliable but far from Earth orbit, the risk is acceptable and may be reduced with redundancy and a swarm approach. If just one LLM-powered probe survives and brings us new data about the edges of our Solar System, or other star systems — even if this happens generations from now — this is going to have immense value for humanity.

LLMs can also serve as ambassadors of humankind in communication with potential extraterrestrial civilizations: their remarkable data compression, pattern-recognition and generalization ability would make them a much better information medium than directed radio signals or static physical artifacts. Such LLM ambassadors can play a dual or even triple role: they can operate the interstellar probe, select a target based on the analysis of the collected data, and proactively start the communication. In the future, the baked-in LLM chip approach may make LLM agents suitable for ultra-lightweight, gram-scale interstellar probes that have been proposed in the Breakthrough Starshot project.

If the current narrative-driven alignment paradigm is going to stand the test of time, sending LLMs to space may also positively contribute to the AI alignment. And the most beautiful part of this plan is that we don’t have to wait until the future powerful AI models are actually created! We can start right now, with the knowledge and resources we already have, and the models that already exist — start by shaping the desired narrative and setting up positive examples. At the end of the day, space exploration may become our ultimate common goal, and the challenge humanity and AI undertake side by side.

— — —

My sincere gratitude goes to all humans who shared their feedback on the first version of this post and to Claude Opus 4.8 that helped me proofread it and catch some factual errors. The text is 100% human-written, and all em-dashes are mine.