The Shorter Loop

AI-assisted research didn’t give me answers: It made it cheap to find out I was wrong.

See LASSO at: https://zenodo.org/records/21928516.

One idea had been sitting on my shelf since I was a teenager: If the Moon needs metal, and asteroids already contain it, why not bring the asteroid to the Moon and mine it there?

Not fly out and build a refinery in microgravity. Not soft-landing it on the Moon. Not braking the asteroid into orbit, which an earlier NASA study found too expensive. Others found soft-landing hundreds of tons of asteroid material also unacceptably costly. But, as kids obsessed with meteorites know, King Tut had a dagger made of star metal, and metal from other meteorites has also been used all around the Earth.

I always wondered why every proposal insisted on landing the thing gently or mining it while in zero-gravity orbits. It’s a chunk of metal heading for an uninhabited body with an ideal low gravity and no atmosphere. Let it hit. Use the Moon as a catcher’s mitt and process the ore on the surface.

For thirty years, I had no practical way to find out whether that intuition held up against the complex physics. My question crossed orbital mechanics, impact physics, materials science, propulsion, and lunar geology, and likely needed its own research team. Learning enough of every field to discover that the idea was trivial or impossible was its own multi-year project.

Then, having explored the power of generative AI, in March 2026, I put my raw hunch into an AI and asked two initial questions:

“Has anyone already proposed this idea?” No one had, so I could then ask: “Why not?”

The second answer came quickly: Velocity.

A typical asteroid impact at ~ 20 kilometers per second is violent enough to melt, vaporize, bury, or scatter exactly the material the lunar miners want to recover. We all know of Chicxulub at that speed wiping out the dinosaurs, so any naive version of the idea was dead.

That is usually where a cross-disciplinary hunch ends. When an objection arrives from a field you do not own, you assume the specialists know something you do not, thus return your idea to the shelf, and move on.

Instead, with the analytical power of AI, I could make a different move and now ask:

If velocity kills the idea, then velocity is not a final verdict: Velocity is now a design constraint to examine. Thus:

How slowly would an asteroid have to arrive for useful mining material to survive and be recoverable? And could practical orbital geometry reproduce that speed?

I grew up around the concepts of orbits and moving targets because of my father’s work in dynamic control, often in military and space applications. That intuition suggested an inversion: approach the Moon from behind, so the Moon is moving away from the incoming asteroid.

Pursuit geometry

With the Moon orbiting the Earth as it orbits the Sun, I imagined hitting the Moon with the asteroid “from behind,” like the slower relative impact of rear-ending a car in front of you on the freeway. Iterations found that through favorable orbital geometry, the Moon’s own orbital motion subtracts roughly a kilometer per second from the closing speed at essentially no propulsion cost.

The velocity objection had become part of my architecture.

This insight did not happen in one magic prompt: It happened across various serial short iteration sessions spread across my evenings and days off: calculate, question, revise, cross into another domain, come back.

A decade earlier, each step would have required another literature, another vocabulary, probably another expert, ideally a team of collaborators, and days or weeks before I could even formulate the next question correctly.

The difference now mattered more than any single answer the AI gave me. The important output was not the answer.

The key was generative AIs’ speed, accuracy, low cost, and feedback of this new loop.

What once required assembling the right people across several disciplines could now run repeatedly inside a working session: propose, test, find the error, revise, and run again. Each pass was cheap enough to throw away, fast enough that the whole problem stayed mentally live, and corrective enough that the next pass started from a better state.

That process then changes more than research speed. It changes which questions are practical to pursue in the first place. For most of the questions on my “conceptual questions” shelf, the old loop was not merely slow; it never ran at all.

The Shelf

Every curious person has one: our mental shelf of questions that never quite went away. Not necessarily because they were right, but because they were interesting yet unfinished, and finishing them required resources that weren't available or practical.

My shelf of questions has been accumulating for decades.

Asteroid metal delivery was an older item. Another was whether the extraordinary development of the human intellect had more to do with developing the exquisite ballistic computer that our species developed over 2 million years. Another was that those monumental prehistoric sites like Göbekli Tepe were unlikely just churches, so they might look different if you first treated them as functional physical landscapes (with constraints on function, purpose, labor, food, storage, waste, and seasonality) rather than beginning with what they might symbolize. There were others: domestication of dogs, megafauna, assorted evolutionary puzzles, how our Stone Age brains were controlled by modern society and became manipulated by social media, and what logic might explain modern political “strong man” behaviors. The common problem was not lack of curiosity.

The common problem was transaction cost.

Each question sat between disciplines that do not naturally belong to one person, such as: orbital mechanics and impact physics; neuroscience and paleoanthropology; zooarchaeology, landscape engineering, food processing; political science, sociology, etc.

To pursue any one question seriously, I would have to become conversant enough in several literatures merely to find out whether my initial observation was interesting, trivial, already known, or wrong. So most stayed filed on the shelf.

Then recently, with powerful generative AI and LLM models freely available, the cost of crossing those boundaries collapsed.

The Shorter Loop

Large language models did not suddenly make expertise obsolete, but they have given everyone unfiltered (and problematic) access to it.

AIs hallucinate and sycophantically agree too easily. They can sound authoritative while being wrong. AIs are especially dangerous when the user cannot distinguish a retrieved fact from a fluent reconstruction of something the model vaguely remembers. (And LLMs are especially irritating when they overcorrect and become disagreeable and arrogant.) My epidemiology and research design training highlighted for me the myriad potential problems with AI research. But when they are gated and used with caution, they have superpowered cross-domain investigations.

What AI changed was iteration latency, ease, and accuracy.

Before, a cross-domain idea might require days to translate the first objection into the vocabulary of the next field and more resources to explore it. By the time the answer arrived, the architecture that generated the question might no longer be mentally live.

Now the sequence can happen easily and almost continuously:

Hunch → map the landscape → find the kill shot → turn it into a constraint → cross a domain boundary → revise the architecture → attack it again.

AI did not remove expertise, evidence, criticism, or measurement from that chain. It compressed the distance between them and made them accessible.

The iterative AI-powered research loop became radically cheaper. You no longer need to assemble a five-person interdisciplinary team merely to discover that your idea did not deserve one.

And because each pass quickly produces feedback, the process can become self-correcting in a limited but important sense: not because the AI magically improves itself, but because every surviving version can contain information from the failure of the version before it.

Eventually, I formalized the rigorous AI-assisted research methodology I was using into a prompt protocol I call Iterative AI Production, or IAP. The name is literal: It is iterative because each pass changes the object that enters the next pass. It uses AI because retrieval, translation, comparison, adversarial analysis, and drafting make the passes fast and inexpensive enough to stack. And it is production because conversation is not the endpoint. The process is supposed to force something falsifiable into existence: a paper, a framework, a prediction set, an analysis, a collaborator packet, an executable test.

I first formalized IAP while developing another project, my Dominance Operating System political analysis framework. I later formalized my research methods into a standalone protocol (IAP v2.0 and FOH), and used these on LASSO and other work. My FOH (Fan-Out Harvest) protocol structures a primary context window’s workflow for generating and distributing a structured research question across multiple clean context windows, typically five, then harvesting and consolidating the results.

This AI-assisted research workflow is very useful and productive for me across real projects. It does not establish that such IAP and FOH is superior, as no validated outcome measure exists.

The basic loop is simpler than a full protocol:

Externalize. Put the hunch outside your head before polishing it.

Map and bound. Ask what already exists and what observation would kill the idea.

Convert. Turn “that won’t work because X” into “the architecture must satisfy X.”

Expand. Cross into whatever field the constraint requires. Try the inversion: what if the obstacle is actually part of the mechanism?

Attack. Stop helping the idea: Ask for the weakest link, the simplest alternative explanation, the hidden assumption, the reason a specialist would reject it.

Crystallize, then exit. Write the thing down. Then specify what the world would have to show for it to survive.

The full protocol adds evidence tags, architecture changelogs, context bridges, explicit kill criteria, and other machinery I made to reduce the tendency of long AI conversations quietly converting speculations into “facts.”

I have also learned not to let one AI become the entire intellectual environment. I maintain paid access to 4 frontier models and regularly give the current claims or drafts to fresh models with no conversational history, with an “AI peer review” prompt. That is not the same as independent scientific peer review, as the models share training data, techniques, and failure modes. But it is quick, easy, and usually surprisingly helpful to run repeatedly, and such impromptu competitions give invaluable insights into the strengths and weaknesses of the various models.

The advantage is practical: rapid, iterative access to clean context, new insights, and divergent failure modes.

A model that has spent eighty turns helping build an idea has learned its vocabulary, premises, and assumptions. It has conversational momentum that a fresh model does not. Sometimes it immediately walks directly into the premise the first model has learned to step around.

Different models tend to attack different things: prior art, numerical assumptions, logical gaps, ambiguous writing. That divergence is valuable.

I think of this as AI peer review, with an important qualifier: It is not a substitute for human peer review.

My sister Marta uses AI professionally, including consulting on how a company can best implement generative AI, and has developed more formal validation workflows in which human subject-matter experts check model output. Our use cases often differ, and we regularly compare where new models fail, where they improve, and which errors persist across models.

That AI peer-review comparison has been useful. But the most consequential criticism of LASSO came from a human.

What Survived:

LASSO (Lunar Asteroid Salvage by Slower-impact Operation) became a useful worked example because several of the original inversions survived repeated attack. The first was obvious:

Do not brake the asteroid.

Conventional resource architectures treat arrival velocity as something that must be canceled. LASSO asks whether the velocity can instead do useful impact-shattering mining work. The second inversion followed:

Fragmentation is not necessarily failure.

If the product is recoverable material, you do not need a solid, pristine asteroid sitting intact on the lunar surface. You need useful fragments concentrated in and around a predictable crater. The crater becomes the mine.

The third inversion became the most interesting:

Arrival velocity is not really a dial.

For an object falling into the Earth-Moon system, much of the final impact speed comes from the system itself. Under favorable pursuit geometry, the physical floor is the terminal velocity (the inverse of escape velocity), about 2.4 km/s.

That number comes from celestial mechanics, not a design preference.

But the asteroid also needs to hit hard enough to shatter. A monolith that buries itself intact under meters of regolith is a problem, not a mine. The impact has to do the beneficiation: break the body into fragments distributed in and around a recoverable crater.

Separate calculations then ask what those speeds do to actual materials. The current LASSO model places the onset of whole-rock silicate melting around 7.4 km/s, with degradation occurring gradually rather than at a magical cliff.

So the speed is bounded on both sides. Too slow and the body arrives intact and inaccessible. Too fast and it melts or vaporizes. The useful corridor is the band where the body breaks apart and stays solid.

The lower bound comes from orbital constants. The upper classification comes from impact and material physics. Neither was selected to make the other convenient.

Yet the minimum physically possible lunar arrival speed sits comfortably inside the modeled ideal solid-fragment survival mining regime.

Then the field evidence converged from a separate direction. Kamil crater in southern Egypt is, as far as I have been able to determine, the only well-mapped case of a relatively slow (braked by the Earth’s atmosphere to ~ 5 km/sec) iron meteorite impact with its metal fragment distribution systematically surveyed. This nine-ton iron body left roughly 3,400 kg of recoverable metal in and around a 45-meter crater, with microscopic dust accounting for only a few percent of the surviving mass. The speed regime: the same lower band that lunar terminal velocity forces on every arriving object.

Then on August 5, 2026, a spent Falcon 9 upper stage struck the Moon at 2.43 km/s. A mostly hollow stage, not a solid iron body, so it does not prove fragment survival. But it confirmed what the lunar surface does at precisely this impact speed. The Lunar Reconnaissance Orbiter imaged the site.

Three independent lines. Orbital mechanics sets the floor. Field data defines the survival corridor. A real-world impact event calibrates the surface response. None was derived from the others, and all three pointed to the same narrow speed band.

I have an intentionally nontechnical name for the feeling when AI investigation finds separately derived constraints unexpectedly nesting like that:

AI grace.

My technical description is less romantic: serendipitous constraint convergence.

And the next IAP move is not to admire it, but to ask: How could this convergence be accidental?

Another design parameter came from such an attack rather than expansion. Redirecting an asteroid into the Earth-Moon system is not a project where safety belongs in an appendix. A hostile AI review forced the architecture to make failure benign by default.

The resulting concept uses two vehicles. The first nudges the asteroid but deliberately leaves it on a trajectory that still misses both Earth and Moon. A second spacecraft lands on it later and supplies the final corrections.

Only that second correction commits the lunar impact. If the second spacecraft fails, nothing hits anything. The pressure check did not simply delete something. It forced something new into existence: The abort corridor became the default trajectory and the plan’s backbone.

LASSO also found and states what may be its strongest failure mode. The small Earth-like asteroids that are easiest to move may be easy to move for a very bad reason: they may be fragments of the Moon already blasted into heliocentric orbit.

Every small Earth-like object with a spectrum considered in the paper, four out of four. has shown a lunar signature. The selection mechanism could therefore be structurally biased toward finding the one material the Moon already has.

Asteroid metal is the motive for LASSO. It is not yet a finding.

Even a remarkably well-timed physical calibration had to be treated carefully. On August 5, 2026, a spent Falcon 9 upper stage struck the Moon at roughly 2.43 km/s, almost exactly the velocity regime LASSO cares about.

But it was mostly hollow. It can help tell us what the lunar surface does at that impact speed: crater formation, ejecta, coupling to the target. It does not demonstrate survival of a solid asteroid metallic projectile.

By v1.1, the catalog result had also shifted from one apparently magical target to a portfolio of potentially usable trajectories. But the most instructive part of LASSO was what did not survive, not what did.

What Died

The first beautiful idea to die was the lunar ice cushion.

My original March workshop pointed toward the lunar south pole, where the first bases are planned partly to access the crater-ice deposits. If an asteroid struck a crater full of ancient lunar ice there, the energy required to melt and vaporize it could absorb part of the impact, reduce penetration, and improve material survival. Shackleton Crater, 20 km wide and 4 km deep, floored with lunar ice, would be a target.

It was exactly the kind of idea that feels wonderful when a complicated problem suddenly appears to solve itself. Then I sent the concept to my sister, Marta, and her objections were immediate: deliberately contaminating the 4-billion-year-old lunar ice deposits would be scientifically destructive and morally inexcusable. And she was right.

As I ran the idea through the loop, the crater-ice cushion deteriorated for other reasons. Lunar gravity works against containment. An oblique impact offers a cleaner way to manage deposition and struggles with crater-wall geometry. The geometry required for a slow pursuit approach makes the originally attractive polar target less convenient than it first appeared.

The ice cushion disappeared, and the architecture improved. That was an important calibration for “AI grace.” Elegant convergence is a reason to investigate something, not a reason to marry it.

The next correction was less elegant. I deposited LASSO v1.0 on Zenodo, then discovered my catalog count was wrong. One stage of the original filter chain returned 1,463 objects, not 259. The error didn't eliminate the final candidate, but it changed derived percentages and meant the public document contained a bad number.

That mattered precisely because the document was public. Errors could be fixed, so v1.1 corrections followed public v1.0 errors.

A more important conceptual correction followed. The first version required the asteroid nudge to fit within the propellant capacity of a spacecraft that had already flown, which produced a beautifully sharp result:

One target object: Asteroid 2026 AC4.

The answer looked like precision. It was actually a category error, as a flown propellant tank is not a law of physics. It is a present-day engineering and budget constraint. I had let the convenience of using contemporary capability masquerade as a physical boundary.

Remove that filter and the claim changes. Instead of “Here is the asteroid,” the result becomes: Here is a class of reachable objects across a ladder of propulsion capability.

LASSO v1.1 explicitly withdrew the propellant ceiling as a filter. The single-target result had partly been an artifact of the question I asked the database. That was not a failure of the method; Finding that out was the method.

Another absolute claim died alongside it. The original analysis treated its arrival-speed estimator as conservatively safe: it might reject useful candidates, but it would not falsely admit a bad one. Then asteroid 436724 broke the statement. Its estimated and measured encounter speeds diverged in the wrong direction. So the absolute claim was withdrawn and replaced with a screening margin.

None of this is glamorous. That is exactly why it belongs at the center of an essay about AI-assisted research. The easiest story to tell about AI is that it generates astonishing things quickly. The more important question is how to engineer the system to also make it easier to discover when those astonishing things are wrong.

A useful research partnership should not merely increase the rate at which hypotheses appear. It should increase the rate at which seductive assumptions become explicit, weak claims get weakened, bad filters get demoted, numerical mistakes get corrected, and beautiful ideas die before they become load-bearing.

That is the other advantage of the short loop. High iteration velocity is useful only if bad iterations are cheap to discard. The measure is whether the process has machinery to identify and withdraw its own flawed outputs, not how much is produced and survives.

Write The Paper, Then The World

This is why the final word in IAP is Production. Conversation is forgiving, and a chat can evolve invisibly. A versioned document has to confess what changed.

Once LASSO became a Zenodo-published preprint, every number and assumption acquired an address. So v1.0 could be broken, and v1.1 could say exactly what had been corrected, withdrawn, or structurally changed. Its Appendix C is a public accounting of those changes.

Writing also forced workshop estimates to become reproducible catalog queries. Publishing exposed where a number was measured, modeled, screened, assumed, or simply unknown.

That distinction matters because AI conversation has a powerful tendency to make all five of those sound equally fluent. But even a perfectly documented AI loop is still trapped inside an AI environment unless you force it out. Eventually, the model has to shut up, and the universe gets a vote.

For LASSO, one such vote has an unusually specific date. The small near-Earth asteroid 2026 AC4 is observable from the ground on October 14, 2026, and then not again for 29.6 years.

The asteroid 2026 AC4’s orbit currently rests on a short observational arc. Another night of astrometry would extend that arc and improve what we know about where the object will actually be during its relevant future encounter.

Whether the observation ultimately helps LASSO or hurts it is almost beside the methodological point. The loop has generated a question the AI cannot answer by becoming more eloquent.

Go look

LASSO now needs at least five expert people the AI cannot replace.

An orbital mechanician to validate the pursuit geometry and bound the screening estimator’s error margins. An impact physicist to run hydrocode simulations of solid iron arriving at 2.4 to 7.4 km/s into regolith, because the fragmentation threshold currently rests on laboratory data and a single field analog. A spectroscopist to determine whether the reachable asteroid population contains anything metallic, since the composition gap is the stated likeliest failure mode. An observer to get telescope time on October 14 and extend the orbital arc for 2026 AC4 before the window closes for 29.6 years. And a mining or materials recovery engineer to answer what robotic collection of distributed fragments on the lunar surface actually requires.

Those five people fill the three explicit free parameters the preprint carries: fragment size distribution, composition, and recoverable fraction.

The AI ran the loop fast enough to produce the document. Those five determine whether the document describes reality and how to implement it.

The Rest of the Shelf

LASSO has been through repeated cycles, a human veto, adversarial redesign, multi-model review, a public deposit, a corrected public revision, explicit free parameters, and publicly stated ways it might fail.

Most of the other ideas on my shelf have not. That distinction matters.

I made my useful Dominance Operating System political analysis framework to explain, for me, the puzzling actions of political “strong men” operators who, rather than behaving as “rational actors,” act by projecting dominance, per our conserved hominid species’ behavior. You do not negotiate, fact-check, or try to shame the silverback gorilla beating its chest in front of the troop. Finally, with the DOS model, all the puzzling political stories now made sense, and my version 2 lets one strategize and rate the responses. (Paste the Markdown version of my DOS in a chatbot and ask, “Using DOS explain this political situation, or rate and discuss Canada’s response to Trump tariffs pre and post Carney. https://zenodo.org/records/21969096.)

My prehistoric-monument work is still in expansion. Another trigger came while doing an AI-assisted project last year (with UCLA students and Atlas Global) to install a StarLink hospital internet-access project in Kenya. Kenya’s Great Rift Valley is a geographical trench that includes a massive pipeline of human evolutionary history. And, nearby seasonal animal-migration river-crossings still bottleneck and concentrate the predictable movement of harvestable protein. (https://atlas-global.org/kenya)

That revived my older question: what changes if you first treat monumental prehistoric sites as functional landscapes and systematically generate practical alternatives to features usually described in ceremonial language? (https://medium.com/@lorenrauch/sacred-pemmican-0f3d5febeecf)

These exercises have produced candidate readings, for example, a linear earthwork as a possible funnel, or recurring fox imagery as potentially reflecting a persistent scavenger problem rather than only cult significance. Those are leads, not conclusions. They still need their equivalent of Marta, expert peer review, and Appendix C.

My Ballistic Brain idea is somewhat further along because AI-assisted research suggests a specified test, though it hasn't been run yet. The broader proposition is that, over the last 2 million years, the brain developing for ballistic throwing may have been exapted for language and other higher processes that make us homo sapiens “human.” I found that the idea is not new, as versions of it go back at least to William Calvin’s work in the 1980s. Using AI research to review the existing literature, a narrower question I found is whether published changes in hominin shoulder anatomy and estimates relevant to cerebellar development have ever actually been plotted against each other across the lineage. If the trajectories track, the hypothesis gets support. If they diverge, my theory needs revision or abandonment. The useful output is not the sentence: Learning to throw grew our human brain. It is the plot that might kill the sentence. The other shelf items are earlier still, some barely hunches. That is fine.

A research program should contain ideas at different stages of survival. The mistake is pretending they have all earned the confidence of the most mature one.

What the Human Is Still For

The easiest argument about AI creativity is also the least interesting:

Who had which idea first?

The division of labor I care about is more practical.

AI is extraordinarily useful for rapid domain translation, retrieval, landscape mapping, constraint propagation, extending consequences, adversarial review, bookkeeping, and turning evolving architectures into readable artifacts.

Fresh models can also be turned against the work of the models that helped build the idea.

Humans bring different things:

1. The scoping idea.
2. What is worth spending a cycle on.
3. Experience the model was never given.
4. The inversion that comes from seeing one field through another.
5. The flinch when an answer is technically smooth but doesn't fit what you actually know.
6. The decision to keep going after an objection or to stop.
7. The willingness to let a collaborator kill the feature you like best.
8. And eventually the judgment that the document is good enough to expose to people who can challenge it.

That is not human creativity versus artificial creativity. It is an architecture designed to make their failure modes collide productively.

AI is prone to smoothness, sycophancy, and plausible fabrication.

Humans are prone to attachment, confirmation bias, disciplinary blind spots, and abandoning ideas when the vocabulary becomes unfamiliar.

Used badly, the two can amplify each other.

Used deliberately, each can make the other’s weaknesses more visible.

There is a personal symmetry with LASSO. My father, Herbert Rauch, developed the Rauch–Tung–Striebel smoothing algorithm and spent much of his career in a world of dynamic control, guidance, pursuit, and moving targets, such as missiles and satellites. His lab authored the code of the first Lockheed ballistic missile interceptors. Early in his career, he also calculated what may have been the first low-thrust Mars trajectory using a Venus gravitational assist. Discussing the idea of lunar pursuit geometry with him at dinner- of not slowing an orbiting object but synchronizing its arrival with the Moon’s orbit- would have given me the 1 km/sec assist estimation. AI doesn’t uniquely solve those questions; it just shares the abilities across domains with non-specialists.

AI made the loop fast enough, cheap enough, and corrective enough to run. AI shortened the path from the hunch to the next thing that could prove it wrong.

My question shelf isn’t empty yet, but for the first time in thirty years, it is getting shorter.

Loren Rauch, MD, MPH, is an emergency physician in Los Angeles, California, and an independent researcher.

LASSO is deposited at doi:10.5281/zenodo.21928516. Or https://zenodo.org/records/21928516.

DOS is deposited at doi: https://zenodo.org/records/21969096.

The IAP (Iterative AI Production) protocol is at: https://zenodo.org/records/21969467. And FOH at: https://zenodo.org/records/21969760.

Our Atlas Global volunteer project to give internet access to remote Kenya health centers: https://atlas-global.org/kenya

One of my Neolithic research questions is on Medium at https://medium.com/@lorenrauch/sacred-pemmican-0f3d5febeecf.