Luca Dellanna posted one of the best replies of the year on X on 10 October. Someone had argued that humans have an intuitive understanding of the world and, unlike AI, do not need to learn from images produced by others. Dellanna answered with six words: "Below, cannon ball trajectories before Galileo." ["Hier: Flugbahnen von Kanonenkugeln vor Galileo."] and the diagram reproduced below. In it, a cannon ball leaves the barrel in a perfectly straight line, labelled the violent motion; bends through a short curve, the mixed or crooked motion; and then drops vertically to the ground, the natural motion. More than 300,000 views later, the point had landed: for most of recorded history, educated people intuitively understood projectile motion, drew it carefully, taught it to each other from images in books, and drew it wrong. It is a perfect joke. It also turns out to be a far better story than the joke, because the history behind that drawing contains three twists that almost nobody laughing at it knows, and all three say something uncomfortable about how European enterprises are buying AI right now.

Figure 1. Luca Dellanna's post on X, 10 October 2026, as shared on LinkedIn. The diagram comes from Samuel Sturmy's Mariners Magazine, first published in 1669. Original post and discussion: https://x.com/DellAnnaLuca/status/2108912889784590494

This is the twenty-eighth piece in a series I have been writing for neuland.ai. Each one has pushed the same underlying argument forward: the value, the risk and the moat in enterprise AI are not in the model. They are in the layer above and around it.¹ This piece is about a different kind of model, the one in our heads, and about the only thing that has ever reliably corrected it.

A model is not judged by how intuitive it feels or how elegant it looks. It is judged by measurement, under the conditions you actually operate in.

The drawing, and the physics behind it

The diagram in Dellanna's post comes out of a tradition that goes back to Aristotle, who divided motion into "natural" motion, heavy things falling towards the earth, and "violent" motion, imposed by force. A cannon ball, in that view, travels in a straight line while the force lasts, then gives up and falls. Medieval scholars refined the idea with the notion of impetus, but the picture stayed recognisably the same: straight, then curved, then straight down. Gunners learned it, engravers illustrated it, and for a very long time it was simply how projectiles were understood.²

So far, so funny. Now the twists.

Twist one: the man with the wrong model got the right answer

In 1537, twenty-seven years before Galileo was born, the Venetian mathematician Niccolò Tartaglia published Nova Scientia, "A New Science", the first serious attempt to apply mathematics to the flight of cannon balls. It began with a practical question from a soldier: at what angle should a gun be raised to fire the furthest? Tartaglia's answer was 45 degrees.³

That is the answer every physics student learns today. But Tartaglia's model of the trajectory still used straight lines and circular arcs, and there was an error in his analysis. He reached the correct number by reasoning that was partly wrong. And he gave gunners something more useful than his theory: he encouraged them to measure their guns' elevation with a quadrant, an instrument for reading the angle off the barrel. The theory was flawed. The measurement was not.

Then comes the part I find genuinely wonderful. The friend who had asked Tartaglia the question thought 45 degrees was too steep. Tartaglia reports that the friend was convinced after some experiments. But in air, the angle for maximum range is actually less than 45 degrees, because air resistance changes the curve. The practitioner's intuition was closer to reality than the mathematician's answer, and the experiments persuaded him out of it.⁴

Twist two: the drawing everyone laughs at is closer to what cannons actually do

Here is the part that should make anyone laughing at the drawing slightly nervous. Those old trajectories, fast and nearly straight on the way out, then a curve, then a steep fall, look absurd to modern eyes. But they approximate the real motion of a projectile quite well once air resistance is taken into account. A heavy ball fired through air does not trace a clean, symmetrical arch. It flattens on the way out and drops more steeply on the way down.⁵

The replies under Dellanna's post turned into exactly this argument. Several readers defended the drawing as realistic; Dellanna countered that a ball fired at around 30 degrees does not fall vertically.² Both sides have a point. In air, the descent is steeper than the ascent, but it is not a vertical drop.

Galileo's parabola, which he published in 1638, is the trajectory a projectile would follow without air: constant speed horizontally, uniform acceleration vertically, the two combined. He established it through experiment, with his patron Guidobaldo del Monte, he reportedly rolled a ball dipped in ink across an inclined plane to record the shape of its path, and through the decisive move of splitting one confusing motion into two independent ones that could each be analysed.⁶ It was a triumph, and it is still taught first for good reason. But for an actual gunner, firing an actual cannon through actual air, it was an idealisation.

So who was right? Galileo was right about the physics in a vacuum. The old drawings were closer to what gunners saw on the battlefield. Both were models. Which one deserved trust depended entirely on the conditions under which it would be used, and the only way to find out was to measure, in the air you actually fire through.

Twist three: the drawing is not from before Galileo at all

There is a footnote to the joke, and it makes the joke better. A reader note added to Dellanna's post points out that the diagram comes from Samuel Sturmy's Mariners Magazine, a practical handbook for seafarers first published in 1669; the "Book V" printed at the foot of the image refers to its fifth volume.⁷ That is thirty-one years after Galileo published the parabola. The physics in the drawing is pre-Galilean. The drawing is not.

In other words, the right answer had been in print for three decades, and a practical manual for the people who actually fired guns at sea was still teaching the old one. Knowledge does not spread because it has been published. It spreads when the people doing the work have a reason, and a way, to check it.

Galileo also learned from images produced by others

There is one more detail, and it answers the original claim directly. Galileo owned a copy of Tartaglia's Nova Scientia, and heavily annotated another of his books. He learned geometry from Tartaglia's Italian translation of Euclid. At the University of Pisa he was taught mathematics by Ostilio Ricci, one of Tartaglia's own students.⁸

The man who broke two thousand years of intuition about projectiles was trained, quite literally, on images and texts produced by others, including some that were wrong. What made him Galileo was not a purer intuition. It was that he decomposed the problem, instrumented it, and measured it. That is the opposite of the claim Dellanna was mocking, and it is the most useful thing in this whole story for anyone responsible for an AI programme.

Now look at your AI strategy

Look at that diagram again. Three phases, neatly labelled, connected by straight lines and one reassuring curve. Violent motion. Mixed motion. Natural motion. Change the labels to Pilot. Scale. Transformation and you have the AI maturity model in roughly half the strategy decks I have seen this year: confidently drawn, beautifully annotated, and wrong about the shape of the curve. The previous piece in this series argued that enterprise value lives several steps deep in a problem, where complexity grows with every additional step rather than in tidy stages. Capability does not move through phases in straight lines. It compounds, stalls and jumps. And if the third twist feels familiar, it should: much of what works in enterprise AI architecture has been described in public for some time, and a great deal of procurement is still buying the old drawing.

The more serious parallel is the one between Galileo's parabola and the vendor benchmark table. A benchmark is a measurement taken in a vacuum: curated tasks, clean inputs, fixed conditions, chosen by whoever publishes it. Your enterprise is air. Your documents are scanned, your systems disagree with each other, your permissions vary by person, and your questions run across five sources at once. A model that leads a public leaderboard may fly beautifully in the vacuum and land somewhere else entirely in your air. An earlier piece in this series on blind evaluation made the case that the only measurement worth trusting is one taken on your own work. Tartaglia's friend would have agreed.

And there is the intuition problem itself, in both directions. Boardrooms carry two intuitions about AI at once, that it does not really understand anything, and that it is about to transform everything, and in most organisations neither has been measured. One comment on the post as it circulated on LinkedIn made a related point well: if your definition of general intelligence excludes the ordinary colleague down the corridor, the definition is the problem. The enterprise version is a double standard I see constantly. AI is held to a standard of perfection that no human process in the building meets, while demonstrations are bought with no measurement at all.

What Galileo actually did, translated

Strip away the cannon balls and Galileo's method is three moves, each of which translates directly.

Decompose. He split one tangled motion into two simple, independent ones. Enterprise problems that look intractable as a single prompt become tractable when split into steps that can each be checked: retrieval, reconciliation, calculation, judgement.

Instrument. Tartaglia's quadrant and Galileo's ink-dipped ball recorded what actually happened rather than what was believed to happen. An AI system needs the same: a record, written at the time, of what was retrieved, which tools were called and what each step produced, not the model's own summary of itself.

Measure in the air you fly in. Evaluate against your own workloads, under your own conditions, with the model's name hidden from whoever scores the result.

I would encourage every CIO reading this to ask three questions of the next AI claim that lands on their desk, whether it comes from a vendor, a consultancy or their own team. Under what conditions was this measured? What would it look like if it were wrong? And who measured it, and could they have been mistaken?

Where neuland.ai stands on this

The neuland.ai HUB is a sovereign AI management and orchestration platform, and it is built around Galileo's three moves rather than around any single model. Work is decomposed into steps that can be checked individually and run in parallel where that helps. Every retrieval, tool call and intermediate result is written to an append-only record by the platform, so that what happened can be inspected afterwards. Models are evaluated on the customer's own workloads before they are admitted to the routing table, and routing stays model-agnostic, open-weight models we host such as GLM, Mistral, Qwen and DeepSeek, and proprietary endpoints such as Claude and GPT where the customer's policy allows, so that the model that wins in a customer's air gets the work, not the one that won in a vacuum.⁹

This is also the honest answer to the question I am asked most often in first meetings: which model is the best? Honestly, it depends on your air resistance. Show me your documents, your systems and your questions, and we will measure it.

Personal take

I want to be careful with the framing here. Dellanna's post is a joke, and a very good one, aimed at a lazy argument about how humans learn. I agree with it completely. Humans learn from images produced by others; Galileo did; every engineer in your company did.

But the joke has a second target, and it is us. Many of the people laughing at that drawing would not know that it describes real cannon fire better than the parabola does, that the "correct" 45 degrees is wrong in air, or that the drawing itself was printed decades after the right answer was known. Most of us are holding a model of AI that we absorbed from images produced by others, slides, headlines, demonstrations, and that none of us has measured. That is not a failing of intelligence. It is the default condition of everyone, including Galileo. What separated him from the gunners was not intuition. It was the discipline to decompose, to instrument and to measure.

The regulatory direction asks for exactly that discipline. The EU AI Act has applied broadly since 2 August 2026, and its expectations on oversight, logging and traceability amount to a requirement that deployed systems can show their working; the Digital Omnibus has moved the high-risk Annex III obligations to 2 December 2027 and Annex I obligations to 2 August 2028.¹⁰

We at neuland.ai would rather measure the curve in the air our customers actually fly through than admire a beautiful parabola drawn in a vacuum.

1. Series articles available at https://neuland.ai/en/resources/insights.

2. Luca Dellanna, post on X, 10 October 2026, https://x.com/DellAnnaLuca/status/2108912889784590494, replying to a claim that humans, unlike AI, do not need to learn from images produced by others; see also his follow-up in the same thread on readers defending the drawing. The diagram labels three phases of a cannon ball's flight as the violent motion, the mixed or crooked motion, and the natural motion. On the Aristotelian distinction between natural and violent motion and its transmission through medieval impetus theory, see "Aristotle, projectiles and guns", https://arxiv.org/pdf/1804.00716.

3. Niccolò Tartaglia, Nova Scientia (Venice, 1537), described by historians as the founding text of mathematical ballistics; the question about the angle of maximum range was put to Tartaglia by a soldier. On the error in Tartaglia's analysis despite his reaching the correct value of 45 degrees, see EBSCO Research Starters, "Niccolò Fontana Tartaglia's The New Science", https://www.ebsco.com/research-starters/literature-and-writing/niccolo-fontana-tartaglias-new-science. On Tartaglia's promotion of the gunner's quadrant for measuring elevation, see the description of the first edition by Sophia Rare Books, https://www.sophiararebooks.com/pages/books/6517/niccolo-tartaglia/nova-scientia-inventa.

4. "Aristotle, projectiles and guns", https://arxiv.org/pdf/1804.00716, on Tartaglia's account that the friend who posed the question considered 45 degrees too great an elevation and was persuaded by experiments, and on the angle of maximum range in a resistive medium such as air being less than 45 degrees.

5. Institute of Physics, "Early perceptions of projectile motion", IOP Spark, https://spark.iop.org/early-perceptions-projectile-motion, noting that older representations of trajectories with an initial straight section, an arc and a final steep descent approximate projectile motion well when drag forces act.

6. Galileo Galilei, Discourses and Mathematical Demonstrations Relating to Two New Sciences (Leiden, 1638). On the experiments reportedly carried out with Guidobaldo del Monte, rolling an ink-coated ball across an inclined plane, see IOP Spark, citing R. H. Naylor, "Galileo's theory of projectile motion".

7. Samuel Sturmy, The Mariners Magazine (London, 1669), Book V. Identification of the diagram as noted in the reader context added to Luca Dellanna's post on X. See also David J. Frost, "The full story of pre-Galilean ballistic", https://davidjfrost.substack.com/p/the-full-story-of-pre-galilean-ballistic, which describes the volume-five illustration and observes that air resistance makes real trajectories somewhat like the older drawings.

8. On Galileo's ownership of Tartaglia's Nova Scientia and his heavily annotated copy of Quesiti et inventioni diverse, see Sophia Rare Books (as above); on his learning geometry from Tartaglia's Italian edition of Euclid and being taught at Pisa by Ostilio Ricci, a student of Tartaglia, see Famous Scientists, "Niccolò Tartaglia", https://www.famousscientists.org/niccolo-tartaglia/, and https://en.wikipedia.org/wiki/Nicolo_Tartaglia.

9. neuland.ai HUB: task decomposition into individually verifiable steps, with parallel execution where useful; append-only audit of retrievals, tool calls and intermediate results written by the platform; evaluation of models against customer workloads before admission to the routing table; model-agnostic routing across self-hosted open-weight and proprietary models; deployment on-premises, in a European sovereign cloud or in a hyperscaler region, including air-gapped operation. See also the earlier pieces in this series on blind evaluation, on provenance as a system property, and on nine loops. neuland.ai AG retains responsibility for content quality and clean delivery of results across all customer engagements.

10. Regulation (EU) 2024/1689 (AI Act), broadly applicable from 2 August 2026. Council of the EU and European Parliament provisional political agreement on the Digital Omnibus on AI, 7 May 2026: Annex III high-risk obligations postponed to 2 December 2027; Annex I obligations postponed to 2 August 2028.

11. See note 2 for the discussion of the realistic trajectory and the counterargument to a vertical descent.