For those of us programming nerds that want to play with aerodynamics, I can't recommend AeroSandbox enough. While the code is pretty obviously written for people who know their way around aerodynamics and not so much around programming, it is remarkably powerful. You can do all sorts of aerodynamic simulations and is coupled with optimization libraries that allow you to do incredible aerodynamic optimizations. It comes included with some pretty powerful open weight neural network models that can do very accurate estimates of aerodynamic characteristics of airfoils in a fraction of the time that top tier heuristic solvers (like xfoil) can do (which are already several orders of magnitude faster than CFD solvers).
He usually posts these brilliant explanations once or twice a year but nothing in 2025. I hope he finds the time to continue because the lessons are really really brilliantly told.
Ok that's long, one top line thing people tend to miss in these flying explanations is that airfoil shape isn't about some special sauce generating lift. A flat plate generates any amount of lift you want just fine. Airfoil design is about the ratio of lift to drag most importantly and then several more complex effects but NOT just generating lift. (stall speed, performance near and above the speed of sound, laminar/turbulent flow in different situations, what you can fit inside the wing, etc)
Exactly. Airfoil is an optimization. There is a common misconception that planes wouldn't get off the ground if you didn't have airfoil. No, most of the lift (depends on the plane but in the ballpark of 80-90%) comes from the overall shape of the wings. ~20% is from leading edge airfoil deflection dynamics.
And if, say, airfoil was never discovered, we'd probably design the whole wing slightly differently to compensate for it, so the actual difference wouldn't even be 20%.
Airfoil is about as important as winglets, and planes fly without winglets just fine. But nobody points to winglets and says that's the crucial bit that makes the whole thing work.
Umm no, at zero degrees AoA as the first diagram on the page shows, a flat plate does not generate lift.
But nobody actually questions that a flat shape can generate lift; we all made paper planes as a kid.
But every airfoil has an equilibrium angle of attack (not always stable with velocity) where it generates zero lift. The chordal angle of attack is for convenience because it depends only on airfoil geometry and not ambient velocity, but it isn't a fundamental physical property of the airfoil.
If we treat the angle where zero lift is generated as the base angle for an airfoil, then all airfoils generate lift depending on their angle relative to that, including a flat plane. As the GP says, other properties are the dominant factor in airfoil geometry.
When introducing airfoils I think it is more useful to start from a plane than a traditional airfoil shape; the math and intuitions are much clearer from there.
And with steady level flight symmetrical airfoils are flown at an angle, a cambered airfoil shape being flown at 0 degrees angle of attack vs its chord line would be an unusual coincidence. Wings are mounted at a small angle relative to the direction of thrust and what one would define as a flat line on the fuselage.
Uncambered airfoils also don't generate lift at zero degrees. What constitutes "0" for curved airfoils is convenience. You want lift, you put a flat plate on an angle, anything fancier is for Lift/Drag, Thrust/Weight, etc.
There are about a million places incorrectly "explaining" that airfoils create lift because the top path is longer and this means the air has to go faster. A flat plate would not create lift in that case. The fact that paper airplanes obviously can fly somehow never stops people from repeating this.
It’s kind of sad IMO. Bartosz has made a ton of these super interesting and meticulously designed explainers. Something thrown together with AI is much more likely to be made by someone who doesn’t know what they’re talking about, and I’m worried that the sheer volume will crowd out actually quality content like this.
Don't think so, and we should stop spread damaging narrative like this. I'd say it's already able to imitate this kind of explainers(badly) thanks to his training data. All the subtle teaching nuances, effort, know-how and visual creativity that people like Bartosz Ciechanowski put on this kind of work is not reproducible if not statistically imitating it
In order to be taken serious I feel like statements like this need to be qualified with who the claimant is imagining to be responsible for generating the anticipated output. The ‘A’ in AI isn’t for ‘autonomous’.
Bartosz Ciechanowski could generate an explainer like this using Claude today if he wanted to. But would he? If someone like him had the mind to do it then they could instead. But where’s it at? These types may hold themselves to a standard above this method. No shame in that.
Good rule of thumb: it should take less time to consume content than it does to create it.
I don’t know how long it takes Ciechanowski to create these explainers, probably a few months? It shows and it’s well worth spending your time reading through his content meticulously.
How long does it take for an LLM to crap out an equivalent explainer? 60 seconds? You should be spending less time than that reading it.
I think it's actually already there. It's definitely possible to make these sorts of explainers with something like a Claude Code, you just have to spend a fair amount of time making sure that it's actually doing what you expect it to do. I think the biggest danger with something like a Claude Code is that you get something that looks functionally correct but that the details are suddenly wrong on. I wrote a little bit about this on my blog for some of the places that I've done visualizations actually, and I think it's remarkably easy to iterate on them now.
It's been said before, but this prediction isn't amazing, imo.
I look forward to Bartosz's articles because they're rock-solid sources of information and the visualizations are both easy-to-understand and surprisingly light on performance. It's all shockingly digestible.
Honestly, as popular science writing goes, this is art as far as I'm concerned, and art is best when it comes from a place of passion and conviction, something AI will never be able to reproduce.
This is absolutely amazing.
For those of us programming nerds that want to play with aerodynamics, I can't recommend AeroSandbox enough. While the code is pretty obviously written for people who know their way around aerodynamics and not so much around programming, it is remarkably powerful. You can do all sorts of aerodynamic simulations and is coupled with optimization libraries that allow you to do incredible aerodynamic optimizations. It comes included with some pretty powerful open weight neural network models that can do very accurate estimates of aerodynamic characteristics of airfoils in a fraction of the time that top tier heuristic solvers (like xfoil) can do (which are already several orders of magnitude faster than CFD solvers).
https://github.com/peterdsharpe/AeroSandbox
Oh man. This guy. His work is simply some of the best explanation content I have ever seen
He usually posts these brilliant explanations once or twice a year but nothing in 2025. I hope he finds the time to continue because the lessons are really really brilliantly told.
https://news.ycombinator.com/item?id=39526057
That's the missing course for the first year of any Aerospace Engineering faculty.
Should be (2024).
I was excited for a moment, and wondered why the RSS feed didn't work, but later realized that the article is from 2024.
Bartosz Ciechanowski, the gift that keeps on giving.
I wish there was an infinite number of blogs that where this good.
Ok that's long, one top line thing people tend to miss in these flying explanations is that airfoil shape isn't about some special sauce generating lift. A flat plate generates any amount of lift you want just fine. Airfoil design is about the ratio of lift to drag most importantly and then several more complex effects but NOT just generating lift. (stall speed, performance near and above the speed of sound, laminar/turbulent flow in different situations, what you can fit inside the wing, etc)
Exactly. Airfoil is an optimization. There is a common misconception that planes wouldn't get off the ground if you didn't have airfoil. No, most of the lift (depends on the plane but in the ballpark of 80-90%) comes from the overall shape of the wings. ~20% is from leading edge airfoil deflection dynamics.
And if, say, airfoil was never discovered, we'd probably design the whole wing slightly differently to compensate for it, so the actual difference wouldn't even be 20%.
Airfoil is about as important as winglets, and planes fly without winglets just fine. But nobody points to winglets and says that's the crucial bit that makes the whole thing work.
Was gonna say where is the debate of bernouli vs. AoA/pforce (p-factor), scatter blast shotgun hitting bottom of wing
Umm no, at zero degrees AoA as the first diagram on the page shows, a flat plate does not generate lift. But nobody actually questions that a flat shape can generate lift; we all made paper planes as a kid.
But every airfoil has an equilibrium angle of attack (not always stable with velocity) where it generates zero lift. The chordal angle of attack is for convenience because it depends only on airfoil geometry and not ambient velocity, but it isn't a fundamental physical property of the airfoil.
If we treat the angle where zero lift is generated as the base angle for an airfoil, then all airfoils generate lift depending on their angle relative to that, including a flat plane. As the GP says, other properties are the dominant factor in airfoil geometry.
When introducing airfoils I think it is more useful to start from a plane than a traditional airfoil shape; the math and intuitions are much clearer from there.
And with steady level flight symmetrical airfoils are flown at an angle, a cambered airfoil shape being flown at 0 degrees angle of attack vs its chord line would be an unusual coincidence. Wings are mounted at a small angle relative to the direction of thrust and what one would define as a flat line on the fuselage.
Paper airplanes do not have barndoor wings, though. Most of them have stepped camber through the way they are folded.
It's not finished but I started writing this to clarify: https://entropicthoughts.com/paper-airplane-aerodynamic-stab...
Scroll down to "trim and angle of attack".
(I hope there's nothing embarrassing in there. It's an old, early draft.)
Uncambered airfoils also don't generate lift at zero degrees. What constitutes "0" for curved airfoils is convenience. You want lift, you put a flat plate on an angle, anything fancier is for Lift/Drag, Thrust/Weight, etc.
There are about a million places incorrectly "explaining" that airfoils create lift because the top path is longer and this means the air has to go faster. A flat plate would not create lift in that case. The fact that paper airplanes obviously can fly somehow never stops people from repeating this.
Where can I find more articles where things are explained in this manner?
The author of the Airfoil article has made several similar tutorials on various topics:
https://ciechanow.ski/archives/
For machine learning, Distill.pub has some excellent hands-on tutorials. For example, here's one on momentum:
https://distill.pub/2017/momentum/
https://explorabl.es collects various
https://ciechanow.ski/archives/
wait what? this is goood!
I was just thinking the other day about how AI will pretty soon be able to create this kind of explainers on everything quite quickly.
Amazing times!
It’s kind of sad IMO. Bartosz has made a ton of these super interesting and meticulously designed explainers. Something thrown together with AI is much more likely to be made by someone who doesn’t know what they’re talking about, and I’m worried that the sheer volume will crowd out actually quality content like this.
> and I’m worried that the sheer volume will crowd out actually quality content like this.
It's a valid fear.
Don't think so, and we should stop spread damaging narrative like this. I'd say it's already able to imitate this kind of explainers(badly) thanks to his training data. All the subtle teaching nuances, effort, know-how and visual creativity that people like Bartosz Ciechanowski put on this kind of work is not reproducible if not statistically imitating it
And the usual corollary: Not just thanks to his training data, but because training data of that kind and for this kind of topic - still - exists.
Exactly and him not publishing any new post in 2025 makes me wonder...
In order to be taken serious I feel like statements like this need to be qualified with who the claimant is imagining to be responsible for generating the anticipated output. The ‘A’ in AI isn’t for ‘autonomous’.
Bartosz Ciechanowski could generate an explainer like this using Claude today if he wanted to. But would he? If someone like him had the mind to do it then they could instead. But where’s it at? These types may hold themselves to a standard above this method. No shame in that.
Good rule of thumb: it should take less time to consume content than it does to create it.
I don’t know how long it takes Ciechanowski to create these explainers, probably a few months? It shows and it’s well worth spending your time reading through his content meticulously.
How long does it take for an LLM to crap out an equivalent explainer? 60 seconds? You should be spending less time than that reading it.
Haven't people been saying this since 2023? Yet to see AI build this kind of stuff "quite quickly".
I think it's actually already there. It's definitely possible to make these sorts of explainers with something like a Claude Code, you just have to spend a fair amount of time making sure that it's actually doing what you expect it to do. I think the biggest danger with something like a Claude Code is that you get something that looks functionally correct but that the details are suddenly wrong on. I wrote a little bit about this on my blog for some of the places that I've done visualizations actually, and I think it's remarkably easy to iterate on them now.
https://estsauver.com/blog/scaling-visualizations
Hey, that's a pretty great article! Thanks for sharing.
(I hope you don't get downvoted by Chichanowski's fanboys. Sad to see people being against innovation, on this site of all places.)
I think it's only a matter of time, AI history has been a cycle of "yeah, but it will never do this", then literal weeks later it does it, lol.
We should think about how each part of the iteration cycle you describe can be improved. This is definitely a problem that can be solved!
It's been said before, but this prediction isn't amazing, imo.
I look forward to Bartosz's articles because they're rock-solid sources of information and the visualizations are both easy-to-understand and surprisingly light on performance. It's all shockingly digestible.
Honestly, as popular science writing goes, this is art as far as I'm concerned, and art is best when it comes from a place of passion and conviction, something AI will never be able to reproduce.