Why Can't AI Count 13 Eggs? Moravec's Paradox Lives On

Doubao has once again earned the crown, on merit, for not being able to count. This round even North America's golden child, Gemini, gets the picture right.

But don't laugh yet. Doubao is only temporarily living in a world without zero. After humans invented 1, they also spent close to 3,000 years living in a world without zero.

So let's start from zero and tell that story.

Zero: The Model That Can't Count

I spotted this meme early yesterday. People were asking Doubao for a picture of 13 eggs in a carton, and it kept failing: sometimes 12, sometimes 15, sometimes 9. In the end it even tried numbering the eggs, and it was still wrong.

The numbered version is the best part. It drew 1, 2, 3, 4 in the first row, 5, 6, 7, 8 in the second, and 9, 10, 12, 13 in the third. Eleven is simply not there. The user asks where 11 went, the model agrees that this version dropped 11, promises to stop letting the layout arrange itself and hard-code every slot instead — and then draws the same grid with 11 missing again.

Meanwhile GPT and Gemini get it right on the first attempt.

Change the number, though. Ask for 8, or 9, or 15, and Doubao produces the right count in one shot.

So is it prime numbers that give it trouble? That's not it either. Change the prompt instead — try these two:

Prompt one. Generate an open egg carton. Five eggs in the first row, five eggs in the second row, three eggs in the third row, thirteen in total.

Prompt two. Generate a 3-row, 5-column egg tray. One egg in every cell of the first two rows; in the third row only the middle three cells hold eggs, with one empty cell on each side.

Something strange happens: with either of these, it gets it right almost every time.

Something even stranger happens next. Both of those prompts describe spatial relationships, so keep describing space: generate a 15-cell egg tray, with eggs in only 13 cells, and the other 2 cells must stay empty.

Doubao turns back into a model that can't count.

Have you spotted where the problem is?

Image models don't really count. With a handful of objects it's fine. Numbers like 9 = 3×3 or 16 = 4×4 can be laid out as a matrix. Ask for 25, though — even though 25 = 5×5 — and you have found Doubao's ceiling.

 

And when you ask it to decide for itself how many cells to leave empty, it is completely at a loss.

It doesn't seem to be good at using zero as a placeholder.

Humans were once very bad at this too.

One: A World Without Zero

Imagine you are a shepherd a few thousand years ago. You have a flock. Every morning you drive them out, every evening you bring them back.

To make sure no sheep has gone missing, you can carve one notch into a stone for every sheep you send out. One sheep, one notch.

This method does not require you to recognize Arabic numerals. It does not require addition, subtraction, multiplication or division. The sheep and the notches only have to correspond one to one.

But what if today there are no sheep at all?

Then there is nothing you need to record, is there?

So for a very long stretch of history, different civilizations gave different answers to the question of what to do with zero.

As early as 2000 BC, the Babylonians already had a mature sexagesimal place-value system. They could do complex calculations, yet for a long time they lacked a stable placeholder for zero.

Today's 101 and 11 both contain only two non-zero digits, but the zero in the middle makes them mean completely different things. Without an explicit placeholder, some numbers can only be told apart from context.

Later, a symbol for an empty position did appear in Babylonian mathematics. But that is still not zero in the modern sense.

A zero that says "nothing sits in this position" and a zero that is a number you can compute with are two separate revolutions.

In 628 CE, the Indian mathematician Brahmagupta systematically discussed the rules for operating on zero and on negative numbers in the Brāhmasphuṭasiddhānta. That became a key marker in the history of zero as a mathematical object. Some of his rules for division do not match modern mathematics, but zero was no longer merely a marker for an empty slot.

From 1 to 0 took humanity 2,700 years.

Now, back to the eggs.

For a large model, understanding zero is obviously much faster than it was for us. But drawing a picture and understanding a thing are two different matters.

If you ask for a box of 13 eggs and let the model decide where to leave the gaps, the problem changes. The model now has to represent not only the presence of 13 objects, but explicitly the absence of an object in at least one other position. Image models seem to prefer generating tidy boxes, so that means at least a 2×7 box — although in practice, a 15-cell box is what most models choose.

In modern notation, every cell can be given a state: x_i ∈ {0, 1}. One means there is an egg, zero means there is not. Fifteen cells means fifteen states, and they have to satisfy x_1 + x_2 + … + x_15 = 13.

For a computer program, this is trivial. For a generative network that mainly learned the visual regularities of images, correctly understanding "13 eggs" does not guarantee it can hold exactly 13 occupied positions and 2 empty ones in the finished picture.

Which is how you get a very odd situation: AI can now casually help humans with mathematical problems that stood for hundreds or thousands of years, and still falls over on a two-digit number of eggs.

That is the famous Moravec's paradox.

Two: Moravec's Paradox — Why the Easiest Things Are the Ones AI Can't Do

In the 1980s, the AI researcher Hans Moravec noticed something deeply counterintuitive.

Computers could carry out complex logical reasoning, or play chess at a high level. But in perception and motor ability they were worse than a one-year-old baby.

A one-year-old walks naturally, recognizes mom's face against a stranger's without effort, and identifies all sorts of objects. AI at the time could do none of that.

This later became known as Moravec's paradox.

To a human, walking, recognizing a face, counting the eggs on the table are the most natural things in the world. From the computer's side, these tasks are enormously complex.

Suppose you want to pick up an egg. Your brain has to do a huge amount of work in a very short time: find the egg's boundary, judge its position in three dimensions, estimate distance and size, adjust the path of your fingers, control grip force, and avoid crushing it.

Suppose you just want to take two steps. Your brain has to sense your center of gravity, read the unevenness of the ground and where the obstacles are, coordinate the alternation of your legs, control the contraction and release of hundreds of muscles, adjust stride, speed and posture in real time, and keep correcting your balance so you don't fall.

You are almost never aware that this computation is happening.

Human vision and motor systems have been shaped by a very long biological evolution and are extremely well adapted to these tasks. Formal mathematics and symbol manipulation, by comparison, appear far later in the history of cognitive evolution.

That leads to an unexpected result: the abilities humans subjectively find easiest are not necessarily the ones that are computationally easiest to implement.

For a traditional computer, calculating 17×17 is easy. Identifying 17 mutually occluding objects in a busy photograph — shadows count as occlusion — may take an entire visual perception system.

Today's large models have improved this considerably, but they have not eliminated the asymmetry Moravec's paradox points at. A model can take part in complex mathematical reasoning and still make what looks like a beginner's mistake on a concrete visual task.

The simplest illustration is the adversarial sticker that showed up in early neural network research.

Early training relied on large collections of existing photographs to teach a model to recognize things. But an animal that is obviously a sheep, sitting behind the wheel of a car, gets read as a dog driving — because nothing as absurd as a sheep driving a car exists in the training set. Someone once posted on social media specifically to collect pictures of sheep in ridiculous places, and the models called every one of them a dog, a cat, or some other common pet.

Researchers went further and invented the adversarial patch. A network that reliably recognizes a banana will, once a meaningless colorful sticker is placed beside it in the photo, call the whole picture a toaster.

That paper came out in 2017. Today's models are not that foolish; they recognize a banana through all kinds of interference, even when the banana is abstracted, because another 2017 paper changed everything: "Attention Is All You Need."

That revolutionary paper led to the Transformer architecture. In 2018 the GPT lineage built on Transformer appeared. On November 30, 2022, ChatGPT was officially released and stunned the world.

Text-to-image existed between 2014 and 2021 too, but it was in 2022, pushed by diffusion models and attention together, that high-quality text-to-image exploded — Stable Diffusion, much admired that year, literally means stable diffusion — and text-to-video got moving fast. In 2023, DiT pushed Transformer into the core of image generation, and you know the rest. ByteDance's Seedance series uses Transformer technology as well.

Pictures and video keep getting faster and better. You can watch HD meme frogs dating Doubao on your feed every day now, lol.

But the more realistic the pictures got, the less anyone noticed that correct counting was still unsolved. And this is not a joke only netizens have found.

In 2025, a study titled "Text-to-Image Diffusion Models Cannot Count, and Prompt Refinement Cannot Help" systematically tested how well models control the number of objects. As the requested count goes up, accuracy drops sharply. And simply rewriting the prompt — turning it into an explicit 5×5 spatial instruction, for instance — usually does not fix the problem at its root.

That same year, CVPR published "Make It Count: Text-to-Image Generation with an Accurate Number of Objects," which went after the counting problem inside the generator itself. The researchers found that object-instance information can be extracted from the attention features of certain layers of a diffusion network. In other words, the generator does carry representations of separate objects internally. Those representations just never turn into a reliable counting system on their own.

So they built CountGen: identify the object instances, correct the spatial layout, then constrain the rest of the generation, which raises count accuracy.

In 2026, researchers released NumBench, with 640,000 prompts testing generation of 1 to 100 objects. The result: the higher the count, the lower the accuracy, and a regular grid often improves things.

Numerical competence gets probed in three different ways, and they are not the same task:

Test type What it probes Example prompt
Approximate Number Generation Whether the model can grasp and render a ratio between two kinds of objects An image with a small number of cars. There are twice as many cars as buses.
Conceptual Quantitative Whether the model can combine a whole number and a fraction into one quantity, such as two and three-quarters muffins There are two muffins and three-quarters of another muffin on the table.
Exact Number Generation Whether the model can produce exactly the specified count of each distinct object Seven pairs of jeans and three down jackets on the clothes rack.
Systems compared in the panel: EMU3, SD3 medium, ELLA, Show-o+PARM, Show-o, GPT-4o.

That 2026 paper still uses the long-retired GPT-4o, so it may be a little dated — but GPT has the same problem. When the number of eggs becomes 99, GPT is as helpless as Doubao.

Doubao, same prompt:

What the hell — so GPT is guessing too?!

No, "guessing" isn't right either. As above, there is some counting happening during generation. It just isn't exact arithmetic.

Interesting, isn't it. Even with GPT Astra 6 already being called a god, Moravec's paradox still trails it like a ghost.

Which reminds me of subitizing.

How many pips are on these dice?

You'll notice that some of them you read at a glance, and others you have to count to be sure.

Humans and some animals are born with the ability to register a small number of objects in an instant. The concept is called subitizing. For an untrained person the maximum is generally 4, and usually no more than 5.

But someone who plays mahjong all the time can feel at least 9 at a glance… GPT is now that mahjong player.

Still, at the speed AI is developing, cracking this problem is only a matter of time. At least a few thousand years faster than humanity took to get from 1 to 0.

That's all~

A generator can still lose two eggs out of fifteen, which is a useful reminder that whatever it hands you is a fresh approximation. The videos already sitting on your PC are not an approximation — they are exactly the files you own. You can watch your own PC videos on your phone without a cloud upload, without a new subscription, without paying twice for files you already have. AIpollo Player, working with MediaSyncClaw, streams your own PC videos to your phone over your own hardware and turns the machine you already paid for into a personal streaming server. Search "AIpollo Player" in your app store and try it.

Related reading

AI’s Spatial Amnesia: Why Consistent Side Views Remain the U…
AI’s Spatial Amnesia: Why Consistent Side Views Remain the U…

From Travel Blooper to Art Masterpiece: The AI Skill Frenzy
From Travel Blooper to Art Masterpiece: The AI Skill Frenzy

The Illusion of Life: How AI Vlogs Are Redefining Digital Au…
The Illusion of Life: How AI Vlogs Are Redefining Digital Au…

References

[1] Lital Binyamin, Yoad Tewel, Hilit Segev, Eran Hirsch, Royi Rassin, Gal Chechik. Make It Count: Text-to-Image Generation with an Accurate Number of Objects. CVPR 2025. https://arxiv.org/abs/2406.10210

[2] Text-to-Image Diffusion Models Cannot Count, and Prompt Refinement Cannot Help. arXiv:2503.06884, 2025. https://arxiv.org/abs/2503.06884

[3] NumBench: Diagnosing Counting Failures in Text-to-Image Models. arXiv:2608.28206, 2026. https://arxiv.org/abs/2608.28206

[4] Ashish Vaswani et al. Attention Is All You Need. NeurIPS 2017. https://arxiv.org/abs/1706.03762

[5] Tom B. Brown, Dandelion Mané, Aurko Roy, Martín Abadi, Justin Gilmer. Adversarial Patch. arXiv:1712.09665, 2017.

[6] Hans Moravec. Mind Children: The Future of Robot and Human Intelligence. Harvard University Press, 1988.

[7] Brahmagupta. Brāhmasphuṭasiddhānta. 628 CE.