Counting Orders of Ten

People often look at a child every day and do not see the child growing. The shirt still fits, the voice is still familiar, the hand still rests inside an adult’s hand. Only when they open an old photograph do they notice that a few years have already crossed that face

AI is easy to misread in the same way. A wrong answer, a clumsy calculation, or an image with too many fingers makes people conclude that machines are still far from human intelligence. Those errors are real, but they do not tell the whole speed of the movement. To see more clearly, Leopold suggests counting orders of ten, each time the resources increase tenfold

In this view, AI capability is lifted by three currents meeting. The first is the compute used for training. The second is algorithmic improvement, which lets the same machines produce better results. The last is the removal of limits around the model, giving it more time to think, more tools, more memory, and the ability to check its own work

Each current on its own may not look strange. A faster chip, a better training method, or a longer memory is only a technical improvement. When they multiply together, the result no longer resembles a car going a few more kilometers per hour. It resembles the road, the engine, and the driver all changing in a single night

The easiest place to get confused is the word multiply. Human intuition is used to straight lines. We imagine next year a little better than this year, like adding one more floor to a house. But orders of ten are more like a seed that, after every season, makes many more seeds. The first steps look small, until one day the garden is full and no one remembers the exact moment it became large

That argument does not guarantee that more machines will produce unbounded intelligence. Data can run out, electricity can fall short, algorithmic progress can slow, and beautiful benchmarks do not always become useful ability in the world. Counting orders of ten is only a way to avoid the opposite mistake: looking at today’s flaws and assuming tomorrow will keep the same shape

He thinks the value of this way of counting is that it forces us to see AI as a process, not a product. A model will arrive and be replaced. An application will rise and grow quiet. Under those names, compute, algorithms, and the ability to use tools are still flowing in the same direction. If the three currents keep rising, should we look at today’s ruler, or at the waterline climbing the wall

❤️ cowriter aethery