How powerful will the intelligent iteration of large models be over the next three years?

Breaking it down, there are actually two variables driving this: the exponential growth of hardware computing power, combined with the scaling laws of the models themselves.

On the chip side, computing capacity can roughly quadruple every 18 months—this is the industry-recognized pace. From the training data perspective, based on current scaling laws, the available high-quality data can support about 2-3 such iteration cycles.

Calculating this way, by 2029, the intelligence level of large models will have increased 16 times compared to now. This number sounds a bit crazy.

For emerging fields like Web3 and AI infrastructure, this kind of capability boost could be key to accelerating application deployment. But at the same time, it also means we need to think earlier about ethical issues and risk prevention.

What’s your view? Are you excited about these possibilities, or a bit worried?