AI and efficiency

OpenAI dropped an analysis this week that quietly undercuts one of the tech industry's favorite talking points. For years, the story of AI progress has been told through the lens of hardware: bigger GPUs, more chips, Moore's Law grinding forward. But OpenAI's numbers say the real engine has been something else entirely — the algorithms themselves.

Here's the specific claim. Since 2012, the amount of compute needed to train a neural network to a fixed benchmark on ImageNet classification has been halving roughly every 16 months. Do the math over that stretch and you get a 44x reduction in the compute required to hit AlexNet-level performance, the network that kicked off the deep learning boom back in 2012. Moore's Law, over that same window, would have only bought an 11x improvement. So software efficiency has outpaced hardware efficiency by a factor of four.

That's a striking reversal of the usual narrative. Chipmakers get the headlines and the stock bumps, but the researchers tweaking architectures, training tricks, and optimization methods have apparently been doing more of the heavy lifting, at least for tasks where labs have poured serious money and attention. ImageNet was exactly that kind of task in the 2010s — a magnet for competitive research, which likely explains why the gains there were so much sharper than what raw silicon improvements alone could deliver.

The implication OpenAI is pointing at, without quite spelling it out in triumphant terms, is that compute budgets and hardware roadmaps aren't the only lever worth watching. If algorithmic efficiency keeps compounding at anywhere near this rate, the cost of reaching a given level of AI capability could keep falling even if chip progress slows down or hits physical walls. That's a quietly important data point for anyone trying to forecast when certain AI capabilities become cheap enough to be everywhere.