The Era of Large AI Models Is Over
Size doesn’t matter
Just kidding: we all know size matters. It’s definitely true for AI models, especially for those trained on text data, i.e., language models (LMs). If there’s one trend that has, above all others, dominated AI in the last five or six years, it is the steady increase in parameter count of the best models, which I’ve seen referred to as Moore’s law for large LMs. The GPT family is the clearest — albeit not the only — embodiment of this fact: GPT-2 was 1.5 billion parameters, GPT-3 was 175 billion, ~100x its predecessor, and rumors have it that GPT-4’s size, although officially undisclosed, has reached the 1 trillion mark. Not an exponential curve but definitely a growing one.
OpenAI was categorically following the godsend guidance of the scaling laws they themselves discovered in 2020 (that DeepMind later refined in 2022). The main takeaway is that size matters a lot. DeepMind revealed that other variables like the amount of training data, or its quality, also influence performance. But a truth we can’t deny is that we love nothing more than a bigger thing: Model size has been the gold standard for heuristically measuring how good an AI system would be.
OpenAI and DeepMind have been making their models bigger over the years in search of hints from the performance graphs, signs in the benchmark results…
