Jason Wei Says Tools Expand Small Models but Do Not Replace Internalized Knowledge

According to multiple reports quoting a post attributed to Jason Wei, he once found the “small cognitive core plus tools” argument persuasive. A small model with adequate access to search and code could, in principle, retrieve obscure facts or perform calculations. He later argued that this theoretical possibility overlooks how much users value fast and natural performance without repeated tool calls.

Wei compared the issue to learning badminton. Being able to perform each movement in isolation does not guarantee that a player can combine those movements smoothly and consistently during a real match. Similarly, a model that has internalized patterns and knowledge can respond more directly and produce broader judgments than a smaller model that must repeatedly search, interpret, and assemble information.

He also argued that repeated retrieval and derivation can reduce reliability in long-horizon tasks because mistakes may compound. Tools therefore remain valuable for extending small-model capabilities, but he believes users seeking the highest-quality intelligence will continue to benefit from larger models. This is an expert argument, not conclusive evidence that larger models dominate every task.