Large Language Models (LLMs) have become incredibly powerful tools, capable of generating human-like text and even passing complex linguistic tests. But are they truly general learners, like humans? A new research paper, "No Such Thing as a General Learner: Language models and their dual optimization," argues that the answer is a resounding no. The paper's authors, Emmanuel Chemla and Ryan M. Nefdt, highlight the dual optimization process that shapes LLMs. First, there's the training phase, where models learn from massive datasets, much like humans acquire language. But unlike humans, LLMs undergo a second optimization: a selection process akin to natural selection. The "fittest" models, those that perform best on benchmarks, survive and are further refined. This engineered evolution makes LLMs highly specialized learners, optimized for specific tasks and benchmarks, rather than general-purpose learners like humans. This distinction, the authors argue, has significant implications for how we interpret LLM performance and its relevance to human cognition. For example, the fact that LLMs can learn certain linguistic patterns doesn't necessarily invalidate nativist theories of language acquisition, as these models have been explicitly optimized for such tasks. Furthermore, benchmarks, while useful for evaluating progress, also play a role in shaping LLM development, becoming implicit objectives in the optimization process. The research also delves into the fascinating topic of "impossible languages"—languages that violate the fundamental rules governing human languages. While some claim LLMs can learn these impossible languages just as well as human languages, Chemla and Nefdt point to research suggesting otherwise. LLMs trained on human languages show a preference for human-like grammatical structures. However, the authors caution against overinterpreting this finding, emphasizing the tailored nature of LLMs. Importantly, the paper reveals that LLMs often develop their own "secret languages"—patterns and structures that are nonsensical to humans but enhance their performance on specific tasks. This further underscores the specialized nature of LLMs. In conclusion, while LLMs are powerful tools, their engineered nature limits their role as models of human cognition. They are highly optimized for specific tasks, far from the general learning abilities of humans.