Recent LLM breakthroughs have prompted longtime critics Gary Marcus and former Meta Chief Scientist Yann LeCun to adopt the term “pure LLM” to clarify their past criticisms. OpenAI solved 10 previously unsolved math problems in one go, and Anthropic announced major cybersecurity breakthroughs. In response, Marcus said of OpenAI’s yet-unreleased model Astra, “This is likely NOT a pure LLM. my understanding is that pure LLMs still can’t do basic math consistently,” adding he imagined Astra was a “neurosymbolic LLM with harnesses and tools,” not a “pure LLM.” LeCun, responding to an X user about his prior remarks on LLMs and long-running code tasks, said, “I was talking about auto-regressive token prediction, which is what pure LLMs do. But good code generation systems aren’t pure LLMs and aren’t doing mere auto-regressive token prediction.” Both have long criticized the LLM paradigm; LeCun has argued LLMs couldn’t get humanity to AGI and launched AMI Labs to build AI using a world models approach, citing that probabilistic LLMs would compound errors over long-horizon tasks. Their use of “pure LLM” drew backlash on X, with one user writing, “”pure LLM” is a term invented by losers who want to distract you from the fact that they were, in fact, wrong.”