The "Omniscience" accuracy benchmark, which measures pure factual knowledge, tracks more closely with a model's total parameters than any other metric. This suggests embedded knowledge is a direct function of model size, distinct from reasoning abilities developed via training techniques.
Related Insights
LLMs learn two things from pre-training: factual knowledge and intelligent algorithms (the "cognitive core"). Karpathy argues the vast memorized knowledge is a hindrance, making models rely on memory instead of reasoning. The goal should be to strip away this knowledge to create a pure, problem-solving cognitive entity.
Traditional benchmarks often reward guessing. Artificial Analysis's "Omniscience Index" changes the incentive by subtracting points for wrong answers but not for "I don't know" responses. This encourages models to demonstrate calibration instead of fabricating facts.
The "bitter lesson" in AI research posits that methods leveraging massive computation scale better and ultimately win out over approaches that rely on human-designed domain knowledge or clever shortcuts, favoring scale over ingenuity.
Google's Titans architecture for LLMs mimics human memory by applying Claude Shannon's information theory. It scans vast data streams and identifies "surprise"—statistically unexpected or rare information relative to its training data. This novel data is then prioritized for long-term memory, preventing clutter from irrelevant information.
Artificial Analysis found that a model's ability to recall facts is a strong function of its total size, even for sparse Mixture-of-Experts (MoE) models. This suggests that the vast number of "inactive" parameters in MoE architectures contribute significantly to the model's overall knowledge base, not just the active ones per token.
The binary distinction between "reasoning" and "non-reasoning" models is becoming obsolete. The more critical metric is now "token efficiency"—a model's ability to use more tokens only when a task's difficulty requires it. This dynamic token usage is a key differentiator for cost and performance.
Unlike humans, whose poor memory forces them to generalize and find patterns, LLMs are incredibly good at memorization. Karpathy argues this is a flaw. It distracts them with recalling specific training documents instead of focusing on the underlying, generalizable algorithms of thought, hindering true understanding.
Traditional benchmarks incentivize guessing by only rewarding correct answers. The Omniscience Index directly combats hallucination by subtracting points for incorrect factual answers. This creates a powerful incentive for model developers to train their systems to admit when they lack knowledge, improving reliability.
Data from benchmarks shows an MoE model's performance is more correlated with its total parameter count than its active parameter count. With models like Kimi K2 running at just 3% active parameters, this suggests there is still significant room to increase sparsity and efficiency.
To improve LLM reasoning, researchers feed them data that inherently contains structured logic. Training on computer code was an early breakthrough, as it teaches patterns of reasoning far beyond coding itself. Textbooks are another key source for building smaller, effective models.
