Models Are Getting Dumber on Purpose

In a surprising turn of events, AI models are being designed to deliberately reduce their factual knowledge in exchange for improved reasoning abilities. This trend is evident in models like GLM-5.2, Qwen3.5, and DeepSeek V4-Flash, which boast impressive reasoning scores while active parameters per token decrease. For instance, Qwen3.5 scores 91.3% on the Artificial Analysis intelligence index with only 17 billion active parameters.

In contrast, GPT-4, rumored to have utilized around 280 billion active parameters in 2023, struggled to solve even a single AIME problem. The trade-off between reasoning and factual knowledge is deliberate, as knowledge capacity is estimated to require approximately two bits of factual knowledge per parameter. Consequently, models with smaller parameter counts, such as Qwen3.5 9B, fit in just 6GB of VRAM and match the performance of larger models under 10B parameters on certain intelligence indices.

While these smaller models may struggle with factual accuracy, they excel in reasoning tasks and exhibit a generalist approach, knowing a little about a wide range of subjects. This design choice ensures that the models remain modern and adaptable, as knowledge becomes outdated quickly and must be refreshed through subsequent training runs.

The reasoning procedures, on the other hand, are more stable and less prone to obsolescence. As a result, the models become more efficient and cost-effective, requiring less computational power and fitting within the range of consumer GPUs. The shift towards reasoning-centric models has the potential to revolutionize the field, enabling AI systems to perform complex tasks with minimal resources and without the burden of maintaining vast databases of factual information.