The week in AI, September 22-28, 2025

optimism

Large language models absorb the worldview present in their training data. Trying to completely remove bias is impossible because every dataset reflects the perspectives and priorities of its creators and sources. Research in 2025 supports this idea. For example, an MIT study on unpacking LLM bias found that scaling up models amplifies small imbalances in pretraining data, which ultimately produces entrenched opinions on topics ranging from politics to environmental impacts. Similarly, a PNAS paper on AI bias showed that models often favor AI generated content over human written material, suggesting a self reinforcing loop in model evolution.  

On the Bitcoin side, the point about language models aggregating a distinct model of the world is especially relevant. Studies have found that mainstream models like GPT frequently hedge or downplay positive aspects of Bitcoin mining, often adding caveats because the training data is saturated with fiat centric narratives. In contrast, models trained on Bitcoin specific datasets tend to emphasize optimism around energy innovation and decentralization.  

