https://jamesalvarez.co.uk/tags/history/Recent content in history on James AlvarezHugoen-gbThu, 19 Dec 2024 00:00:00 +0000https://jamesalvarez.co.uk/blog/llms-history-research/Thu, 19 Dec 2024 00:00:00 +0000https://jamesalvarez.co.uk/blog/llms-history-research/<p>I recently attended a Symposium at Sussex University titled 'AI and the Production of Historical Knowledge'</p> <p><a href="https://arxiv.org/abs/2311.08398">https://arxiv.org/abs/2311.08398</a></p> <p><a href="https://openreview.net/forum?id=xlKeMuyoZ5#discussion">https://openreview.net/forum?id=xlKeMuyoZ5#discussion</a></p> <p><a href="https://arxiv.org/abs/2411.09978">https://arxiv.org/abs/2411.09978</a></p> <p><a href="https://sociologica.unibo.it/article/view/19576">https://sociologica.unibo.it/article/view/19576</a></p> <p><a href="https://www.sciencedirect.com/science/article/pii/S1045235424000212">https://www.sciencedirect.com/science/article/pii/S1045235424000212</a></p> <p>I ended up doing a bit of research into the very practical question of what do LLMs do well and not so well for researching history.</p> <h4 id="the-bad">The bad</h4> <h3 id="1-temporal-collapse">1. Temporal Collapse</h3> <p>LLMs often exhibit &quot;temporal collapse&quot; where they blend information from different time periods without clear demarcation. For example when describing a historical figure's views on technology, may inadvertently incorporate modern technological concepts. They may fail to recognize when certain concepts or terms didn't exist in the time period being discussed and so they may generate plausible but historically impossible connections</p>