Media
5 items
Description
Highlight the informative words. Read faster. — Facing a wall of text, almost everyone finds reading difficult to some degree, especially those with ADHD. The way humans read has not fundamentally changed in a thousand years: the eye still goes word by word. Speed-reading methods exist, but they are limited by the physiology of the brain and the eyes. In the age of AI, there is a new way to read faster: let a large language model highlight where your eyes should linger, and skip the rest. Surprisal Theory in psycholinguistics has been repeatedly confirmed: the more unexpected a word is in the context, the longer the brain takes to process it. Surprisal and reading time are nearly proportional, across languages. Recent cortical experiments also show that humans and large language models share the same computational principles for language: both continuously predict the next word. The Info Highlight extension uses a large language model to compute the surprisal of each word and paints it as a heatmap on the webpage. The harder a word is to predict from the preceding context, the darker it is, and the more your eyes should linger. Lighter stretches are already guessable from what came before, so you can skim. Both theory and intuition support that this can speed up reading. For example, in “I opened the door and saw a UFO”, “UFO” deserves more of a pause than “cat” in “I opened the door and saw a cat.” After the author’s own small-scale use, Info Highlight really does help you read faster — it’s hard to do without it now. Features: • Works with most kinds of text, including PDF • When your computer supports WebGPU, analysis can run on-device and the page text is not uploaded • Optional cloud analysis; the cloud path is handled with full privacy, and the code is open source ———————— 高亮信息量大的字，加快阅读。 — 面对大片文字的时候，每个人都或多或少感到阅读困难，甚至阅读障碍，特别是ADHD人群。人类现在的阅读方式和一千年前没有本质区别：肉眼逐字阅读，有快速阅读的方法，但是受限于大脑和眼睛的生理上限。在AI时代，有一种新的加速阅读的方式：让大语言模型来帮你高亮眼睛应该停留的地方，快速跳过其它地方。 心理语言学里的惊讶度理论（Surprisal Theory）已经反复验证过：一个词相对上文越意外，大脑处理它就越慢；惊讶度和阅读时间几乎成正比，而且不限语言。近年的皮层实验还表明，人类理解语言和大语言模型一样，都在连续预测下一个词——双方共享同一套计算原则。 Info Highlight插件通过大语言模型，计算每个字的惊讶度，在网页上通过热力图的方式绘制出来。越难从上文预测出来的字，颜色越深，眼睛越该停留；对于浅的地方，既然通过上文已经能猜到，可以快速扫过。它在理论和直觉方面都支持可以加快人类的文本阅读速度。比如“打开门看见了一只飞碟”里的“飞碟”，就比“打开门看见了一只猫”里的”猫“更该停留。 经过作者小范围的试用体验，Info Highlight的确可以帮助读的更快，现在已经离不开它了。 特点： * 对大部分文本类型都适用，支持 PDF * 本机支持 WebGPU 时，可在本地分析，正文不上传 * 可选云端分析，云端链路全隐私处理，代码开源公开