https://fuckai001.com/technology-inspired-thinking/Hugozh-CNTue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/01-problems-and-solutions/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/01-problems-and-solutions/最初人们研究语音识别，是把它当作语音学、语言学问题，总不得要领，识别错误率很高。后来从信息论、概率论、统计学方向入手，把识别转换成概率计算，很快就把错误率从30%降低到了10%以下。 比如识别一句语音：今天下雨了。 早期思路：先研究发音的声学特征，然后制定规则：这种频率、这种共振峰应该是i，那种应该是t……一步步拼成字。由于现实中的口音、语速、噪音等问题，这种规则识别很容易出错。 概率思路：先让计算机把每个词识别出来，识别不准没关系，今天还是金天？过程中不断计算词的概率，今天的概率>金天，于是就用今天；最后是今天下雨了，还是今天大雨了？按概率很可能是前者，就用前者。https://fuckai001.com/technology-inspired-thinking/02-information-to-computation/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/02-information-to-computation/https://fuckai001.com/technology-inspired-thinking/03-ai-and-skill-acquisition/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/03-ai-and-skill-acquisition/https://fuckai001.com/technology-inspired-thinking/04-neural-networks-and-individuality/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/04-neural-networks-and-individuality/https://fuckai001.com/technology-inspired-thinking/05-other-machine-learning-insights/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/05-other-machine-learning-insights/https://fuckai001.com/technology-inspired-thinking/06-does-education-still-matter/Tue, 20 Oct 2026 00:00:00 +0000https://fuckai001.com/technology-inspired-thinking/06-does-education-still-matter/