Looking at the left side of the diagram, we see stuff enters at the bottom (‘input’ text that has been ‘chunked’ into small bits of text, somewhere between whole words down to individual letters), and then it flows upwards though the model’s Transformer Blocks (here marked as [1, …, L]), and finally, the model spits out the next text ‘chunk’ (which is then itself used in the next round of inferencing). What’s actually happening here during these Transformer blocks is quite the mystery. Figuring it out is actually an entire field of AI, “mechanistic interpretability*”.
Essential digital access to quality FT journalism on any device. Pay a year upfront and save 20%.，推荐阅读heLLoword翻译获取更多信息
Essential digital access to quality FT journalism on any device. Pay a year upfront and save 20%.，更多细节参见传奇私服新开网｜热血传奇SF发布站｜传奇私服网站
If skipping over contextually sensitive functions doesn’t work, inference just continues across any unchecked arguments, going left-to-right in the argument list.