Book 02
How a reply is chosen
At each step the model holds a list of likely next words. How it picks from that list is why two replies to the same question can differ.
After this: Say why two runs differ. · 7 chapters · about 40 minutes
- 01The list of next wordsA percent for every possible token5 min
- 02The usual pickAlways take the highest percent5 min
- 03SamplingChoose among the high percents5 min
- 04How wild the pick isTemperature5 min
- 05Only the plausible wordsTop-k and top-p5 min
- 06Why two runs differSame prompt, different landings5 min
- 07The first word can refuseTone sits on the same list as facts5 min