Comment on Pool of data
nutbutter@discuss.tchncs.de 12 hours ago
Technically, you can control the randomness by using the same seed value. Although, it still remains a text completion or a next word (token) prediction model.
Comment on Pool of data
nutbutter@discuss.tchncs.de 12 hours ago
Technically, you can control the randomness by using the same seed value. Although, it still remains a text completion or a next word (token) prediction model.
urushitan@kakera.kintsugi.moe 11 hours ago
At the lowest level yes. But you’re talking about millions of bits of entropy seeds being constantly gathered and spread across parameters with every new token. The industry approach to balancing stochasticity with determinism is Temperature: www.ibm.com/think/topics/llm-temperature