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.
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: https://www.ibm.com/think/topics/llm-temperature
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.
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: https://www.ibm.com/think/topics/llm-temperature