4 ms·
Also continuous learning. The training step is currently separate from the inference step, so new generations have to get trained instead of learning continuous
by SAI_Peregrinus 3y ago
Also continuous learning. The training step is currently separate from the inference step, so new generations have to get trained instead of learning continuously. Of course continuous larningin a chatbot runs into the Microsoft Tay problem where people train it to respond offensively.
- keepamovin 3y agoYeah, evolution multiple generations. Necessary for sure. Things have to die. Otherwise, there’s no risk. without risk, There’s no real motivation to live and without that there’s no emotion no motivation to learn and without that there’s no AGI.
- SAI_Peregrinus 3y agoI'd say you need it to learn within each generation as well. Human & animal learning doesn't require evolution, we learn without replacing the whole. Connections within brains change while the organism is still alive, without even individual neurons dying off. The current process is purely iterative; GPT 1 was "done" and GPT2 got made. Then GPT 3, etc. If ChatGPT4 gets corrected about some particular thing by many users it won't learn that for the next user to come along. Of course learning has an inherent trust issue; if the inputs one is learning from are incorrect then incorrect information will be learned, and incorrect output will result. Microsoft experienced that with Tay, which quickly learned how to respond to people via training from 4chan users. That was predictably disastrous, and what OpenAI are hoping to avoid with the generational learning strategy.