3 ms·
Citation needed. A lot of neural-net based AIs actually get better when trained on their own output[1]. [1] https://en.wikipedia.org/wiki/AlphaZero https://en.
by jmcphers 4y ago
Citation needed. A lot of neural-net based AIs actually get better when trained on their own output[1].
[1] https://en.wikipedia.org/wiki/AlphaZero https://en.wikipedia.org/wiki/AlphaZero
- ladberg 4y agoAlphaZero is entirely different than training an LLM. AlphaZero can play against itself as a simulated opponent to make itself better, ChatGPT using its own output as training data will just cause future iterations to optimize for a lower (and less human) bar than the original.
- moyix 4y agoHere’s a more directly comparable example: https://arxiv.org/abs/2210.11610 https://arxiv.org/abs/2210.11610
- m00x 4y agoBut then you can use a GAN to create a bunch of chatGPT data and try to detect it vs known human data, which will optimize chatGPT to generate even more human-like content. This isn't a huge problem, it's barely a problem.
- im3w1l 4y agoThat's a completely different problem, trained in a different manner and optimization for a different thing. One difference is that alphazero tries to beat itself, whereas LLM's will try to mimic itself. Because alphazero is playing a game there are rules that are constantly enforced during the training, which ensures that it stays grounded. LLM's have no rule other than "look similar".
- bordercases 4y agoI'm going to be a bit less nice than your other responses and ask you to give something a bit more thought before providing a glib "citation needed".
- berkle4455 4y agoYou need a citation to understand that an AI which ingests human thought will plateau as new human thought production slows down due to the prevalence of AI-generated drivel? It’s a standard feedback loop.