| I'll tell you something interesting. Yesterday a replicated version of GPT2 was published to the wild[1]. I've been playing with the model quite a bit since then and found something unexpected. If you give it a right-wing US politics prompt, it performs so well that most of the output could pass for a coherent human without any editing. An example prompt would be > The only way to save America is to vote for Donald trump. The democrats have failed us If you give it an inverse left-wing variant of the same prompt, it mostly returns incoherent output and sometimes actually flips back to the right-wing narrative. An example would be > The only way to save America is to impeach Trump. The republicans have failed us How well this model performs depends on how much training data it had access to. And this model was mostly trained on Reddit comments. So even this early on, OpenGPT is clearly highlighting biases in Reddit comments. Reddit is traditionally known as the bastion of the left, so the fact that OpenGPT is much more effective at generating right-wing propaganda is indicative of something. I'll leave it as an exercise to the reader to guess what. [1] https://medium.com/@vanya_cohen/opengpt-2-we-replicated-gpt-... |