4 ms·
>> How does one train an AI to make up such convincing lies? Oh, that's simple. It's a free bonus of training on human-produced texts, which are often imprecis
by pk-protect-ai 3y ago
>> How does one train an AI to make up such convincing lies?
Oh, that's simple. It's a free bonus of training on human-produced texts, which are often imprecisely defined. The extrapolations also produce various data with assigned probabilities, which may or may not be true in the future. Therefore, it's not surprising that AI generates lies since it generates and merges tokens in a probabilistic manner.
And here is what GPT-4 (phind.com) tells about it:
Training an AI to generate convincing lies is actually a byproduct of training on human-produced texts, which are often imprecisely defined¹. As the AI learns from these texts, it extrapolates and generalizes information, creating a variety of data that may or may not be true.
This process involves generating and merging tokens in a probabilistic manner, which can result in AI-generated lies. The AI doesn't intentionally create lies, but the nature of its learning process leads to the possibility of generating false information¹.
[1] https://towardsdatascience.com/real-artificial-intelligence-understanding-extrapolation-vs-generalization-b8e8dcf5fd4b https://towardsdatascience.com/real-artificial-intelligence-...
- visarga 3y agoI see LLMs as "language simulators". They just execute language in - language out, but they have no space to memorise all the facts in the world. Each token "touches" the whole network, so you could say it visits the whole culture before being created.
- fastball 3y agoCalling human-produced texts "imprecisely defined" is being generous. The truth is that many (most?) human texts are riddled with inaccuracies or straight up lies. I have to imagine that a GPT trained on a similarly sized corpus of text as GPT4 but only 100% factually accurate (to the best of our knowledge) would be pretty good at sticking to the facts.