Randomly Generated Ai Quotes

Just as AI lacks the causal frames to win at Dota 2 and needs them encoded by people, so too computers can’t generate counterfactuals on their own but require people to supply them. Carcraft’s rare scenarios were not the result of a machine dreaming alternative worlds, or randomly generating extreme events. Rather, humans came up with them.

Part of the issue is the characterisation of generative AI as a human replacement. This makes people treat the tool as a hyperintelligent magical being that deserves reverence. Recent research, however, shows that AI tools get the law wrong between 69 and 88 per cent of the time, producing 'legal hallucinations' when asked 'specific, verifiable questions about random federal court cases'. A human lawyer or judge with that kind of error rate would undermine public faith in justice. Automation bias means we are more likely to believe the machine than the person who questions it, but also more likely to cut it some slack when we know it has got things wrong. Automation bias's little sibling, automation complacency, means that we are also less likely to check the output of a machine than that of a human. The problem is not the technology; it is the human perception of it that leads us to put it to utterly unsuitable uses which makes it dangerous.

This characteristic of LLMs is called stochasticity. It refers to the randomness inherent in the way LLMs generate responses. These models don’t produce the exact same answer every time, even when asked the same question.

LLMs are connection machines. They are trained by generating relationships between tokens that may seem unrelated to humans but represent some deeper meaning. Add in the randomness that comes with AI output, and you have a powerful tool for innovation. The AI seeks to generate the next word in a sequence by finding the next likely token, no matter how weird the previous words were. So it should be no surprise that the AI can come up with novel concepts with ease. I asked AI to: Find me business ideas that would incorporate fast food, patent 6,604,835 B2 [which turned out to be for a lava lamp that included bits of crystal], and 14th century England.

Once trained, the LLM is ready for inference. Now given some sequence of, say, 100 words, it predicts the most likely 101st word. (Note that the LLM doesn’t know or care about the meaning of those 100 words: To the LLM, they are just a sequence of text.) The predicted word is appended to the input, forming 101 input words, and the LLM then predicts the 102nd word. And so it goes, until the LLM outputs an end-of-text token, stopping the inference. That’s it! An LLM is an example of generative AI. It has learned an extremely complex, ultra-high-dimensional probability distribution over words, and it is capable of sampling from this distribution, conditioned on the input sequence of words. There are other types of generative AI, but the basic idea behind them is the same: They learn the probability distribution over data and then sample from the distribution, either randomly or conditioned on some input, and produce an output that looks like the training data.

But in many ways, hallucinations are a deep part of how LLMs work. They don’t store text directly; rather, they store patterns about which tokens are more likely to follow others. That means the AI doesn’t actually “know” anything. It makes up its answers on the fly. Plus, if it sticks too closely to the patterns in its training data, the model is said to be overfitted to that training data. Overfitted LLMs may fail to generalize to new or unseen inputs and generate irrelevant or inconsistent text—in short, their results are always similar and uninspired. To avoid this, most AIs add extra randomness in their answers, which correspondingly raises the likelihood of hallucination.