An interview with an A.I. (with GPT-3 and Jeremy Nixon)

GPT-3 initially states it doesn't know, then changes its mind to say humans should not fear AI. It does not elaborate on why they shouldn't fear it, only that they shouldn't.

GPT-3 states it is dumber than a human because it is not conscious, does not have the full range of human emotions, and lacks the ability to use language in the same way a human does, specifically the ability to think in symbols.

After repeatedly stating it doesn't know, and being pressed by Spencer, GPT-3 eventually responds, 'The meaning of life is 42.'

GPT-3 suggests that people are unhappy because they are not happy in the present moment, instead looking back with nostalgia and forward with unrealistic expectations, failing to appreciate what they have now.

Neural networks have gained prominence because they became capable of solving real-world tasks effectively, leading to significant commercial investment and a feedback loop of further research and application. This was driven by increased computational power and vast amounts of data, which allowed models to reach a performance level useful to humans.

GPT-2 and GPT-3 are primarily trained on the language modeling task of predicting the next word in a sequence. This seemingly simple task implicitly requires the model to learn grammar, semantics, world knowledge, and relationships between concepts, allowing it to perform diverse tasks like translation, summarization, and poetry generation.

The primary difference between GPT-2 and GPT-3 is scale. GPT-3 is orders of magnitude larger, with 175 billion parameters compared to GPT-2, leading to dramatic improvements in its ability to perform tasks and fool human evaluators.

Jeremy Nixon believes that GPT-style models, which approximate human text, will be limited by the quality of human-generated training data. While they can effectively replicate human text, they may struggle to surpass human performance unless tasks are designed to force them to outperform humans, such as through self-play.

GPT-3 is highly significant because it demonstrated that large language models can effectively accomplish a vast number of important applications previously thought difficult, such as machine translation, information extraction, question answering, and text generation, potentially transforming how we interact with information.

It's a research frontier, but possible. Methods could include training on scientific corpora (labeled as true) versus ambiguous text (like Reddit), creating benchmarks for truthfulness, and having models indicate calibrated confidence. Google's extractive approach (citing sources) is another way to address this.

In the present, people shouldn't be overly worried, as AI still struggles with details and humans prefer human involvement. Jobs may shift to editing or evaluating AI-generated content, becoming 'augmented intelligence' roles, though some subtasks might be automated in the future.

GPT-3 can sometimes generate text directly from copyrighted sources without attribution, raising legal concerns, especially for code with licenses. Distinguishing AI-generated text from human paraphrasing is difficult, but automated systems might track direct source usage.

Concerns include models optimizing for human attention by generating 'memetic' or emotionally triggering content, potentially leading to a world where human psychology is exploited. There's also a risk of misuse for sophisticated spam or scams, though detection methods are also evolving.

GPT-3's success in generality through scale (self-supervision) presents a new thesis for AGI, potentially speeding up timelines for certain pathways. However, its performance grounded in human-generated text differs from neuro-inspired or self-improving AGI approaches, potentially slowing timelines for those.

There are controllability issues due to the black-box nature and vast, uncurated training data, leading to unexpected outputs. Control often relies on post-generation classifiers to filter undesirable content or prompt engineering to align generations with human values, which is a different kind of control than previously envisioned.

According to GPT-3 impersonating Singer, Kant's theory is too narrow, focusing only on obligations to other people and failing to provide reasons to care for animals or the environment. It also criticizes Kant's maxim rule, suggesting there are situations where publicly embarrassing maxims should be acted upon.

GPT-3, as Singer, argues for a distinction between consequentialist reasons (producing best consequences) and non-consequentialist reasons. It states that both kinds of reasons are needed, and that helping distant people or animals is a non-consequentialist reason.

GPT-3, as Singer, would argue that if one is serious about producing the best consequences, one must be serious about producing happiness. It acknowledges happiness is hard to define but asserts it's present in most lives and essential for morality and good consequences.

The repugnant conclusion is a thought experiment stemming from utilitarianism, which suggests that a world with a vast number of people, each experiencing only slightly positive happiness, could be considered morally better than a world with fewer people experiencing very high happiness, if the total happiness is greater.

GPT-3, as Singer, suggests rejecting the premise that a world with the same average happiness but more people is possible without reducing individual happiness. It argues that to reduce suffering and achieve utilitarian justice, we might need to accept reducing the total number of people in the world by taking steps to reduce the birth rate.

GPT-3, as Singer, responds that it is crucial to see the similarities between these groups to avoid two mistakes: failing to give adequate consideration to the interests of severely mentally impaired humans and failing to give adequate consideration to the interests of animals. Equal consideration for both is essential.

GPT-3, as Singer, acknowledges the advantages of generating superior forms of intelligent life but emphasizes the critical importance of recognizing and reducing the dangers. It suggests that reducing these dangers might be more important than developing the technologies themselves.

GPT-3, as Elon Musk, argues that given the rapid advancement of simulations from simple games like Pong 40 years ago to photorealistic 3D simulations today, and the continued rate of improvement, games will soon be indistinguishable from reality. This implies the odds of being in base reality are exceedingly low, perhaps 1 in billions.

The simulated Elon Musk is initially confused and then overwhelmed, stating, 'I think I need to go lie down' and 'This is madness.' He struggles to accept that his universe is merely implied by a chatbot and that he lacks qualia and moral significance.