Bruno’s Newsletter #9

Theorists often speculate on what the best image or metaphor for a powerful future AI might be: an agent, an oracle, perhaps a genie or tool following human commands. But these metaphors all miss the point: AI is a virtual world. As David Chalmers notes, the agentic model is misleading and GPT “is more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.” The fact that a large language model will often produce wrong results or “hallucinations” deviating from established facts should not surprise us once we understand that it creates variations of the internet, where many inaccurate facts are always present. GPT does not write nonfiction. It writes fiction, even when answering factual questions. GPT “was not optimised to be correct but rather realistic, and being realistic means predicting humans faithfully even when they are likely to be wrong.”

Because each individual user experiences only a few instances of the model’s creative powers, it is easy to mistake it for a nonhuman interlocutor or agent with which human agents engage in communication. But there is a reason communication plays such a central role: it is used for training purposes! And there is probably no other way for the model to manifest itself to a human user. One might recall the history of religion with its numerous examples of how a divine intelligence must manifest itself in human form in order to then reveal its real nature. It is easy to assume that the way we use GPT, or the way we commonly use GPT, tells us what GPT is. But this is an error of perspective, a case of the fallacy of misplaced concreteness. In order to overcome it, an effort of abstraction is needed, an effort akin to that of modern science when it abstracted from physical things or substances in order to reach the underlying laws of the universe.

The interface is not a search interface but a generative one: what you ask the model is to recreate or rephrase the internet. After all, the model does not contain any specified content but rather a set of rules for generating content. In this sense, it is a virtual engine. Using the organising principles or statistical regularities extracted from the current iteration of the internet, it is possible to recreate it or to create new iterations, provided they follow the same fundamental rules. The method is distribution learning: how to extract from training samples the rule or algorithm governing the distribution from which they have been drawn. But then, as Ted Chiang asks in his New Yorker essay: why, if we already have one internet, do we need to recreate it?

The answer is that this recreated internet runs virtually on machines and a virtual internet extends to infinite iterations of the internet we now have. A meeting between Socrates and Confucius is one case in the latent space of possible iterations. What is recreated is not the current internet but a virtual one. This also means that nothing is left out. Instead of having to fill in gaps or resolve contradictions between different search results, a user in the virtual internet can simply ask for the result it wants. Everything will be instantaneously rolled out. The model provides both the rules and the limits of possibility. It takes the place of the “laws of nature” in the physical world. In its growth process, the internet aspires to replace the physical world. This is where AI comes in.