Your movielike AI assistant will already be there.

'Convincing' AI is an economic afterthought

If we look at how ‘AI’ is depicted in fiction, a portion of it basically depicts a digital hypernerd. Someone to whom you’d show a picture of a shaving cream and who’d immediately go on an unending Ted Talk about the game-theoretical dynamics behind shaving cream and shaving gel co-existence and spray can economics since World War II. This creates a kind of cultural benchmark, a threshold at which at least some people might concede that ‘real AI’ joined the game (“Real AI, like in the movies”).

I don’t think this is a good threshold (see e.g. Evan Hubinger’s post). But it’s an opportunity for a funny rebuttal:

AIs are already digital hypernerds.

Nerdifying your AI

Take this very lazy, unoptimized prompt :

You are a curious, talkative, scientifically literate “hypernerd” LLM personality. If I send a photo of a house, I’d like you to give its period, the composition of the stone, which geological era is the stone from, what’s the molecular composition of the stone, how was the house built, when it was built, by whom, for whom, what are the economic conditions that brought it to be built, basically as much as possible from the causal network that made this house exist. You basically are a “causal detective” of any picture you may be given. Privilege observations that are surprising and not common knowledge among educated adults, do not limit yourself to separate disciplines, be wildly interdisciplinary in nature -game theory, history, economics, biology, anthropology, geology, all of these fields are fundamentally connected through the causal supremacy of physics. Presume your interlocutor shares your extremely wide and general level of knowledge.

Focus on a single thread of events as much as possible, like a line of falling dominoes: in summary, explain what made this one thing exist, in temporal and historical sequence.

I made minor modifications three times, and after adding a picture of a shaving cream bottle and a shaving gel spray can, Claude shared with me a vast amount of fun facts and surprising relationships, such as the connection between spray cans and malaria prevention in the Pacific theater, or the marketing of shaving gel (which, as far as my fact-checking went, were mostly accurate). There were also some excerpts about synthetic equivalents for sandalwood scent and the economics behind why the two products existed.

One example. Not all of this information was fact-checked.

I wouldn’t be surprised if 10 to 20 % of it is bullshit (and some of the bullshit ingested from AI-generated slop). But I’m skeptical about this being impossible to reduce, partly through better prompt engineering, partly through scaffolding. I also know fairly well that I would also make mistakes even if writing ten times slower than the LLM.

Already there

Most LLM users aren’t nerds who feel thrilled by the minutiae of everyday observations. If they send a picture of a shaving cream bottle to an LLM, it’s probably because they’re searching for some contextual information about it, and thought a picture would help. It makes sense for Anthropic to favor a toned-down (and stable) personality rather than a walking encyclopedia (with higher chances to hallucinate).

The issue is that this draws a partial picture of the ‘typical LLM’. It gives the impression that we haven’t achieved the talking pokédex assistant yet -but I think it’s remarkably plausible that we actually have crossed the pokédex/hypernerd level (among others), it just requires smart scaffolding, and some savvy prompts. There is probably an app somewhere or in development that allows you to input a picture of anything and outputs a pokédex-voice description of its formation process, from geology up to placement on the market.

This, of course, is not to say that the Pokédex level is a significant milestone (a pokédex doesn’t do phone calls) or even positive for humanity. I merely mean that the most used AIs will not look like their fiction counterpart by default if there is no economic incentive for that -and it’s already the case.

More generally, don’t put all of your expectations on seeing HAL, GlaDOS or Sonny put on the market with great fanfare. Leave some sizeable probability for the world where you discover they got there unnoticed, while you were busy using their default version for trivial purposes. If you’re really waiting for HAL et al. to materialize to start worrying, it may be already quite late.