The world of “AGI,” as it is broadly advertised, depends economically on inference becoming extremely inexpensive. (That is part of the backwards Jevons story Frontier AI tells, which I discussed in my last post.)

AI inference is the process by which the AI applies its model to new input and produces an output. Otherwise, in the dictionary, an inference can be the process of using logical reasoning to reach a conclusion. That said, we often use inference to say that we made a judgment or formed an opinion. And that more common usage will prove to be more relevant than it may appear now.

Frontier AI’s value proposition relies on our accepting the AI operation of inference as reasoning itself. For frontier AI’s success, inference cannot be simply a proxy for judgement; computational inference must become a judgment machine. The slippage occurs when an inferential output is treated as a judgment, especially when that judgment organizes or directly initiates action. However, if AI inference is going to be cheap and ubiquitous then it will need to be accepted as a general purpose judgment machine. But I have gotten a little ahead of myself. To be fair, when it comes to frontier AI, it can become confusing remembering which tail is wagging whose dog.

The Jevons-esque theory for frontier AI is that when inference becomes too cheap to meter then we will use it everywhere. However Jevons’ theory doesn’t insist that a commodity becoming cheap will result in it becoming ubiquitous. It only suggests that we would continue to demand more of it in the contexts where the commodity would have value. The problem for frontier AI labs isn’t only cost. For AI to become ubiquitous, individuals and institutions will need to invent many more sites for machine judgement. Not every commodity works the way electricity does. For example, if cocaine became insanely cheap I still don’t think we’d put it in everyone’s deodorant or toothpaste or even back into Coke. That said, back in the 70s and 80s, corporations did find a lot of uses for cocaine. But maybe that proves my point. Cheapness isn’t enough. There also needs to be an institutional culture that encourages its use.

Realistically, what would it mean for AI inference to become so inexpensive that we would not only use much more of it in the places where AI already exists but that it would begin to permeate life on a grander scale? There are a lot of speculative answers to that. The criti-hype response is that inference is not presently inexpensive in any straightforward economic sense. Its apparent cheapness to users depends on enormous capital expenditures, investor subsidies, and unresolved business models.

Instead, we might consider in what sense inference or judgment can become too cheap to meter. That is, we can ask what might be at stake in making judgment cheap. The easiest example is with an autonomous vehicle. The inference process operating within in the vehicle is basically free to employ. Then the vehicle causes a traffic fatality. Is the inference still too cheap to meter? Clearly the problem with the concept of free inference is that while inferences can be freely made, acting on inferences has consequences. A traffic fatality is an obvious example, but every choice produces counterfactual costs that are unaccounted for. In fact, it can become difficult to account for the hidden costs of continuous judgments that do not create ruptures like car accidents. The fact that AI might not “exert” itself to output an inference does not mean that those inferences do not have material consequences.

Then there is the internal contradiction. Each inference may conclude, but an economy organized around cheap inference cannot. It must continually create new occasions for judgment. Technically each output concludes its inference, but when the output generates further inference and those generations happen at greater speeds, then the dynamics have shifted I would say. Instead, incessant inferences generate tasks, actions, outcomes, and consequences that become new inputs for further inference. This sociotechnical exchange among AI, machines, and humans continuously updates, infers, outputs, acts, and samples the consequences of its previous actions.

Finally, without dipping into criti-hype, we can consider the existing network of AI inputs… I’m afraid that would be a very long list. These AIs don’t work on their own. Continuous machine judgment requires continuous machine perception. Consider video surveillance: AI systems depend on the technical and institutional capacities of cameras and their networks. Here we might consider the current discourse around Flock and Axon (artist formerly known as TASER International) cameras. Understandably, inferences made from video require the video to be communicated first. I am “picturing” a series of odd-angle images, temporally stamped across a landscape that becomes stitched into an AI generated narrative and a judgment. So we might start considering the time-critical media and infrastructural operation of inference as a judgment machine built upon a weird digital temporality that is part sample and part inference.

The frontier AI narrative suggested, at minimum, that AI would do better than we do–for example making quick and cheap  judgments. But AI’s economic narrative depends on inference becoming ubiquitous. And continuous machine judgement will require infrastructure that supports both continuous inference and continuous input. 

What kind of intelligence can only realize itself through making the world part of its machine?