Infinite interns versus cantankerous colleagues

Two visions of AI applications

The march of public progress in large language models has slowed some, and we now have a little time to turn our attentions from the newest programs to how we might actually use them.

Of course, many of us have been using ChatGPT since it came out, and have been finding new uses for it along the way. Myself, I probably spend half the time I used to writing code. This is a huge boon, but I also use it to help me edit writing that I don’t feel particularly confident about — cold emails and journal submission cover letters always make me feel like I have no idea what conventions I should follow, and I use LLMs to help me at least produce a sensible version that I feel okay about sending. And of course, I use it as a tutor of infinite patience, who’s always ready to come up with a new metaphor to help me understand some concept or other.

Even before that, though, a lot of people were thinking about how we’d integrate LLMs into our work and lives. Benedict Evans wrote in 2018 about one vision that has taken hold partly for its pragmatism and partly for its universality. His essay “Ways to Think about Machine Learning” introduced the concept of infinite interns:

Talking about ML does tend to be a hunt for metaphors, but I prefer the metaphor that this gives you infinite interns, or, perhaps, infinite ten year olds. […]

That is, machine learning doesn't have to match experts or decades of experience or judgement. We’re not automating experts. Rather, we’re asking ‘listen to all the phone calls and find the angry ones’. ‘Read all the emails and find the anxious ones’. ‘Look at a hundred thousand photos and find the cool (or at least weird) people’.

This is a level-headed view of AI based on a sober view of AI’s present capacities, and it resonates with sober, level-headed people. We’ll just keep on doin’ what we’re doin’, but AI will help us do it more efficiently. Evans’s vision of infinite interns is sensible because it allows AI to simply slot into the status quo. Vitally, it keeps the domain of human judgment almost entirely intact. The human mobilizes AI to fulfill some goal of the human’s choosing, and the machines’ only judgments are on the level of the single phone call or photo. It has no say in the goal itself.

Given most of a year’s experience with LLMs, though, I think it’s clear that AI should cause us to fundamentally rethink some major aspects of how we do things. Sure, AI should take on a lot of tedious tasks, however we reorient ourselves to the presence of our new silicon-based companion species. So there’s no use doing away with the infinite interns concept. Instead, I want to augment it with what I see as the more important and more disruptive use case for AI — what I call the cantankerous colleague.

Serving as a cantankerous colleague, AI would be an ever polite, yet ever skeptical peer who questions what you’re doing. Think of it as a kind of automated red team. With every diagnosis a physician makes, for example, the AI might say, “Hold on, I think it’s something else because [plausible chain of reasoning].” Or imagine that you’re working on an essay: when you pause, the AI might pop up and let you know it thinks that your argument is predicated on some false assumptions. This could be especially valuable for planning interventions or analyses, where every document could become an adversarial collaboration.

The medical application comes most naturally to mind for me. As an example to see how AI as a cantankerous colleague could be beneficial, let’s take the example of radiologists screening mammograms for breast cancer. They detect about 87% of breast cancer cases, while missing the other 13%. Using AI as infinite interns would, in a best case scenario, identify the same 87% of cases that radiologists would. This improves efficiency, but it doesn’t advance what we can do with humans. Instead, if we imagine an AI that causes radiologists to reconsider and correctly identify that 13% of cases that they would otherwise miss, it would allow us to do something we couldn’t accomplish before.

The benefits to epistemological praxis would be tremendous… at least among those who could tolerate working under those conditions. That’s both the challenge and the potential of this approach, though. It could allow people to cultivate a much better calibrated sense of uncertainty that might ultimately permeate into treating nearly every action as an experiment. This is a skill, though, and not everyone would be equally adept at it. Already, there’s some evidence that the ability to distinguish between true and false advice labeled as being from AI doesn’t correlate well with overall performance.

Still, if Evans’s concept of infinite interns took off in part because it offered a realistic take on what we can do with AI’s current capacities, does that mean that my concept of AI as a cantankerous colleague has to wait indefinitely? I don’t think so, although in a brief experiment, ChatGPT didn’t seem to take on that role with much enthusiasm.

The major LLMs were developed using RLHF, reinforcement learning with human feedback. Trainee models undergoing RLHF provide multiple responses to each query, all of which are then ranked by humans on a number of criteria such as helpfulness.

My idea is that RLHF could also provide a way to train cantankerous colleague AIs: rather than rewarding the eager to please and quick to apologize attitude that it currently displays, why couldn’t we design criteria that rewards polite but firm attempts at refuting ideas? The highest performing of these attempts will then allow users to shift the tone by adapting or clarifying their statements, or even by providing evidence to support them.

AI provides us with an opportunity to expand our abilities. Yes, a major part of that is freeing us by taking on drudgery as infinite interns. But if we allow ourselves to ignore the potential to use it to help us do things we couldn’t do before, we leave a lot of its benefit untapped.