The Abstraction Dial

A few weeks ago, I asked Claude to pull together a literature review for a paper I am working on. Maybe half an hour later I had roughly a hundred pages of well-structured article summaries as well as a synthesis. I skimmed the packet and moved on. Then it hit me that I had just done the exact thing I keep telling my students not to do. Basically, I had accepted work I had not actually read, because there was too much of it to read. While I didn’t send it off to a journal or a conference, I moved on to other tasks. This means that in principle there had been a “human in the loop,” yet it wasn’t practically possible for me to validate the entire output.

In my post on human-embedded agentic design (HEAD) I argued for flipping this around, to put “AI in the loop” of a human designer rather than dropping a human into the loop of the machine. Last week, I wrote about how to engineer the context architecture within a HEAD process. This week I’ll be discussing what I call eval artifacts, temporary scaffolds that help you examine the ongoing process and outputs of agents working under your supervision. I think such eval artifacts are critical for human designers to be able to engage critically and creatively with an agentic design system.

Information overload

It is tempting to read my literature review story as a personal failure of discipline: I got lazy, I should have read all 100+ pages, shame on me. Boohoo. I don’t, however, think that is the right lesson, and the evidence for the whole “just add a human” idea is less than flattering. A recent meta-analysis found that on average, humans and AI working together performed worse than either working on their own. When tasks were focused on making decisions, working with AI hurt human performance. The silver lining is that working with AI seemed to boost humans’ performance on creative tasks.

So, what is the lesson? I think it is that putting a person next to a machine does not reliably produce quality outputs, and sometimes it produces worse outcomes than if the machine (or the human) was left to its own devices. A big part of the explanation here is the cognitive effort it takes to work with AI. When an agent hands you hundreds of pages of prose, or perhaps thousands of lines of code, the price of validating the work skyrockets, and going along with it (or completely ignoring it) becomes rational, albeit ultimately ineffective. Hence, I think that human-in-the-loop often is “governance cosplay” rather than a serious effort at governing AI in a way that creates better decisions in the long-run. I’ve spoken in the media about this previously.

The abstraction dial

Therefore, we need tools to manage the information load that AI is inflicting upon us, and this is where the idea of an “abstraction dial” comes in. At 0 sits the raw artifact, the 100+ pages or the thousands of lines of code. Turn the dial and the same artifact becomes a five-page synopsis, then a single slide, then one figure, then a sentence you can hold in your head. What you are moving is the level of abstraction, while the deep structure of the underlying content remains the same.

LLMs operate on the deep structure of information and can re-cast that structure into almost any surface form. Indeed, some researchers are building systems that help to do what is referred to as abstraction matching, which is about modulating the level of abstraction of prose or code so that it communicates well with an LLM (and humans, of course). It’s possible to translate model output, such as code, back into plain English so the user can see the level the model is operating at. This is a form of “re-representation” which allows different users to work with code (or prose for that matter) at a level of abstraction that works for them. I call the outputs of this practice eval artifacts, because their job is to let you evaluate what the machine produced. These artifacts function as scaffolds for temporary abstractions that help the designer throughout the process of working with agents.

Fluid artifacts

Eval artifacts are things such as summaries, figures, tables, or visualizations. Most of us are used to produce these things as Word or Powerpoint documents, but increasingly I think we can create more interesting, more dynamic artifacts to share information with each other. Anthropic is pushing this through allowing anyone to publish artifacts—essentially little web apps for custom purposes—instantly from Claude. Increasingly I expect we will send each other small web apps rather than documents and spreadsheets, and the person on the other end will re-represent and edit them with an agent rather than reviewing them from top to bottom. Hence, digital artifacts are becoming even more fluid than before.

When deployed throughout a design process, such as HEAD, these eval artifacts can be helpful for dealing with what I increasingly think is one of the more difficult problems related to AI—the reduction in creative outputs that they seem to engender. For example, recent research has shown that when writers used generative AI for ideas, their individual stories were rated as more creative, especially for less creative writers, yet the stories as a set became markedly more similar to one another compared to stories written by humans alone. LLMs pull toward the center of gravity within their training data. As we move towards more abstraction in the eval artifacts that we create to understand what is happening in our work processes, an LLM will throw information away, and replace it with something that moves incrementally towards the “modal account” encoded in its weights. So, working with abstraction levels when creating eval artifacts does not remove the need for judgment, but if we can figure out how to do it right, they can be helpful for stopping the information overload produced by LLMs, thus helping human designers to leverage their creative capacities.

The literature review is still sitting in one of my Claude Cowork folders. I have not approved it yet, what I did instead was build a small dashboard around it, so I can move between the one-figure version, the 100-page version, and the layers in between, and decide for myself which level of abstraction works best for my work process.