The Beholder's Share in Idea Recombination

I was recently in Poland (notes and photos here for the curious), and in Warsaw we visited the Marie Curie museum. I studied biochemistry way back in the day which made this more fun, and maybe I wasn't paying attention in class, but I suppose I never thought deeply about element 84: polonium.

Marie Curie was Polish and, at the time, was working in Paris. In 1898, she and her husband Pierre were studying pitchblende, a uranium-rich ore, when they made a strange observation: it appeared to be more radioactive than they'd expect given the amount of uranium inside.

They speculated that the ore contained another unknown element but hadn't yet isolated it. Even without a confirmation from a characteristic spectral line, they went ahead and published a conditional claim: if the new metal was confirmed, they wanted to name it polonium, after Marie's motherland.^1,2 She later wrote that she chose the name "in memory of her native country."³

There are at least two stories here: the first is about chemistry and the discovery of a new element on the periodic table, the second is about the name polonium, and how it's situated within the history of Polish identity and partition (at the time, Warsaw was under Russian rule).

The transmission of ideas through language, like any art, is never finished by the writer. Often, the audience completes the meaning of the piece using some combination of life experiences and imagination. This interpretive act is what art historians refer to as "the beholder's share."⁴

The environment around the beholder is both physical and digital. In The Architecture of Intelligence, de Kerckhove describes an "architecture of connectivity" that joins mental, physical, and networked space.⁵ What someone can understand depends not only on the record in front of them, but on the people, places, interfaces, ideas and even feelings their information environment makes available.

Which aspects of polonium's discovery matter depend on which questions we bring to the situation. By changing the question, different connections become more or less useful. Scientific findings work similarly: their future meaning and usefulness are partly completed by readers who bring questions and contexts the original authors could not anticipate.

I previously suggested that ideas flowing through innovation systems might be modeled at a high level like this:

In order to be buildable, findings have to be generated, found, validated, and understood. Finding ideas is insufficient: our ability to recombine it with something else is, in part, bottlenecked by whether or not we can understand why and how it matters to us.

Where writers, editors, and institutions bind a few contextual paths early in the creation process, readers have to reconstruct alternatives, sometimes at considerable cost. In a world of abundance, we may need agents to help us construct and challenge more of those paths later, raising the odds that a finding becomes understandable for a new problem, and sometimes, buildable in a new combination. This quickly creates an interesting set of design challenges around contextualization.

First, does our agent understand what we care about, and is it able to find interpretations of ideas that are relevant to us, and demonstrate why?

Second, if what we read is what an agent presents, not the source text, how does that change how we interpret ideas presented to us, and our capacity to imagine new possibilities?

Creating Frames: Contextualization on Write and on Read

Writers already do quite a lot of contextualization for us readers. Scientists choose which prior work to mention and explain why an experimental finding matters. Editors apply similar judgements for what they imagine to be the core readers of their journals. Museums and their curators elevate a room of objects into an exhibition by creating context for attendees.

This "context on write" is useful of course, museum visitors need some idea of what they're looking at. But fixed artifacts can only bind so many paths through their material.

Frames are fairly simple to understand in practice: who is asking, what are they trying to figure out, and what are they treating as relevant? A chemist reading the Curie paper has one view; someone trying to explore Polish independence or identity under partition might have another.

Marketing has a wonderfully practical example. If you worked at a milkshake company and wanted to grow revenue, you might wonder how to make a better milkshake and then ask how it compares to other direct competitors. Asking a commuter what a milkshake accomplishes for them, though, might expand the consideration set of alternatives (bananas, bagels, donuts), cluing you into the idea that duration and one-handed use matter more than milkshake flavors.⁶ In the process of reframing, you find other solutions by refining your understanding of the problem.

A frame is useful precisely because it leaves certain things out, creating focus in particular directions. Omission becomes a problem when what was left out would materially change the answer, stakes, or available action, which, as we'll see later, is why imagination is crucial to frame generation. Ideally, surfacing another frame should therefore not just show other interpretations, but what the first one considered and what it excluded.

This is all ultimately at the core of how we recombine ideas.

Imagination and Recombination

If contextualization modulates what is relevant, imagination asks what might follow if a new relationship holds.

Peirce's account of scientific inquiry demonstrates one version of this. A scientist might find a surprising observation, which prompts a process of abduction to posit a hypothesis: a possible explanation worth testing. Deduction then works out what else should be true if the explanation is right, while induction tests those consequences against experience.⁷ Imagination, here, is a process of explanation-generation to be tested later.

Let's go back to the Curie and polonium example. Pitchblende was unexpectedly radioactive, prompting the Curies to ask whether the presence of another element would explain the anomaly. That "what if?" opened up a sequence of experiments with which they tested the possibility.

This pattern happens in innovation systems generally, not just in academic scientific discovery. I've written previously about Martine Rothblatt and the formation of United Therapeutics.

When her daughter was diagnosed with pulmonary arterial hypertension, Martine Rothblatt started searching the literature and chasing through citations. She eventually found a paper about a molecule that was originally tested for congestive heart failure. Under the paper's frame, the molecule had failed, but one of its effects was to reduce pressure in the pulmonary artery without similarly reducing pressure elsewhere in the body.⁸ Exactly what she was looking for.

She then found a way to license the drug, spun up the company, and United Therapeutics was formed. Martine Rothblatt, as a reader, supplied a frame to a paper written under a different framing, and was able to identify an opportunity that otherwise went unnoticed. I recommend listening to the full interview, it's rather wonderful.

This is the imaginative part of contextualization: a finding is produced inside one frame and someone approaches it with another and asks what else it could mean or do. Naturally, any system that increases the rate of frame generation also needs a tight loop around verification to test these candidates.

Letting Agents Ask: What If?

This feels to me like a more concrete way to conceptualize the value of agents in innovation systems. Instead of summarizing papers, agents might be armed with a problem and approach existing findings through that frame. Which negative results demonstrate an effect that's useful under a different clinical objective?

Rather than simply generating a long list of questions, I think the real value here is in constructing a defensible path between different frames.

For example, in the Rothblatt example, an agent would need to understand the original congestive heart failure trial, figure out how to represent what pulmonary arterial hypertension requires (decreasing pressure), identify the pressure result in the published paper as the bridge, and explain why the relationship might be worth testing. A classic search system might miss the paper because the relationship Martine cared about wasn't the relationship under which the original paper was written and indexed.

This gives us something more useful to evaluate than how imaginative an agent sounds, creating a forcing function with which we can ground and benchmark its abilities. By holding the source material steady and supplying different problems, we could, for example, see whether it preserves original findings while proposing relationships that experts judge coherent and testable. A genuinely different frame should change which properties, sources, comparisons, or experiments matter, rather than merely changing the language of an answer.

These systems also pose risk through their reflexivity. As suggestions shape what I notice and what I ask next, my ability to form interesting directions might either be delegated to the system, or worse, be subtly narrowed in ways I don't fully realize (if we apply de Kerckhove's thesis, the risk is really in the changes to the shape of my own information environment). Ideally these systems should make the frame itself visible: what it considered, what it excluded, and which alternative frames might lead us somewhere else.

This is one key aspect of the famous "create context" warning that Hideo Kojima had at the end of the video game Metal Gear Solid 2 (I highly recommend watching this).

As I've argued previously, our ability to scale this depends on building a tight loop around ideation and verification. Agents could expand the number of frames brought to an existing idea and ask far more "what if?" questions than any one person could. The bottleneck then moves from producing possible relationships to deciding which ones deserve our attention and belief.

Cheers,
Ashish

P.S. We've now covered most of the raw system components, so we'll start getting a bit more technical in future essays!