Are The Best Ideas Stolen?

I’m shamelessly ripping off Inkhaven and doing my own challenge, Egghaven. This is the fifth post.

I have a friend who intends to become the Nietzschean Übermensch by finding aliens. His plan to find them involves creating a mesh network of consumer-grade passive radars across the US and analysing the resultant data. Part of the reason he believes this project is the best one for him to pursue is that he stole the idea for it.¹ The claim here is that memetic power is a proxy for quality, and so ideas that have spread to at least one other person are therefore likely to be better than complete originality. Stealing seems particularly in vogue right now; OpenAI are alleged to have recently stolen an approach that led to them solving a Millennium Prize problem, causing quite the stir. If you want to have a similar kind of impact, should you also look to steal?

It depends.

We could take the “marketplace of ideas” overly literally. The venture capitalists hunt for new ideas, the private equity firms find ideas that need more competent implementors, hedge funds allocate additional resources to underpriced ideas, and arbitrage traders take good ideas from one domain and apply them in another — a kind of stealing.

There are many roles that contribute to making this market efficient. Efficient markets are good. However, there are some ways in which our ideological marketplace analogy breaks down; for example, it assumes that ideas are verifiable.

Math is a domain that is verifiable, thanks to Lean, and similarly, you can easily check the share price of a given company to understand whether it has changed. With many domains, this is not the case. You might not know whether your idea has increased or decreased in value. There are many kinds of ideas whose merits are hard to evaluate.

Ideas are (probably) getting harder to find, and most new ones are (probably) combinations of old ones. If you believe both of these claims, then stealing becomes mandatory, and your bottleneck is verifying whether the chimeric idea is a good one. My naive understanding of OpenAI’s recent Navier-Stokes proof is that it involved combining ideas from multiple subfields in a new and unintuitive way.² It’s easy to verify mathematical chimaeras; human mathematicians have already been doing it for decades.³

A less-verifiable domain salient to me is the field of AI alignment; in practice, it’s difficult to understand how aligned an AI model is, and therefore almost impossible to evaluate the quality of a research agenda that aims to produce more aligned models. Another more universally applicable case study might be philosophy. What are the outcomes that a good philosophical idea might be trying to produce? They’re all somewhat contrived; they cannot really be quantified or compared to one another. Indeed, philosophy is essentially vibes, and yet there are many who would passionately argue that some philosophical ideas are better than others, or that you ought to try to develop your philosophy from first principles rather than simply stealing whatever “good” ideas others might have.

It seems like these fields really could do with having some mechanism that makes them more like a market. Markets are good tools for exploring idea space rather than having all participants cluster around a single idea; there is a reward signal built in that will encourage resources to be allocated to new ideas when it seems like new ideas will be valuable. Even monopolies can crumble in marketplaces, cf. Kodak or Nokia. Contrast this with alignment, where researchers and funding cluster around a small number of agendas. There is no baked-in reward for uncovering a new promising research agenda. The only legible signals of success are joining a project or getting funded or being recognised by someone who already has some credibility, and all of these are attached to existing agendas.

In more verifiable fields, the stealing is good because it improves the efficiency of the market; ideological kleptomania is likely the optimal strategy. Stealing ideas in these unverifiable domains is particularly tempting because it gives you a built-in proxy; something to aim for. But in those fields, originality is likely to be much more valuable. If you’re in the field of alignment, or indeed ufology, I am unconvinced by stealing; surely there are low-hanging fruit left yet unplucked by those who can resist the temptation to strive for legibility.