Deep tinkering vs deep coordination

In my post on hackers, wizards, and scientists I ended with the following tension:

1. Fields where individuals can cheaply iterate on technical ideas with tight feedback loops tend to see explosions of innovation. These kinds of deep cycles of tinkering, scaffolded by quick validation, are often only made possible as a byproduct of existing commodity tooling. For example in computer science: early innovations in both computer graphics and AI were significantly accelerated by the proliferation of cheap personal computers, and later by affordable graphics cards.
2. This mode of iteration only really scales horizontally¹. However, in many fields, the requirements for deeper coordination make this more challenging. Individual tinkerers start to block on each other for resources and their quick feedback loops come apart. For example in medicine: innovations are currently mostly driven on the margin by enabling existing teams to improve what they are currently doing, rather than by adding new tinkerers.

Right now, many interesting parts of computer science can feel like they are moving from being "deep tinkering" fields to "deep coordination" fields. There are exceptions to this in both directions, but one area which dominates as an example is frontier AI research, where even well funded university labs have been losing top talent who are willing to trade less academic freedom for more access to resources.

I think there are two main gaps in this picture.

Firstly, having fewer resources isn't a bottleneck on its own. For example, part of what made demoscene work (see also image below) so impressive were the cool visual effects which graphics hackers were able to achieve with such limited resources. While individuals were resource constrained, the benefits they were getting from tight, fast iteration more than made up for this. Similarly people working on open source AI models have been able to figure out some very clever techniques to squeeze out surprising amounts of performance from smaller amounts of compute.

Secondly, even in the most romanticised periods of 20th century "hacker" culture, tight tinkering loops only got you so far. Specifically they mostly only got you as far as being able to build good demos². Even in a field like software with very low marginal costs, as soon as you need to take work into production, distribute it to many users, maintain it, keep it secure and safe etc. coordination very often becomes the limiting factor once more.

As far I am concerned both the original tension I've drawn and these two gaps I have named point towards the same moral: a healthy field is one where there are low barriers to deep-tinkering work with smooth handoffs to deep-coordination work at precisely the right points. I think there are three main things required in order to achieve this:

• Figuring out when the handoff should take place
• Making tinkering up to the point of handoff cheap, safe, and accessible
• Making the handoff possible and attractive

There is much to say about all three of these barriers, but I think the last one is especially underrated. In many cases work built by deep-tinkerers is inscrutable to larger teams who try to make sense of it, and they don't spare much attention to trying to make it legible. Conversely, deep-coordinators will occasionally invite outsiders to contribute to their research, but don't properly think through how they will interface with the work that gets sent their way, nor how to make the barriers low enough that good work gets sent their way in the first place.

Closing Thoughts

I've spoken about the science fiction genre and general aesthetic of cyberpunk in several contexts this year. The genre has many features, but one recurring theme is the dramatization of the contrast between the two cultures I describe above. It portrays: on the one extreme: deep-tinkering as the domain of a seedy low-life underworld of alienated hackers who are limited to working on scrap hardware they steal or repurpose; and on the other extreme: deep-coordination as the domain of a corrupt and bureaucratic system of bloated but ruthless organisations who are deeply adversarial towards acknowledging the risks, or sharing the benefits of, the technologies they develop.

The reason I bring these exaggerated fictional archetypes up is not in order to dismiss them as such, though I do think recognising where they have been exaggerated is important. Instead I think we often miss the point of the context in which they were originally introduced. That point is that the core takeaway from cyberpunk, as with any genre of dystopian fiction, is that we can hope for better.

Something I've learnt both as a result of spending more time recently with resource constrained academic researchers and with non-academics who blog about their diy-science work is that we already do have it a lot better than this. Somehow good work done by deep-tinkerers eventually does find its way to the right places, even if this happens more slowly and less often that I would like. In fact, a part of what is so cool as that in spite of the pipelines from deep-tinkering through to deep-coordination being so imperfect and broken in so many ways a lot great work still manages to bubble through.

This makes me optimistic that we can improve these pipelines even more, through a combination of better vibes, and better infrastructure.