Toke

Evan as a skeptic, I will admit, it was interesting to hear about how Claude Code was created and how it is being developed now in this interview with its creator Boris Cherny:

Cherny’s instructions to build for the model they will have in six months, coupled with the seeming lack of understanding of what model they will have in six months (either software development goes away or an ASL-4 level catastrophe) was to be expected I guess? Maybe he knows and just isn’t saying? Maybe there isn’t very good understanding of whether one model is working better than another? The question of how these models are being evaluated for particular types of work, like software development, is actually interesting to me.

Of course, Anthropic employees would like nothing better than for people to forget how to develop software, and to become utterly dependent on them in the process. Indeed they are happily leading the way, high on their own supply of limitless tokens. They are counting on employers to follow suit, paying subscription costs to give their employees tokens to spend instead of having software developers on staff. This is following in the footsteps of what we’ve seen happen with cloud computing.

In some ways this is nothing new. Software developers have been dependent on the centralized development of compilers and interpreters for some time. So you could look at the centralization of software development into platforms like Anthropic and OpenAI as the natural next stage of development in information technology. Indeed, I think this is the argument currently being made (somewhat convincingly) by Grady Booch about a Third Golden Age of Computing which got underway with the rise of “platforms” more generally, and which includes recent genAI platform APIs and tooling.

But the big difference, that they want us all to forget, is the amount of resources it takes to build a compiler compared to an LLM and our ability to reason about them, and intentionally improve them. They also want us to forget that we need to, you know, give them all our data and ideas as context for them to do whatever they want (thanks cblgh). And as with cloud computing, they want us to forget about the materiality of computing, where computation runs. Ironically, I think computer programmers are particularly susceptible to this rhetoric of abstraction, or the medial ideology of the digital and the cloud (Hu, 2015; Kirschenbaum, 2008).

From a sociotechnical perspective I am curious how prompt data is being used to try to improve these models, as people start using them for ordinary tasks, and also in attempts to intentionally shape the model motivated by greed and malice. I guess the details of this process must be well hidden? Pointers would be welcome.

Hu, T.-H. (2015). A prehistory of the cloud. MIT Press.

Kirschenbaum, M. G. (2008). Mechanisms: New media and the forensic imagination. MIT Press.