Unbundling AI and Graffiticode


Benedict talks about unbundling AI in this newsletter. It seems self evident that this will happen and that the consequence is that LLMs will be in every information product to automate knowledge processing.

His argument is that this will lead to a dramatic increase in scale in the same way that SaaS does. Not necessarily with more users, but with more use by the same number of users.

If true, then what are those uses? And how are they implemented? How specialized are the LLMs that enable them (how unbundled?)?

How might we use Graffiticode as programming interface for LLMs? This would amount to using a formal language to generate natural language prompts to cause LLMs to produce reliable outputs.

The compiler could produce different prompts for different models so that the client is insulated from changes in a particular model, or uses of multiple models.

Here is one possible use case: https://x.com/tylertringas/status/1754947979910234546?s=20