I’m currently building cell, a programming language/environment (and if you read any other recent article of mine, I must apologize for the ad nauseam repetition of that fact). Unlike the ustack, I do want someone else to use it (hopefully, many people, with many different use cases). The big problem is that if you build something new, it is unfamiliar, and it’s quite difficult to drive adoption of unfamiliar technology.

I think that, this time, I have two saving graces that might shift the impossible into barely possible territory.

• LLMs: LLMs quickly understand unfamiliar technology, as long as you explain the rules. If the unfamiliar technology has an LLM integrated into it, then half the problem disappears.
• Results are shown: next to each call, cell also shows its result. That means that even if you don’t know what a particular function does inside, you can see that it’s called X, receives Y, and returns a value Z. That’s usually enough to figure out what’s going on, at least as a first approximation.

In the time-honored tradition of extrapolating patterns from a single data point, I state the following:

• Any LLM will halve the hardness exponent of adopting new technology.
• There has to be a particular aspect of that new technology that further reduces that exponent. I wouldn’t be surprised if that aspect was always related to feedback from the system to the user.