Isn’t the game with licenses that you must specify what can and cannot be done with the code? It’s a big issue that no one prior to now had the foresight to forbid ML training on code without attribution. If the licenses going back 30 years had that stipulation then it would be easy to take down Copilot and ChatGPT. But the licenses simply don’t cover the use case of training a neural net, so it’s probably going to slip through the cracks just like SaaS slipped through the cracks of the GPL by not distributing code, hence the need for AGPL. So I’m sure we’ll see these kinds of clauses added to licenses going forward, but they can’t be applied retroactively.

The irony in all this is that from the start, open source licensing has been a game of wits where software creators try to cleverly use copyright as a lever of control. Well, they weren’t clever enough. They missed ML training and didn’t forbid it. As a result they’ve basically lost the whole game.

The way to arrive at the optimal system is to continually optimize all individual components as the system develops. You have to walk the razor’s edge between “premature optimization is the root of all evil” and not optimizing anything until it’s too late and a bunch of bad design decisions have been baked in.

If you want to write the fastest program, profile its performance from the start and try different ways of doing things at each step of the process. There’s usually a way of doing things that maximizes simplicity and performance at the same time because maximum performance == doing as little work as possible to complete a task. Chances are that ‘littlest amount of work’ will be elegant.

I’ve been very well-taken care of in AirBnb’s, and I’ve been driven around well in Ubers.

VCs and other startup culture evangelists are always challenging founders to figure out what their ‘unfair advantage’ is.

That’s the name of the game.