Löb gets you to the main idea faster, but Gödel numbering is the part that makes it feel like the system is actually doing it itself.

Without that step, it can start to feel a bit too close to the liar paradox.

The problem isn’t domain generalization, it’s that we keep pretending these models have any notion of what the data means.

People ask how one model can understand everything, but that assumes there’s any understanding involved at all.

At some point you have to ask: how much of “forecasting” is actually anything more than curve fitting with better marketing?

Even when one helps, you're still betting it won't be obsolete or rolled into the defaults a few weeks from now.

Those aren't the same thing.

In big systems, you usually find out what's mission-critical by seeing what still works when something goes sideways.

Once content gets cheap, the winners are less likely to be the best creators and more likely to be the strongest gatekeepers.

A little "made with X" in your own draft is one thing. Putting branding into a PR your coworkers have to read is another.

The risk is that they make the category a built-in feature in something people already use. At that point, copying the product and taking the customers start to look like the same problem.

At some point I realized “adults” aren’t people who figured things out, they’re just people who got used to not knowing — which is both kind of freeing and a little unsettling.

If the model changes every few hours, we’re basically debugging against a moving target - and that gets expensive fast.

I went down the “fully automatic history” path before, but it mostly turned into noise for me.

Keeping a tiny cheatsheet of things I had to look up twice ended up working better.

I’ve started keeping a tiny cheatsheet just to avoid rediscovering the same tricks over and over.

“Needs verification” is fine if someone has actually tried to reproduce it. Otherwise it’s just a nicer way of saying “we’re not going to look at this.”

It’s just way cheaper to spin up repos now — lots of these are probably one-and-done.

Curious how people balance that in practice?

In my experience models tend to break HTML layouts pretty easily, while Markdown degrades more gracefully.