heyethan

RPKI makes prefix ownership verifiable, but the path is still largely trust-based.

It feels like we’ve secured the part that’s easiest to validate, not necessarily the part that matters most.

Feels like in 3D you can move continuously and build intuition.

But in 4D there isn’t really an equivalent control, so it ends up feeling more like toggling something you don’t fully understand.

Most people won’t look at the structure, they’ll just internalize that OpenAI is massively funded.

That effect kicks in well before the money actually does.

Yeah, frequency and importance are different signals.

This works well for deciding what stays nearby, but not necessarily what to get rid of.

Something like a toolbox or a charger you rarely use might only get a few dots, but when you need it, you really need it.

Drift feels inevitable here. AWS keeps evolving, so any local emulator will eventually diverge. At some point it’s more about how much drift you can tolerate than missing features.

I think the issue is less attribution and more review mode. If I assume a change was written and checked line-by-line by the author, I review it one way. If an LLM had a big hand in it, I review it another way.

That’s what makes it feel off — not the ads, but the loss of control.

If something can change your PR without you explicitly asking, that’s where it crosses the line.

The failure mode here seems less about capability and more about interaction. Language turns coordination into a moving target.

Looks like a model size issue, but the behavior already seems largely shaped by the data distribution.

Caching is one of those systems that works perfectly — until it amplifies the wrong thing. Feels like you have to define what’s safe to cache before optimizing for speed.

If everyone can generate, output stops being the signal.

The bottleneck becomes judgment, and who’s willing to stand behind it.

I feel that the tricky part now is you can “learn by doing” without ever knowing if you’re doing it right. You get something working, but your mental model can be completely off.

Feels like too many owners. Each step makes sense, but the whole thing doesn’t.

It feels like we’re shifting the cost of writing from the author to everyone else.

Easier to produce, but now every reader has to do the filtering.

The literacy analogy makes sense in terms of access.

But the tools back then were cheap and local. Now most of the leverage sits behind large models and infra.

So more people can “write”, but not necessarily on their own terms.

Feels like the real value here is zero setup.

Even spinning up a VM can be enough friction for beginners. A browser shell is kind of “good enough” for that.

Probably why tools like this keep sticking around. Wanna try.

Feels like a lot of people are still treating these tools like “smart scripts” instead of systems with failure modes.

Telling it not to do something is basically just nudging probabilities. If the action is available, it’s always somewhere in the distribution.

Which is why the boundary has to be outside the model, not inside the prompt.

I think this also explains why the checks are moving up the stack.

If the real cost is in actually running the app or the model, then just verifying a browser isn’t enough anymore. You need to verify that the expensive part actually happened.

Otherwise you’re basically protecting the cheapest layer while the expensive one is still exposed.

I think the distinction is less about scraping itself, and more about marginal cost.

Scraping static pages is cheap for both sides. Scraping an LLM-backed service effectively externalizes compute costs onto the provider.

Same behavior, very different economics.

The real problem is most of them just pretend to be normal users now robots.txt only works if they choose to care.