Maybe macOS 26 was lot bigger... but I went straight from Sequoia to Golden Gate and freed up like 30 GBs. I didn't have enough disk space to download the update prior, so the extra 30 GB wasn't that... but could've been caches I guess. I assumed it was dropping Intel?

They already knew to drain the basin. Control-F in a PDF is way less annoying than watching a 9 minute long (for the algorithm) YT video.

Sure, you can search the video transcript... but all the YT video is doing is regurgitating the manual.

This was my thinking. It seemed easier to find the exact instructions for my dishwasher and most of the advice available was for newer models that had different control layouts.

Plus, the repair manual is full of flowcharts for diagnosing and resolving issues, which is very useful.

Did you not catch the emphasis? They did end up writing about it.

Sure, not for the same reason, but... got'em.

Tailwind is so pervasive it's 100% in pretraining _and_ RLHF.

I've done several Tailwind projects and rarely did models from various providers need to reference Tailwind documentation.

> go after profitable niches and avoid unwinnable fights

If you know Apple's history, this should stand out to you.

I desperately miss their old XULRunner app. No longer worked through the API as of ~3 years ago and was a sad day for me.

He didn't say 90% of the coding jobs. He said LLMs would write 90% of the code. As in be LLM generated.

If you've ever been confused about which turn is which with top down navigation, then this is for you. While driving a car, being able to at a glance know exactly what turn is what would have saved me countless wrong turns, my parents probably doubly so.

One classic: there's a fork in the road on the way to my parents. One side is "straight", the other side is an exit. I took the wrong, exit side of that several times while my app said to go straight before I just knew the route and didn't need GPS at all. Top-down isn't as good in these cases because of densely packed roads and back to back exits.

That said, it's "good enough" and I doubt this takes off.

For me, confusion in top-down navigation happens either because there's no bitmapped imagery being displayed, forcing me to guess where the turn is by mentally wrapping vague 2D geometry over the scene viewed through my windshield, or because the trace isn't drawn precisely enough over the imagery that is available.

I don't see much added value with this particular nausea-inducing approach. It falls into the category of "You should've thought less about whether you could and more about whether you should."

while I think that the widgets telling you which lane to put yourself in largely solve this kind of intersection confusion, I understand the problem you are describing

but this seems to me just a topdown and "third person" at the same time through the distortion, I don't see how it helps with knowing exactly where to turn, is it because it shows the way further down the turn (like, sideways after the turn in the demo) ?

I had to login the other day to use it. Probably getting slowly rolled out.

If you just want it to generate UI out of nothing, the benchmarks aren't really for that.

If you want to generate a UI based on specific user input of some kind, then they are.

I'd suggest using one model for UI and another model for tacking onto that UI. LLMs are great at pattern matching, and benchmarks don't really capture one-shotting desirable UI.

That said, benchmaxxing is a thing and your experience with models is a thing. Benchmarks are fuzzy and should be taken with a grain of salt.