preyneyv

Wrote this after a lecture on Hays & Efros' scene completion using just lookup on 2.3M images. Made me think about where the "simple model + lots of data" bet actually leads. The article traces that thread through the Scaling Hypothesis to ARC-AGI-3, where frontier LLMs score under 1% on novel interactive tasks. I think LLMs built intelligence in the wrong order and I'm curious where people think I'm wrong.

ooh these are excellent, thank you for sharing!

Yeah, it's definitely one of those things where your eyes are "opened" when you use something faster. Kind of like headphones, what you have is good enough until you try something better.

Hard agree on VR / AR latency requirements. It's the first consumer use case in a while that's ahead of the average user's compute power, and I really hope it will mean good things for performant software.

I think you have it backwards. We get away with longer distances because of faster automobiles.

But imagine for a moment what would happen if we had faster automobiles and shorter distances: things would just be better.

This is exactly the issue. We can do things faster => we can do more things worse!

Honestly, this was a big one for me too. I remember being a kid and using Office 2007 on a terrible Windows XP desktop, and things just worked. Now I'm on a "modern" machine with "modern" software, and right-click takes half a second to pop up the context menu.

In the grand scheme of things, maybe not a big deal? But that's also how things get worse.

complete satire

really fun to build though