Agent threading. Insane. It makes sense that this would happen eventually but it really means you have to fully trust each return if it were to actually have a benefit in terms of time saving.

Have you considered giving some free letters for newbies? What are your thoughts on the minute cryptic people?

"I'm handing a spoon to you"

What's wrong with the guardian way of doing it? Is it because you want one app for all your different news sources?

Well said. I think this flag represents them best

Contexts (basically independent text files) that you can choose to append to and re-summarise etc with their own version control is what I want. Ideally you can use multiple at once for input.

I saw someone you had on your YouTube with similar they'd done themselves via command line

Nice. I've been meaning to do some of these so every time somebody (see GPT5) changes everything I have a way of rationalising how things might change for me practically. I really want an "explain <some topic I can explain very well> in x words or less" one to check style

Really shows the worth of your test to be honest. I think there should be a name / classification for these kind of tests. Subjective but clear consensus on a good answer. Ideally I would run ~10 of these kind of things for every new model based on what I use models for.

Also can be repurposed as a Saddam depending on demand.

All the people waking up, remembering what career they have,and screaming

Thanks so much! I absolutely love the puzzle. I thought I had nailed it and looked up the answer and I was wrong. 🤦‍♂️

I just wonder if that's all you need. Good memory and good inference. Is intelligence that simple? Is inference more important than memory? It's all getting very interesting that's for certain!

Wow, that's bonkers. Food for thought. It very much feels like more of an ARC prize style question in many ways. Slightly exaggerating but you don't really need to 'know' anything other than what a rectangle is.

Indeed. But I bet you could say that about the kids doing the test too. Guarantee their training data / regime is pretty good.

Also particularly interesting which question they left out. I would say it's the most visual of the lot and the most accessible for humans. I imagine there aren't enough similar questions out there in training data for that one too.

Maybe they're just being nice and I'm a miserable cynic, I hope so!

Totally a ploy so they can invalidate your opinion if you play less than four hours (everyone except people who dig it)

Cool, no documentation it is then!