Was it inspired by this?

https://housing.over.pizza/

I've never seen this, but this is amazing!

Discussed recently here:

A city-building game in which the city would prefer you didn't - https://news.ycombinator.com/item?id=50036864 - Oct 2026 (201 comments)

Tried to use it and the error messages and documentation were overwhelmingly unhelpful. Spent about an hour on it. Maybe it will get better though, good luck!

I wish the useful improvements would be ported to https://github.com/open-annihilation/open-annihilation which seems a better open source project.

It feels lke they replace older models with "optimised" versions which are cheaper to run, but keep the same name.

You mean like quantized?

Yes; or something which has a similar effect.

It would describe the observed behavior. Especially if there were an internal quant/efficiency team that wasn't as diligent about regressions as the primary model team.

That's exactly what it feels like and it makes perfect sense, given the known compute constraints.

For pixel owners: the Google photos app cannot be uninstalled, but you can disable it using adb.

The only reason to buy a pixel (new or refurbished) is to run Graphene or Lineage anyway :)

There are reasons to remain on Google Android for now, though. For example, QuickShare compatibility with AirDrop[1,2], or compatibility with certain banking apps[3].

[1] https://www.android.com/quick-share/with-iphone/

[2] https://github.com/GrapheneOS/os-issue-tracker/issues/6550

[3] https://grapheneos.org/usage#banking-apps

> "Instead of improving an agent against a fixed test, we let the evaluation evolve alongside the agent"

This quote should have been highlighted earlier in the article.

Yes, this is key

It defines and introduces a lot of concepts/acronyms in the thinking blocks which we normally don't read.

> If the model was designed specifically to quantize down to 1.58b, then it's different.

> AFAIK, there's no large models designed for this yet.

Isn't BitNet b1.58 2B4T what you are looking for? (haven't tried it myself though)

2B is pretty small...

No 100B+ param (certainly no 2T+ param) models have been trained natively to quantize down to 1.58b.

> Quis custodiet ipsos custodes? is a question that is roughly 2000 years old. Still no good answer.

I feel like we are moving towards a very specific answer, without much thought: AI.

Did anyone calculate this for OpenAI?

Yes, in the linked sources, SemiAnalysis did test for Codex too: https://x.com/SemiAnalysis_/status/2064815044085318040

Similar effect, but even 70 times higher compared to Claude, 40 times.

Looking forward to giving this a try with llama.cpp. I’m watching the open-weights competition with high expectations.