There is some research on layer duplication on existing pre-trained models. This paper also references layer skipping: https://aclanthology.org/2025.findings-emnlp.967.pdf

nice

The cost argument is wrong. Compare Gemini 2.5 Pro to Gemma 4 31B.

Released only 10 months later, a tiny open-weights model outperforms what was once SOTA. Fable and 5.6 Sol will be outperformed by laptop-class models next year.

Except no one will be able to afford those laptops. The hyperscalers will buy up literally all the RAM and high power GPUs. So you'll be stuck with 8GB of unified memory on anything affordable. A laptop of better capability will cost 50% more than today (at least). That won't be your biggest problem because some C-suite goon with stars in their eyes laid you off because of AI.

I’ve always used lagom.nl/lcd-test/

Haven’t needed it in years, but when it was still common that VGA was the only input into an LCD, this website had the images for easy calibration to perfect pixel alignment.

> Cloudflare will not let you run an image heavy site on its own

Cloudflare R2 object storage is S3-compatible, offers free egress, and does not share their CDN product's fair use policy against image-heavy use. Storage is 65% the per-GB cost of S3.

yes, you are right i didn't see their update: https://blog.cloudflare.com/updated-tos/

Thanks for sharing, this was a great read.