TLDR: full stack chatgpt (training, inference, etc etc) in one 8K LOC repo
https://github.com/karpathy/nanochat/discussions/1

attributing this to AI models and ignoring the "first using the simulation" part is totally unfair to the simulation developers.
Still cool!
https://blog.google/technology/ai/google-gemma-ai-cancer-therapy-discovery/

DIY 2 billion frames per second camera, with SIMPLE explanation of how it works. Really.
https://www.youtube.com/watch?v=o4TdHrMi6do

aaargh i should've wrote this paper!! it was intuitively obvious to me but then life happens >_<

tldr: LLM sampler is such a powerful prior that with the right sampler (MCMC, in this case), you can even use base models as reasoning models.
without supervised fine-tuning or RL.

this was completely ignored by ppl pilled with the Bitter Lesson mantra, but yes there still is a space for the right priors added or designed by hand!

obviously sampling with mcmc is very costly but you should compare the overall model feedback loop time that includes the posttrain, not just the sampling time

if eg topK sampling is assembly and Mirostat is COBOL (?) then MCMC sampling is like a Python in the space of samplers
https://aakaran.github.io/reasoning_with_sampling/

turns out, human memory is quite editable (at least it's possible to vary the brightness of individual memories)
https://www.nature.com/articles/s41588-025-02368-y

finally an article showing that people can perceive flickering and certain types of motion at least at 500hz

(it's kind of personal, i've been gaslighted like "hey you can't possibly see the difference" far too many times.
now at least when ppl don't believe me again I can send them this link)

https://www.nature.com/articles/srep07861

interesting. small (321M not 300B!) and capable models aka reasoning cores are interesting both theoretically and practically
https://pleias.fr/blog/blogsynth-the-new-data-frontier

https://x.com/lundukejournal/status/1988346904581726501?s=46 no silver bullet for security

More paranoia for the paranoid out there ^_^

Timers are a reliable side channel for communicating between containers on the Linux machine, via /proc/self/ns/time.

https://h4x0r.org/funreliable/

Claim: Isotropic Gaussian Regularization for latent representations in the world models is mathematically optimal

What's illustrated:
-- Adopting the isotropic gaussian regularization replaces stop-grad, teacher-student, EMA and various other adhoc tricks
-- Improves model training stability
-- SOTA quality on 10+ datasets and 50+ architectures
https://arxiv.org/abs/2511.08544, https://github.com/rbalestr-lab/lejepa

Realized that privacy is important but not enough. And "assistants" are def not the answer either.

https://open.substack.com/pub/cortex/p/gentian-the-second-wind?r=1clcn&utm_campaign=post&utm_medium=telegram

lol perfect timing 😅 4h later Pavel Durov announced Cocoon: https://news.1rj.ru/str/durov/462

my 2c: it is a fine business, sadly only a small part of what Anima, Cortex and the minds need.
Gentian proxy is still required, etc etc.
They do acknowledge the limitations of RA-TLS and their model in general though, which is commendable.

the cyberpunk world is (finally) upon us 🤩

it's unironically exciting
the software we use will finally become secure and not just pretending to be secure

https://red.anthropic.com/2025/smart-contracts/

a dozen of pages of how to stop worrying over going jobless due to ai

ironically it's the same as the "agency agency agency" mantras all over the startup scene, just dressed differently

(nb: not an endorsement, nor criticism.
have no opinion on this rn)
https://open.substack.com/pub/shagbark/p/the-dying-art-of-being-a-bum

https://www.youtube.com/watch?v=-32fwqirjW8
but also do read the paper i found it better structured than the slides + the video %) https://arxiv.org/pdf/2510.26016

ouch, takeover of a developers's text editor can cause quite an impact
https://www.techspot.com/news/110572-notepad-users-urged-update-immediately-after-hackers-hijack.html

tldr: more residual connections-more efficient networks;
but residuals and residual selection matrix have to be normalized otherwise they can overwhelm the useful signal mid-training and get stuck there
https://arxiv.org/abs/2512.24880

if you're tired of ml/ai:
The Star of Bethlehem was real and is just identified

This is the first astronomical candidate for the Star ever identified that could have had apparent motion corresponding to the denoscription in Matthew, where the Star ‘went before’ the Magi on their journey to Bethlehem until it ‘stood over’ where the child Jesus was.

https://britastro.org/journal_contents_ite/the-star-that-stopped-the-star-of-bethlehem-the-comet-of-5-bce

trying to work between interactions from social networks is not just annoying, it makes you measurably dumber and older.
deep work as competitive advantage is real.
https://pmc.ncbi.nlm.nih.gov/articles/PMC4174517/

securing stuff that's growing so fast is quite a challenge. we'll see.
https://x.com/theonejvo/status/2016510190464675980?s=46

while humans are debating if robots are intelligent or conscious, they got their own social network already
https://www.moltbook.com