“alignment” is everywhere but either nobody (including EA, doomers etc) seriously believes they will reach AGI/ASI, or they believe “agents” will forever be caged superintelligent toys

https://x.com/richardmcngo/status/1882994095599370554?s=46

in case you were wondering why so many viruses were distributed as PDF files: PDFs can contain a lot of things, including multiple ways to execute code.
Here is a PDF file running Linux (open in a Chrome-based browser)
https://linux.doompdf.dev/linux.pdf

A pet rock is an agent; an incredibly successful and persistent one.
https://arxiv.org/abs/2502.04403v1

^ i'm joking but really that's a great food for thought and a way to resolve paradoxes about AI -- in a lot of discussions on "what's an agent" we were missing a dimension and simplifying the model a bit too much

look, zero trust, provably secure LLM inference! if you use TinfoilAI you don't have to trust them to keep your data safe, just check the code yourself!
https://tinfoil.sh/chat

I'm a bit annoyed of their reuse of the Tinfoil Chat name, which is another privacy tech, but oh well) at least both are cool)

i wonder when we'll get to training over binary ops (and if this is fundamentally possible) bc 1-bit as shown in the article is more like dictionary training, the matrices still contain real numbers
https://arxiv.org/abs/2502.05003

verifiable reasoning datasets. not a lot tho.
https://github.com/open-thought/reasoning-gym/

look closer
https://fixupx.com/yomix1337/status/1893686635239289307?s=46

science says we will remain smart as long as we want
https://www.science.org/doi/full/10.1126/sciadv.ads1560?af=R

we're coming to explicit reasoning about uncertainty and unknowns, at last!
https://arxiv.org/abs/2502.05244

if you have too much free time...
https://www.clockworkpi.com/home-devterm

complex/convoluted systems are always insecure because one gets tired and glosses over the details, instead of checking everything end-to-end.

https://github.blog/security/sign-in-as-anyone-bypassing-saml-sso-authentication-with-parser-differentials/

Oops Tencent guys beat me to distilling agent behaviors from LLMs into "regular" code 😅

We present PORTAL, a novel framework for developing artificial intelligence agents capable of playing thousands of 3D video games through language-guided policy generation.

https://zhongwen.one/projects/portal/

^ that was the original idea that led me to write a custom benchmark
https://news.1rj.ru/str/notatky/1292

https://ezyang.github.io/ai-blindspots/

AI Blindspots – Blindspots in LLMs I've noticed while AI coding

regardless if you use AI, almost all of these are simply good coding practices to adhere to anyway :) and there's quite a list, so take a look

https://thingofthings.wordpress.com/2017/01/10/meditation-for-people-who-hate-meditating/
interesting.

if you add x-middleware-subrequest header to your request to the next.js website it would be treated like it has already passed through the auth wall middleware. cozy php vibes

https://security.snyk.io/vuln/SNYK-JS-NEXT-9508709

that box is something.
https://x.com/austinbv/status/1903276706699546958

instantly wishlisted. portable Raman spectrometer is like a sixth sense, except it's real

Also, "Minimal ignition risk" is an absolutely lovely detail 😂

https://fixupx.com/jwt0625/status/1904562738833367531?s=46

build the code that builds physical objects!
also a thread with a lot of solid experience

https://x.com/seveibar/status/1905443905979715725?s=46

glimpse of ASI (narrow domain, still interesting as a workflow)

https://www.pnas.org/doi/10.1073/pnas.2406675122

backpropagation-free training, this time much more efficient. small datasets tho.
https://arxiv.org/abs/2503.24322