Nous Research designers are really cracked
https://distro.nousresearch.com/

turns out the backdoor is a door after all ¯\_(ツ)_/¯
https://tuta.com/blog/china-salt-typhoon-worst-us-hack

writing things down helps if you want to commit;
but also attracts too much attention to what you just wrote and distracts from other possible options
https://arxiv.org/abs/2412.06769

AutoReason: Automatic Few-Shot Reasoning Decomposition
TLDR: explore simpler assertions in first prompt, draw conclusions in the second
https://arxiv.org/abs/2412.06975

cooperation and coevolution is the real alignment, not any kind of standardized testing (which will just get overfit)

Cultural Evolution of Cooperation among LLM Agents
https://arxiv.org/abs/2412.10270

https://x.com/conspiracyslyr/status/1868737204916552052?s=46
flat earthers visited the south pole, confirmed the earth isn’t flat and got expelled from the flat earth community 😅

Tachikomas (from the Ghost in the Shell universe) are now real!

this is how you do observability. or logging, or tracing, whatever you call it. preaching this for years, now somebody wrote the guide so I don't have to.
https://jeremymorrell.dev/blog/a-practitioners-guide-to-wide-events/

matches my experience 100%. I take a lot of notes, and they are super useful, but...

when I don't care how I might write something down that's like 3x boost to thinking speed and ~2x boost in available "context size", in LLM parlance

upd: actually, one of my biggest disappointments in childhood was that the words don't map 1:1 to what we now call "concept embeddings"

PDF: https://wz.ax/2024/language-is-a-tool-for-communication

TLDR: Lesswrong discovers B2B sales tooling of 2020
(yes LLMs made it a bit easier to built but it's not news really. probably existed years earlier)
https://www.lesswrong.com/posts/GCHyDKfPXa5qsG2cP/human-study-on-ai-spear-phishing-campaigns

this is an important reminder that all digital electronics is really analog but with good correction circuitry.

and run-time cpu and memory error rates are always nonzero too, though orders of magnitude lower than chip yield rates

https://cerebras.ai/blog/100x-defect-tolerance-how-cerebras-solved-the-yield-problem

I got tired of AI puppets everyone calls "agents", followed the lead of Nous Research and made a bit more autonomous pet rock.
This time without OpenAI and with entire LLM in the TEE.

https://open.substack.com/pub/cortex/p/cortex-stand-alone-complex?r=1clcn&utm_campaign=post&utm_medium=telegram

with deepseek R1-distill you can now run an o1-level reasoning model at home 🤯

https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero

realtime special effects in games are much harder to make and it's cool to peek under the curtain
https://www.youtube.com/watch?v=HPqGaIMVuLs

maybe thinking is a bicycle built for two 🤔

If we were designed to think solo, monologue would be easier than dialogue.

Dialogue involves INCREDIBLY complex acts of prediction, coordination, task-switching and mind-reading--yet we find it MUCH easier than monologue.

Why? Maybe thinking is a bicycle built for 2.
https://x.com/AgnesCallard/status/1882866517077114888

so most deepseek rants are noise as always but this is real shit: complex prod code editing example, with all prompts

https://github.com/ggerganov/llama.cpp/pull/11453

Robots with tentacles!
https://arxiv.org/pdf/2303.09861

“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