Worth stressing.

geohot.github.io//blog/jekyll... Very nice balanced post IMO

Ran this with Rays Bradbury's "The Veldt" Feels nice but way to short, Sol 5.6 must think im a trash-gamer

To the surprise of no one:

(bad finnish accent)

I do have to say, there's no feeling quite as good as hand-coding raw-dogging some basic stdlib C (if you dont care about memory-safety...) strlen/memcpy all the things

finally finished packaging my first slop-coded three.js app check it out at jakobsachs.blog/projects/tra... (data might load for a sec)

Made the mistake of checking up on RTX 3090 prices... wtf? these were 800 like 2 months ago And now 1.3k ????

I know the guy is already not popular in this slice of the internet, but like writing in 2024 that its laughable to write up that order of API revenue... and then getting body checked by reality.. Naive question, but what has Ed been verifiably correct on? In this whole field ..?

It took me a while, but i might be getting fable-pilled.. fuck

This summer: European dark ages 2.0 neural bogalo

Sometimes i forget that a "default Sonnet' out of the box is still very .. yeah

First time i ever managed to max out my 20$/month Codex sub, crazy value you get from that one

Do people do anything special for managing their smartphones (screentime etc.)? Currently writing a blogpost about phone restrictions/dumbphones, and curious if I'm the only one with no impulse control with these things

Im only ~40% through "the dream machine" but already i feel like i need a "Uhm actually Licklider already wrote about this in the 60s" sticker or something

The more and more i try to make Python like "production ready" i just end up writing a weird mixture of Go, TS and Rust Maybe i should just choose one..

OOP wants your program to be built like LEGOs, but its built like a Fugue sent skeet

Some last more tweaks and i might actually host this somewhere If anyone has a hint for cheaply hosting these "slop"-projects

This is it after running 5.5 on xhigh with the task to "make it more feature complete and visually coheren" via playwright for ~45 minutes

Idk if i just got lucky, but 5.5 oneshotted itself a "not horrible" UI without me telling it to

Okay i see why this is the case, but it still irks me As much as i like the ergonomics of it, i feel like stuffing Pydantic into each "official package" is a bad standard to set

Okay but beyond this, holy hell the model is good ?? I thought i knew roughly how well open-weight models where (long time Kimi fan) but wowie its just ... going ? And feels alot more eager then a Sonnet/Opus at this point

Crazy how it seems theres no MCP Client SDK that goes under 40MB RSS /20 MB USS

Thank you GLM5.2 , i try my best

I cant wait for the next era of Qwens, those O(1B) models gonna be the last piece i need for my total home-served assistant setup

Many such cases

the other problem is that it seems to make it much harder for people to rewrite a prompt when the output isnt what they want, and they start arguing with the chatbot instead if the output is wrong, arguing makes the problem worse with context pollution

I do think this is actually pretty adjacent to the factual truth

This is just the social media version of the Sophons from 3BP

whoa there 5.5. "need to say yes" ?

I wrote this while at the dentist, and looking back at it I maybe should've taken just a minute to do some spell checking..

Maybe a bit of crackpot speculation but SMIC has had relatively bad yields I heard, so the weird core per cluster count might just be due to heavy binning

For ML/AI people, its not terribly good though at ML workloads. Roughly 10℅ of an H100s BF16/FP16 flop/s The top500 list uses Linpack which is all Double Precision floating point

Okay so the story behind this is quite strange to me:

Fun fact, china just bumped of all american systems on the TOP500 list of HPC centers with an absurd amount of Cores Gotta sleep direly, but gonna look into what info I can find on that processor tomorrow

God I love being a computer toucher Instead of having to manually toil in my bank records (boring and lame), I can build a pipeline to pass the PDFs to local OCR model and then my own analysis ontop Instead of 20 minutes scrolling PDFs, its just 3 hours fixing Pytorch MPS issues Great success!

On top of this, i've been having alot of fun exploring the early scientific works in computer history. And one of von Neumanns significant works in this space is quite literally titled: "Probabilistic Logics and the Synthesis of Reliable Organisms from Unreliable Components"

Im not sure why everyone is so hyper concerned about indeterminism in LLMs/agents for coding. Its not like prior compiler builds where all that predictable, or OS specs Theres a reason "reproducible builds" is an ideal devs aspire to (and its not just called "builds")

Bought a random chinese label printer, but it only comes with a Phone and Apple App Store app (which looks sketchy af) So lets see how long it takes GPT 5.5 to crack it and write me a python script that copies the behaviour

Okay so this is pretty similar to what i envisioned here, but their evaluation setup isn't even close to what i would try arxiv.org/html/2405.15... Thinking how expensive it might be to run a more uptodate version myself 🤔

Since input tokens are more expensive then output tokens... is there a good argument for generating code with haiku and then verifying/reading with opus ? Since maybe only ... 10% of the tokens in the code might be something Opus would change.. do you just save hard-core on cost ?

This just feels very, nothing? Seems like ritsuko just is the normal peoples choice

The new stalker sequel is looking a bit different then i remembered, but I guess that's just what happens with Hollywood

With the new DiffusionGemma dropping i had to try myself (& Opus) at slop coding my first Three.js app. Tried to show/tease out the candidate-denoising process of the diffusion LLM

I'm so excited for the Chinese open weight distillation of Mythos without the middleware guardrails that will probably drop in like a month 😍

Our statement on the UK gov't's demand that all content on all devices sold or used in the UK be scanned, on the presumption of nudity, using a dystopian combination of age verification & content scanning. This proposal will not safeguard children. It endangers us all. signal.org/blog/pdfs/20...

The fact that speculative decoding works at all still feels magic to me Also shows that decoder-only transformer is unlikely to remain the default LLM serving architecture in the long-term

This but unironically louder for the people in the back when I have to explain what an agent is for the N-th time