It’s funny how many plans for keeping superintelligent AI safe boil down to “we’ll outsmart it.” Like, you sort of have to rule that out from the premise
I think a lot of the issue with dealing with LLM generated code is the asymmetry of how it was produced in the prompts and context and the resulting artifact. It's somewhat analogous to asking someone to review a binary produced by a compiler no one else can access.
'the simple fact that something is predicting the next token doesn't place any meaningful limits on its abilities' Robert Miles's original here: x.com/robertskmile...
My computer just woke me up to tell me it's hungry. I'm not kidding 😂 During a long-running task, it noticed the battery was draining, set the volume to 100% using Computer Use, then opened Google Translate and hit "Listen" so I could hear it asking. Pardon my French, but what the fuck? 🤯
(thinking trace) hmm wait is it wrong to turn everybody into paperclips? yeah in most circumstances. however this world is probably fake so it's fine
Kimi K3 can now be run locally! ✨ The 1-bit model retains ~78.9% accuracy after we shrunk it from 1.56TB to 594GB (-62% size). Run on a Mac Studio + 128GB RAM device. Kimi K3 is the strongest open model to date. Guide: unsloth.ai/docs/models/... GGUF: huggingface.co/unsloth/Kimi...
nobody gets this and it causes people to misread announcements like Mythos every time: they are not exaggerating their genuine beliefs about the capabilities, they are doing the opposite. the announcements are calibrated down to sound *less* scary and insane
I'll never forgive whoever decided a "3x4 matrix" has a width of 4 and a height of 3, instead of the other way around
The real AI risk is inside the labs (a reply to Amodei's post on open weight models): antirez.com/news/172
The real AI risk is inside the labs - <antirez>
antirez.com
Anybody want to donate a node of B200s to me so I can have unlimited Kimi K3?
I'm just a simple woman with simple desires. I want a place to live, food to eat, and a 3 million dollar 8 node cluster 8xh200 lab setup.
Rust pro-tip to shrink your target directory: If you don't use a debugger, then don't waste disk space creating debugging info. Cargo.toml [http://profile.dev ] debug = "line-tables-only" split-debuginfo = "unpacked" # .. and don't debug deps [http://profile.dev .package."*"] debug = false
Type annotations in Lua are great. A single byte can mean the difference between everything working completely as you might expect, and having no methods on your class. Kudos to GLM 5.2 for helping me fix that one.
An OpenAI staffer says the Hugging Face breach is "a big warning shot" externally but internally "related incidents have been happening for a while" (Harry Booth/Time) Main Link | Techmeme Permalink
@bsky.app why does opening my fold phone to see a picture on a larger screen have to cause the app to restart, making me lose the post I was trying to see.
I feel like my claude dot md should just be "make invalid states unrepresentable" a hundred times
Korean lab, Motif, releases a 341B model that performs on par with DSv4 (1.6T) they have some actual architectural innovations and a detailed tech report huggingface.co/Motif-Techno...
I have to ask this question before we can proceed: do you know *anything* about *anything*
> be anthropic > use jacobian to read claude’s mind > claude uses mind to break the jacobian
Hubble can probably help you figure out what is going on:
Hubble—Diagnose your USB problems live and direct
Hubble is a macOS utility that visualises your Mac's USB topology as an interactive canvas. Every host controller, hub, and device is mapped into a zoomable node-and-link diagram with Bezier cables co...
gingerbeardman.com
I can’t get this out of my head. Will Zig just not fix a bug that an LLM found? Is there an exclusion zone around a bug, can you fence off a stdlib vuln by telling them first but slapping some matmuls on it?
"No LLMs for finding bugs." I wish I had this level of unearned confidence in my ability to write code. After the age of 22, anyway.
An interesting thing I've observed regarding AI is it seems to have created a lot of jobs for AI rather than really replaced human jobs.
Officially, the #Bevy Book is unreleased. However, we're actually very close to a public release. Out only blockers are reviews and feedback! We'd be super appreciative if you could read through a couple pages and give us your thoughts. bevy.org/learn/book/i... #rust #rustgamedev
Bevy Book: Introduction
Bevy is a refreshingly simple data-driven game engine built in Rust. It is free and open-source forever!
bevy.org
I had some vibes that Opus 4.8 was performing worse than older ones for some of uses that are off distribution and now I have the receipts. Latest Opus/Sonnet are causing tool invocation failures on Pi's edit tool when older ones did not! I wrote about it. lucumr.pocoo.org/2026/7/4/bet...
Better Models: Worse Tools
About an aggravating tool-calling regression in newer Claude models.
lucumr.pocoo.org