Devstral thinks so

Is it normal for tool calls to launch a new context and session, feeding back messages from the original context, instead of running with the same kvcache as the user session calling the tool? It seems kind of expensive.

All 128 cores melting :( lol I might have to make the fans a bit more aggressive

Haha. I keep running out of Claude capacity. Can we build "Claude" at home? Barely. Tool calls seem to work. Can it fix itself? 😂😂😂 Messing around with llama.cpp and mitmproxy. Is there a better way to do this?

LLVM Clang Adds Support For @AmpereComputing "Ampere1C" CPUs - Presumably AmpereOne Aurora Ampere1C based on Armv9.2 https://t.co/ljyNHz2P44

Microsoft has been under fire for what is becoming of the desktop os....clearly attributed to AI.

Torvalds is being cool. Rust in the kernel. Him and his team are carefully reviewing things for slop. So I think the kernel will be in a great place. I'm very worried about userspace.

I saw a terrifying video of AI actually generating OS use images - like you can open a browser and it was just the AI dreaming up a browser window and opening up web pages or other apps.

Other actors want to reinvent the wheel or keep old stuff going. And it's now possible. When there are competing standards, what happens? Do we all get bespoke custom systems?

Some people are playing ball - building Wayland compositors to established standards, probably accelerated with AI, and really cool stuff is coming from it.

I can't tell if AI coding agents are a new beginning or the end for FOSS. It's clear there is a ton of fragmentation - coding agents can keep up with big money from the usual sources of standardization, ie Redhat/IBM/Oracle/etc.

Claude Code is the new Factorio. Absolutely one-shotting every single autist I know.

@shade_engine Back to mining coal

Why are people preferring Xlibre, or X in general, over Wayland? Some desktop environments don't have full Wayland support I guess but most do now.

Lol

Some of this is lol I wrote bad C++ code, skill issue...but interrogability in C++ is sometimes annoying, where things expect to be passed references to std::vectors, char arrays, etc. It quickly can turn into self-immolation.

The C code is actually doing LESS heap allocations...which actually means probably less surface area for memory bugs. And like opaque struct + inline storage provides some clear sense of ownership of objects, allocated on stack or static. No passing around smart pointers...

Removed dependency on flatbuffers, the serialization\deserialization code became simpler and shorter, because flatbuffer isn't exactly what I needed, I just originally bent to serialize\deserialize into my own objects...

I wanted something that serializes uint8\16\32\64, blobs, converts to network byte order where appropriate, and copies to an array to send out over the network. I typed up some basic function signatures, claude filled in the blanks...I wrote test cases and read the code...

Was using libwebsockets...flatbuffers... I'm porting to C with Claude, the resulting code is looking a LOT better. It's so far actually less code too. uthash replaces std::map and std::vector where I really needed it. Flatbuffers are nice but I always end up copying data