Andrew Lamichhane

I've never had this much fun playing Elden Ring Nightreign with the boys, it's a peak game, but it's kinda disappointing that it's not getting updates anymore. I've started playing a few mods and it's insane to see what modders can really do.

@ppok24 Yes its a node binary, there's react core components which is not web agnostic I think, and the react Dom for web renders, but for tui they use ink as the renderer

@azeus333 technically it's a node binary right? since react is a framework for the frontend (correct me if i'm wrong)

why is claude code written in react ?

This week's focus 1. Finish the GPT-2 paper (read the BPE paper last week) 2. Complete my ROS2 UR5 simulation + sorting task 3. Figure out how to properly model Gymnasium environments 4. Hopefully get back to the gym (that minor injury slowed me down)

@andrewbrown I prefer the dnd version

If we are redesigning GitHub lets do it in the style of Final Fantasy / Dragon Quest.

@itsabhiyan A single session of claude code might be 300+ mb and when you add mcp into the equation that's just crazy, shipping fast should be priority for any company but they gotta have a set of standards too fr

@azeus333 yeah... if it was a startup about AI agentic CLI tool, then may be you could consider, because you NEED to ship fast. but this just doesn't make sense. from any point of view. People at claude can't be this bad...

@itsabhiyan Honestly this is just bad engineering, and all companies and engineers accept it due to its popularity. Javascript for everything isn't a joke anymore and now we have next js shipping with a compiler

@azeus333 yeah... looking at bugs people are facing and all... it shows and codex took the route of minimalism claude code took the route of cramming all the features not testing them properly before shipping facing consequences

@itsabhiyan From what i have gathered it's mostly because of the vast amount of data available on typescript so it's easy to vibe code and ship faster(claude-code, opencode etc) The approach that codex took is way better, i guess its just poor decision of javascript frankenstein

@azeus333 yeah... I never got that reason, and they boast making it faster. You are building a TUI, not just CLI. If CLI, it would make sense... to just use for convenience. But TUI doesn't make sense to me. Not in the slightest. Unnecessarily heavy and slow

We can finally retire the 'Java is too slow' meme. It's 2026 and Netflix is running 3,000–4,000 production services on Java with near-zero GC pauses Amazing talk! youtu.be/ucJTPda_zx0

When interview is coming up, it seems so hard to focus on any other thing, had to pause all my projects

Fixed the issue. Turns out it was a Gazebo sim bug: camera_info follows the REP 103 standard, but point cloud generation in Gazebo sim doesn’t. There are two possible approaches: 1. Use a additional frame to make it align with rep 103 2. Use depth_image_proc(i chose this one)

Buh my emacs theme is generated by claude and the image by chatgpt, i freaking love it

Okay i don't understand why this is happening, i set the camera according to rep 103(z forward[points to the scene], x right, y down)

I don't think i'll be doing anything other than fixing my emacs config this weekend

水面に映る雲を擬似的に表現してみた。全然リアルじゃないけど雰囲気良いい！ この視点でも空の高さを感じる #indiedev #ゲーム制作

@KAIJIUUUUU buh this lools beautiful

I have been absolutely binging on higanjima by Koji Matsumoto, just ended up on "higanjima 48 nichijo" and found out that the English translations are behind atleast 400 ch 💀, might as well learn my kanjis, if anyone knows some good horror managas drop em #manga #horror

Set this up a while ago, but switching to the new Gazebo Sim from Classic has been a real hassle. I want to learn proper pick-and-place with the arm and eventually do Tower of Hanoi. Tried something similar in Humble but never finished it, so this time I’m documenting it as i go.

ともだちコレクションでスペックを描きました。

I also read Karpathy’s recent post on his LLM Wiki / knowledge base approach and I’m tempted to experiment with it as the core of long-term memory combined with a minimal RAG system. Anyone else playing with similar memory designs? Would love thoughts. (3/3) #buildinpublic

I also sketched out a basic idea of what i want my agents memory to look like : 1. User Details (profile, preferences, background) 2. Conversation (short-term chat history) 3. Long-term (important facts, memories, knowledge) 4. Active Context (what’s currently ongoing) (2/3)

Day 2 of working on my personal agent for fun Spent a good chunk of the day debugging a streaming issue in llama.cpp built a simple but functional memory manager from scratch. It handles conversation sessions, message history, and persists everything to a JSON file for now (1/3)