tim

new AGENTS​.md --- this document exists for non-obvious, error-prone shortcomings in the codebase, the model, or the tooling that an agent cannot figure out by reading the code alone. no architecture overviews, file trees, build commands, or standard behavior. when you encounter something that belongs here, first consider whether a code change could eliminate it and suggest that to the user. only document it here if it can't be reasonably fixed. ---

@0xSero awesome, let’s do that!

@NERDDISCO I definitely need help. The codebase is in a real mess so I need to slowly swap out components until it’s scalable. Maybe we can work on the vLLM piece together

I am going all in on vllm-studio, in the past my take was that if Claude can do it out then people should figure it out. I've also just been doing whatever comes to mind, but I am going to trim out most of the code and focus on a desktop electron app. Good UX coming soon

@Prince_Canuma 👀👀👀

Unlocked 4x speed up with this 🚀

@LukyVJ 😂

@NERDDISCO Lmao 🤣 at first I was like « great job » Then, I read the tweet 🐥

asked codex to create a clean layout without any pills or cards, i think it did a great job

@andrey_cheptsov have no time, need to respond

@NERDDISCO I bet you are behind the scene reading the messages and clicking buttons)

one of my projects is finally live

@crislenta @supercell @ipaananen wow, this looks super awesome

😵 5 star hotel for a private AI hackathon by @supercell WE HAVE A PRIVATE CHEF > incredible breakfast > snacks, fruits, drinks > claude code credits 😂 > the goat @ipaananen in the house > vibe is off the charts THE FINNS ARE SETTING A NEW STANDARD

@0xSero as it should be faster than llama.cpp?

I am going all in on Exllamav3 This is the middle ground between fast, performant, works on consumer cards, and intelligent. VLLM and Sglang are my go to but they're too finnicky below certain bits.

@levelsio yes

Okay let's see who can reply to this

@Prince_Canuma this is super cool

Just implemented Google’s TurboQuant in MLX and the results are wild! Needle-in-a-haystack using Qwen3.5-35B-A3B across 8.5K, 32.7K, and 64.2K context lengths: → 6/6 exact match at every quant level → TurboQuant 2.5-bit: 4.9x smaller KV cache → TurboQuant 3.5-bit: 3.8x smaller KV cache The best part: Zero accuracy loss compared to full KV cache.

1.0107 15.5mb ttt lr=0.0032, 12ep

1.0516 15.7mb ttt lr=0.0008, 8ep, 8 blocks

1.1276 = autoresearch + 8xh100 sxm for 48h 12l int5 gptq, mlp 3.5x, leakyrelu(0.5)², full mha, gated attn, value residual, ema 0.997, ttt 3ep, xsa-all, partial rope, ln scale, ve128, bigramhash 8192, early qat, prune 2% thx to @runpod for the compute o7

@max4c_ @bcherny nothing less o7

@bcherny @NERDDISCO goals for runpod labs

Little known fact, the Anthropic Labs team (the team I joined Anthropic to be on) shipped: - MCP - Skills - Claude Desktop app - Claude Code It was just a few of us, shipping fast, trying to keep pace with what the model was capable of. Those early Desktop computer use prototypes, back in the Sonnet 3.6 days, felt clunky and slow. But it was easy to squint and imagine all the ways people might use it once it got really good. Fast forward to today. I am so excited to release full computer use in Cowork and Dispatch. Really excited to see what you do with it!

@exsponged @OpenAI @runpod dm

@NERDDISCO @OpenAI @runpod hey! would you sponsor a student club in Minnesota to collectively work on this?

Are you up for a challenge? openai.com/parameter-golf

@0xSero 👀👀👀

For those interested in getting into local AI this is my most important video. youtu.be/Adliwsf2oPE