limoce
China's DeepSeek developing its own AI chip, sources say
China Blocks Meta's $2B Acquisition of AI Firm Manus
Claude Code Source Map
NeurIPS retracts overly broad sanctions list
Re:$500 GPU outperforms Claude Sonnet on coding benchmarks
on Thursday March 26, 2026 @09:47PM (#47538189)
The title should be "Adaptive Test-time Learning and Autonomous Specialization".
NeurIPS Statement on Sanctions
Declining to review for NeurIPS due to policy on sanctioned institutions
PetClaw: A Desktop Pet Companion That Works 24/7
Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search
Cursor's 'Composer 2' model is apparently just Kimi K2.5 with RL fine-tuning
Add NPU support for AICore for Gemma4 model
SteptronOss: Lightweight, AI-native training framework for large language models
Re:GLM-5 was trained entirely on Huawei chips
on Thursday February 12, 2026 @02:00AM (#46985649)
> This site is an independent informational resource and is not officially affiliated with Zhipu AI.
Re:Ship Types, Not Docs
on Tuesday February 10, 2026 @09:11PM (#46969914)
What about using LLMs to automatically update docs?
Adding Support for Qwen3.5
Reduce Complexity of Matmul to O(N^2logN) by Asymptotically Optimal Quantum Algo
Learning from context is harder than we thought
Residual Context Diffusion
Step 3.5 Flash
LongCat-Flash-Lite 100B A3B Technical Report [pdf]
Re:Which programming languages are most token-efficient?
on Sunday January 11, 2026 @10:13PM (#46583593)
I think separating thinking tokens from "representing" tokens might be a better approach, like what those thinking models does
Jinja.cpp: A single-header C++11 Jinja2 template engine for LLM chat templates
Mini-SGLang: A lightweight yet high-performance inference framework for LLM
Google drops Gemini 3 Pro image preview
Re:"One Student One Chip" Course Homepage
on Monday November 17, 2025 @08:43PM (#45960460)
Japanese kanji is not the same as Chinese characters.
Re:Precompiled headers and why Squid won't be using them (2023)
on Tuesday November 04, 2025 @09:35PM (#45818351)
Do C++ modules solve this problem?
DGX Spark may have only half the performance claimed
You Can Cool Chips with Lasers
Re:NVIDIA DGX Spark In-Depth Review: A New Standard for Local AI Infer...
on Tuesday October 14, 2025 @03:46AM (#45577276)
> ollama gpt-oss 120b mxfp4 1 94.67 11.66
This is insanely slow given its 200+GB/s memory bandwidth. As a comparison, I've tested GPT OSS 120B on Strix Halo and it obtains 420tps prefill and >40tps decode.