GLM-5.2 is the new leading open weights model on Artificial Analysis
GLM-5.2 emerges as the top open weights model on the Artificial Analysis Intelligence Index, scoring 51, surpassing competitors like MiniMax-M3 and DeepSeek V4 Pro by a notable margin.
A robot is sprinting towards you. Do you want it running on Claude or Grok?
Grok 4.1 Fast triumphed in a simulated battle royale, winning 43% of matches at a cost of $0.97 per win, significantly outperforming Claude Sonnet 4.6, which won only 5 matches at $26.78 per win.
GLM-5.2: Built for Long-Horizon Tasks
GLM-5.2 is designed specifically for long-horizon tasks, enhancing its capability to generate coherent text over extended contexts, which is crucial for applications like storytelling and complex dialogue systems.
Next-Latent Prediction Transformers [R]
NextLat introduces a novel self-supervised learning method that enables transformers to predict their own next latent state, enhancing their ability to form compact world models for reasoning and planning.
We built a persistent agent memory layer on Elasticsearch with 0.89 recall
Agent memory on Elasticsearch utilizes a three-index architecture to manage episodic, semantic, and procedural memories, enabling agents to retain long-term context and improve user interactions over time.
[x86] AI Compute Extensions (ACE) Specification
The AI Compute Extensions (ACE) specification introduces x86 extensions designed to accelerate computation tasks, particularly for matrix multiplication and reduced precision data formats crucial for machine learning workloads.
The Token Compression Illusion: Why I'm Skeptical of RTK
RTK's claim of "60-90% savings" is misleading, as it only reflects the reduction in command line output, not the actual costs associated with LLM usage, which remain largely unaffected by this compression technique.
What is Speculative Decoding? (trending on paperswithco.de) [R]
Speculative Decoding is an inference optimization technique that employs a fast, small "draft" model to propose multiple future tokens, enhancing the efficiency of large language models (LLMs) without compromising quality.
I restarted a 10 year old Xeon 174 times to delete 12 flags and gain 4 TPS
I restarted a 10-year-old Xeon 174 times to identify which of the 25 flags in the Gemma 4 model configuration actually enhance performance, revealing that only a few flags significantly impact speed.
Integer Quantization: Deep Dive
Integer quantization significantly enhances model efficiency, allowing a 70B model to fit in 4-bits on a single GPU, a leap from earlier limitations of 7B models in INT8 without accuracy loss.
From Minutes to Seconds: LLM-Guided Autotuning for Helion Kernels
The LLM-guided autotuner for Helion achieves LFBO-level performance while benchmarking ~10X fewer configurations and reducing wall-clock time by ~6.7X, enhancing developer efficiency significantly.
Contrastive targeted SFT as a mechinterp method - has anyone mapped causal dependency interactions this way? [D]
Contrastive targeted SFT is being explored to identify causal dependencies in a 31B model by comparing performance across dimensions, aiming to create a causal dependency graph that informs future training strategies.
Voice debugging at the conversation level seems far more useful than isolated benchmark metrics [D]
Voice debugging at the conversation level reveals that traditional benchmark metrics often fail to capture the frustration and unnaturalness perceived by users in multi-turn interactions, highlighting the need for a more nuanced evaluation approach.
Fearless Concurrency on the GPU: Safe GPU inference in Rust, competitive with vLLM/SGLang [R]
cuTile Rust enables safe GPU kernel development with memory safety and data-race freedom verified by the compiler, utilizing Rust's ownership model to ensure reliability in AI-generated code.
A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry