Evaluate an Open LLM with the EleutherAI Eval Harness: A Hands-On Guide
Evaluate open LLMs with the EleutherAI lm-evaluation-harness: run standard benchmarks, compare models fairly, and read the results like a practitioner.
Read Full News
Serve an Open Model at Speed with vLLM: Setup, Configuration and Benchmarking
Serve an open model with vLLM behind an OpenAI-compatible API: installation, the configuration flags that matter, and honest benchmarking with vllm bench.
Build a Simple AI Agent with Hugging Face smolagents: Tools, Code and Local Models
Build a working AI agent with Hugging Face smolagents: give a CodeAgent web search and custom Python tools, then run it locally or via the HF API.
From Vision to Language: Where VLMs Actually Fuse What They See
A Findings of EMNLP 2026 paper uses causal interventions to pinpoint where vision-language models fuse visual information — mainly when reading answer options.
The ‘Threat-Model Coverage Gap’: Why Automated Red-Teaming Isn’t Enough
A new paper argues automated red-teaming beating humans on cost and count doesn’t retire humans — benchmarks measure search thoroughness, not harm coverage.
“Don’t Drop Dropout”: The Case for Layer Sparsity in Efficient LLM Training
A Cerebras paper argues layer dropout belongs in state-of-the-art LLM training: up to 25% training FLOP savings and 1.5× faster inference, 2,400 experiments.
VoiceMem: Open-Source Memory With a ‘Left Brain’ and ‘Right Brain’ for Voice Agents
VoiceMem v0.0.1 gives voice agents a two-sided memory: a ‘left brain’ for facts and a ‘right brain’ for emotion and personality. Apache-2.0, from Tsinghua.
LG and NVIDIA Sign MoU: Bipedal Humanoid on Isaac GR00T Due Q1 2027
LG will build a next-gen bipedal humanoid on NVIDIA’s Isaac GR00T, with a public unveiling set for Q1 2027, under an MoU signed August 13 in Santa Clara.
Anthropic Open-Sourced Its AI Safety Harness — Here’s What’s Inside
Anthropic published the Automated Alignment Researcher harness on GitHub: the full agent loop, 10 benchmarks, integrity monitor — and one license caveat.
vLLM 0.28 Ships: 584 Commits of Inference Firepower
vLLM 0.28.0 ships with 584 commits: Kimi-K3 optimization push, DeepSeek V4 sparse MLA, tiered KV-cache offloading, and a new Rust/gRPC frontend.