A developer built a 10-scenario benchmark testing five LLMs against indirect prompt injection, where malicious instructions are hidden inside tool outputs rather than user input. Claude Sonnet 4.5 and…
AgentBrew, an offline training framework from researchers publishing on arXiv (2609.05837v1), learns tool-use policies from a single batch of raw interaction trajectories without task verifiers or ite…
Bridgewater Associates' AIA Labs and Thinking Machines Lab reported in June 2026 that fine-tuning an open-weight model, Qwen3-235B, on expert-labeled data achieved 84.7% accuracy on six financial info…
By mid-2026, open-weight AI models have closed the performance gap with proprietary APIs for 90% of enterprise tasks, making the choice a matter of unit economics and infrastructure rather than raw in…
AirLLM, an open-source inference library, claims to run a 70B-parameter large language model on a 4GB GPU without quantization, distillation, or pruning. The claim is technically true, achieved by loa…
AirLLM, an open-source inference library, now supports running the 2.8T-parameter Kimi K3 model on a single 4GB GPU, using only 3.72GB of VRAM on an RTX 6000 Ada, by streaming one expert at a time for…
A new study testing GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Flash, and Qwen3-235B on 101 World Values Survey questions across 13 language-country slices found that prompt framing is a first-order deter…
A new study on arXiv (2607.22925v1) shows that frontier language models can perform invisible reasoning using semantically irrelevant filler tokens, with accuracy improvements of up to 13 percentage p…
Bridgewater's AIA Labs and Thinking Machines Lab fine-tuned a Qwen3-235B model on the hedge fund's internal investor judgement, achieving 84.7% accuracy on triage tasks versus 78.2% for the best front…
Bridgewater Associates and Thinking Machines Lab fine-tuned a Qwen3-235B model for financial tasks, achieving 84.7% accuracy and outperforming GPT, Claude, and Gemini at roughly one-fourteenth the cos…
NVIDIA's DGX Spark workstation can be linked via a single 200 GbE cable to form a two-node AI cluster, aggregating 256 GB of unified memory to run frontier-scale models like Llama 3.1 405B locally. Th…