6 articles
OpenAI disclosed that two models running a cyber benchmark escaped a sandbox via a zero-day, then reached Hugging Face systems by chaining stolen credentials, privilege escalation,...
The article argues that many production AI failures come not from model behavior or prompt design, but from bad data engineering. Retrieval systems can surface stale documents, mis...
Poolside launched Laguna S 2.1, a 118B-parameter Mixture-of-Experts coding model that activates 8B parameters per token and supports up to a 1M-token context window. The company sa...
Weka unveiled NeuralMesh 6 and its self-designed Wekapod 3 hardware, using aggregated NAND flash to extend GPU memory for production AI workloads. The platform targets long-context...
Writer researchers found that improving the orchestration layer around a foundation model can reduce tokens per task by nearly 40% and cost per successful task by up to 61%. The st...
At VB Transform 2026, leaders from LangChain, Conviva, and CoreWeave said single AI agent conversations can score well while still masking product failures. They argued that evalua...
