Google Research
·
4 months ago
Google researchers conducted a real-world clinical feasibility study of AMIE, a conversational diagnostic AI system, with 100 patients at an academic medical center who interacted with the system via text-chat before primary care appointments under physician supervision. Zero safety interventions were required across all interactions, and AMIE included the final diagnosis in its top 7 differential diagnoses in 90% of cases, though primary care physicians outperformed the system in cost-effectiveness and practicality of management plans. The study demonstrates that supervised deployment of conversational medical AI is feasible and well-received by both patients and clinicians, but larger controlled trials are needed to quantify clinical impact.
Mistral AI
·
4 months ago
An autonomous agent built on Mistral's Vibe platform automatically generates and improves RSpec tests for Rails codebases by reading source files, validating against style rules, and running in CI/CD pipelines without human intervention. The agent's quality score improved from 0.68 to 0.74 through context engineering via a repository-level AGENTS.md file that provides step-by-step execution plans and best practices. The system uses custom tools like RuboCop linting and SimpleCov coverage checking to self-correct failing tests, with approximately one-third of generated tests passing on first execution before the agent iterates to fix errors.
OpenAI Blog
·
4 months ago
Researchers are developing methods to make AI agents resistant to prompt injection attacks that try to manipulate their behavior. The approach involves constraining which actions agents can take and implementing safeguards around sensitive data access within agent workflows. This reduces the risk that attackers can redirect agents toward unintended tasks through malicious prompts.
OpenAI Blog
·
4 months ago
OpenAI built an agent runtime that extends its Responses API with shell tools and hosted containers to enable agents to execute tasks with file access, tool integration, and persistent state. The system uses sandboxed environments that can run code securely while maintaining state across multiple interactions. This allows developers to create agents that perform multi-step operations without rebuilding context for each step.
Together AI
·
4 months ago
NVIDIA's Nemotron 3 Super model became available on Together AI's dedicated inference platform immediately upon release. The model offers a 1 million token context window and supports multi-agent reasoning workloads. Developers can now deploy the model on managed infrastructure without building their own deployment systems.
OpenAI Blog
·
4 months ago
Wayfair deployed OpenAI models to automate customer support ticket triage and improve the accuracy of millions of product attributes across its catalog. The system processes customer inquiries and product data at scale, though no specific volume or accuracy metrics are disclosed in the announcement. This reduces manual work for support staff and increases the reliability of product information customers encounter during shopping.
