OpenAI CEO Warns: Don’t Blindly Trust AI – Insights from New Open‑Source Models

OpenAI CEO Sam Altman recently warned that users should be cautious about trusting ChatGPT completely, noting that AI can produce "hallucinations"—confidently incorrect or misleading answers.

Since its launch in November 2022, ChatGPT has become widely used in education, research, and even parenting advice, but concerns about accuracy, bias, and potential risks have grown, prompting Altman to urge critical evaluation of AI outputs.

In Germany, data‑protection officer Meike Kamp demanded the removal of the Chinese AI chatbot DeepSeek from Google Play and the App Store after it was found to illegally transfer user data to China, violating GDPR requirements.

Tencent announced the open‑source release of its new large language model, Hunyuan‑A13B, built on a Mixture‑of‑Experts (MoE) architecture with 800 billion total parameters and 130 billion active parameters, enabling deployment on a single low‑end GPU while reducing inference latency and computational cost.

The model excels in mathematical, scientific, and logical reasoning tasks, offers fast and slow thinking modes for simple and complex queries, supports agent‑style applications, code evaluation, and natural‑language tasks, and provides an API on Tencent Cloud for easy integration.

Key Features of Hunyuan‑A13B

Low‑Resource Deployment: MoE design allows operation on one mid‑range GPU.

Mathematics & Logic Reasoning: Accurate decimal comparisons and step‑by‑step analysis.

Fast Thinking Mode: Quick, efficient responses for simple tasks.

Slow Thinking Mode: Deeper, more thorough reasoning for complex tasks.

Agent Applications: Tool‑calling capabilities for travel planning, data analysis, etc.

Code Evaluation & Optimization: Uses the open‑source ArtifactsBench dataset for code‑related tasks.

Intelligent Q&A: Supports text generation and question‑answering.

Open‑Source Support: Code available on GitHub.

API Access: Available via Tencent Cloud.

Technical Foundations

MoE Architecture: Selectively activates relevant model components per input, reducing latency and resource consumption.

Pre‑training Data: Trained on 200 trillion high‑quality tokens covering diverse domains, enhancing general capabilities.

Multi‑Stage Training & Optimization: Improves inference, supports 256K context window for long‑form tasks.

Project links: GitHub repository and HuggingFace model hub . The team also released two datasets: ArtifactsBench for code evaluation and C3‑Bench for agent‑scenario assessment.

Google introduced version 2 of its Data Commons Python client library, enabling developers to query over 200 public datasets covering demographics, economics, education, environment, energy, health, and housing. The library supports custom instances, Pandas integration, API‑key management, and type‑safe operations via Pydantic, and is hosted on GitHub.