Tencent Open-Sources Hy4-preview MoE, xAI Sued for Child Abuse Content in Grok Training | AI Daily Brief
Audio in Mandarin Chinese · English transcript below
⚡ Tencent drops 770B Hy4-preview LLM, igniting open-source race; xAI hit with class-action over training data; ex-OpenAI Barret Zoph bolts back to Google.
The AI industry is accelerating the transformation of LLMs into Agents capable of executing complex tasks. Tencent's open-sourcing of a 770B-parameter model, paired with dedicated scientific evaluation benchmarks, and Amazon's construction of agentic creative workflows through MCP together demonstrate that "getting things done" has displaced "parameter size" as the central competitive dimension. Notably, this leap in capability stands in stark contrast to the lawsuit alleging xAI's use of illicit content for model training—the deeper technology penetrates production workflows, the more non-negotiable the ethical baseline for training data becomes.
Today's Top 3 Headlines
- AI
🤖 Tencent Hy Team Open-Sources 770B-A49B MoE Weights & FP8 Quantized Version
Tencent's Hy Team open-sources Hy4-preview 770B-A49B MoE weights on Hugging Face with FP8 quantization. Developers gain direct access to 100B+ parameter MoE models for fine-tuning and deployment, slashing LLM adoption barriers.
Source ↗ - AI Industry News
🤖 First lawsuit accuses Musk's xAI of training Grok on child sexual abuse material
Jane Doe files class-action lawsuit alleging Elon Musk's xAI trained Grok on child sexual abuse material, after Canadian Centre for Child Protection found AI-generated CSAM on the platform. AI industry faces unprecedented training data compliance challenges; companies must rebuild illegal content filtering and review mechanisms.
Source ↗ - AI Industry News
🤖 Terminal-Bench-Science: New Benchmark for Evaluating AI Agents in Scientific Research
Terminal-Bench-Science benchmark drops, testing Agent performance in real scientific workflows. For devs, evals shift from toy tasks to real-world challenges; HN discussions show Claude 3 Opus is more stable but still carries hallucination risks.
Source ↗
+6 more headlines
- 🤖 Amazon Quick & fal Build Agent Creative Workflows with MCP
- 🤖 Barrett Zoph leaves OpenAI after just 5 months, joins Google as VP of Research
- 🤖 Amazon Bedrock India Launches OpenAI GPT-5.6 Terra/Luna with Local Inference
- 🤖 AI reads mammograms to predict heart disease in women: 86% stroke detection accuracy
- 🤖 Google AI Unveils Gemini 3.5 Transcribe: 85+ Languages, 2.6% WER
- 🤖 Court rules Trump admin blacklisting Anthropic unconstitutional
Podcast highlight
Right, there's no standard answer in AI—just endless trial and error, and iteration.