Social Media Briefing

vLLM project announces day-zero verified support for Tencent Hunyuan Hy4-preview on NVIDIA GPUs, detailing architectural specifics including 770B total/49B active parameters, 256 routed experts plus one shared, 1M context with sparse attention to 2048 tokens, HPC-Ops kernels, and a 10B MTP speculative decoding layer.

vLLM announces day-0 support for Z.ai's newly open-weighted GLM-5.3 model: 744B total parameters with 40B active, 1M context length, 128K max output. Notes that Z.ai kept the GLM-5.2 base and scaled post-training, allowing vLLM to serve it on the existing GLM-5.2 inference path with unchanged parsers and FP8 KV cache.

Anthropic reports that Sonnet 5 was used to post-train an early checkpoint of Opus 4.8, reaching safety scores approaching those of production Opus 4.8.

Andrew Ng shares an AI Engineering Skills map covering how software engineering fundamentals are evolving in the era of agentic coding, framed as a learning roadmap for practitioners.

Anthropic reports that across 10 alignment failures Claude reliably improved safety scores without degrading capabilities, and methods generalized to models up to 4.7x larger.

Gary Marcus breaks down Nvidia's Q2 financials with detailed analysis: $96B revenue contrasted against $366B in future commitments (supply/capacity $279B, cloud $29B, leases $25B, equity $25B, capex $8B), drawing attention to the massive gap between current revenue and forward obligations.

Anthropic describes how Claude hill-climbed safety benchmarks for misalignments such as deception and sycophancy while preserving general capabilities, then validated the best methods on held-out benchmarks for generalization.

Detailed list of talks from the AI-Native Enterprises Track at AI Engineer World's Fair 2026, covering agentic SDLC at Uber, Figma MCP Server, multi-agent code review, LLM gateways, and AI-native organizations.