Easy AI Daily News Digest | January 9, 2026: OpenAI Healthcare, GLM-4.7, Qwen3-VL, vLLM Milestones

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily digest for January 9, 2026 covering policy, models, infrastructure, agents, research, and industry news. OpenAI launched ChatGPT Health and OpenAI for Healthcare (HIPAA-compliant, deployed at AdventHealth, UCSF, MSK, HCA). A Stanford paper claims frontier models memorize training data, with Claude 3.7 Sonnet reproducing ~95.8% of Harry Potter Book 1. Zhipu's open-source GLM-4.7 Reasoning scored 42 on Artificial Analysis (355B MoE, 32B active, 200K context, MIT license) as Z.ai pursued a HKEX listing. Alibaba released Qwen3-VL Embedding and Reranker; Baidu ERNIE-5.0 and Tencent Hunyuan-Video-1.5 entered LMArena rankings. AI21 open-sourced Jamba2; TII released Falcon-H1R-7B; Lightricks open-sourced LTX-2 audio-video generation. vLLM hit ~16K token/s on B200 and merged a KV offloading connector; AI-generated kernels entered vLLM with ~40% single-kernel speedups. Hugging Face shipped Transformers v5. WSJ reported Anthropic raising ~$10B at a possible $350B valuation, and NVIDIA skipped new GPU announcements at CES for the first time in five years.

Key points

Policy, Governance & Safety

OpenAI launches ChatGPT Health / OpenAI for Healthcare: HIPAA-compliant products for clinical Q&A, paperwork, and knowledge retrieval, live at AdventHealth, UCSF, MSK, HCA. Raises privacy and "AI replacing doctors" concerns. (announcement | ChatGPT Health)
Stanford paper on memorization: Multiple frontier models can leak copyrighted training content; Claude 3.7 Sonnet reportedly reproduced ~95.8% of *Harry Potter and the Philosopher's Stone*, while GPT-4.1 was much lower. (thread)

Models & Capabilities

Zhipu GLM-4.7 tops open leaderboard: Scored 42 on Artificial Analysis (up 10 from 4.6), strongest in coding, agents, and scientific reasoning. 355B MoE (32B active), 200K context, MIT license, ~710GB BF16 weights. Z.ai also announced HKEX listing plans. (benchmarks)
Alibaba Qwen3-VL Embedding & Reranker: Two-stage multimodal retrieval supporting text, images, screenshots, video, 30+ languages, adjustable dimensions; tops MMEB-V2/MMTEB benchmarks. Available on Hugging Face/ModelScope; vLLM nightly support. (intro)
Baidu ERNIE-5.0 & Tencent Hunyuan-Video-1.5 on LMArena: ERNIE-5.0-Preview-1220 reached #8 on Vision (1226 score, only Chinese lab in top 10); Hunyuan-Video-1.5 entered text-to-video (#18) and image-to-video (#20). (Vision | Video)
AI21 open-sources Jamba2: Enterprise-focused hybrid SSM+Transformer with KV-cache optimization, Apache 2.0 license.
TII Falcon-H1R-7B: Small hybrid Transformer+Mamba reasoning model; solid results on Humanity's Last Exam, τ²-Bench Telecom, IFBench.
Lightricks open-sources LTX-2: Production-oriented audio-video generation model runnable on consumer GPUs; weights, code, trainers, LoRA, and docs included; NSFW/copyright training restrictions. (model page)
Gemini 3 shines on PokerBench: 21,000 hands of poker; Gemini 3 Pro posted the highest overall profit, though Gemini Flash was stronger head-to-head. (site)

Infrastructure & Hardware

vLLM hits ~16K token/s on NVIDIA B200; merged KV Offloading Connector (with IBM Research) asynchronously offloads KV cache to CPU RAM—up to 9x throughput gains on H100, 2–22x TTFT reduction on cache hits. (milestone)
AI-generated kernel lands in vLLM: An LLM-generated fused RMSNorm kernel gives ~40% single-kernel speedup (~1.6% end-to-end). Author argues system-level benchmark suites are key for AI-written kernels. (deep dive)
CuteDSL Flex Attention ~30% faster on H100 forward pass; SM100 backward supported, SM90 in Flash-Attention PR #2137.
Hugging Face Transformers v5: Unified tokenizer backend, PyTorch-first refactoring, better serving/quantization/deployment; plus swift-huggingface and AnyLanguageModel for Apple ecosystems. (blog)
Epoch AI: Global AI compute exceeds ~15 million H100-equivalents; chip power alone surpasses 10GW.

Agents & Tooling

LangChain file-based agents & VS Code Agent Skills: Harrison Chase proposes describing agents via agents.md, subagents/, skills.md, mcp.json; VS Code ships "Agent Skills" based on Anthropic's open standard.
DSPy multi-turn rewrite planned: Conversation history currently stuffed into system prompts; maintainers will rework multi-turn handling and history serialization.
MCP community discusses standardized staging for side-effectful tools: Dry-run calls before execution for auditing; debate over SEP vs SDK best practices.
Claude Code power-user techniques: Hooks-based skill routing, error log systems, /commands as local mini-apps, Opus-forced subagents, disciplined context compaction.

Research & Methods

MAGMA: Agent memory via multiple graph structures (semantic, temporal, causal, entity) with policy-controlled retrieval; clear gains on LoCoMo and LongMemEval.
SPOT @ ICLR 2026: Workshop on scaling post-training (SFT/RLHF); submissions due Feb 5.
Artificial Analysis on evaluation: Real-task benchmarks (GDPval-AA) plus an Openness Index; discussion of evaluation fragility.
"Dead salmon" artifacts in LLM interpretability: Interpretation methods (probing, sparse autoencoders, causal analysis) produce plausible explanations on randomly initialized networks (arXiv:2512.18792).

Products & Industry

Gmail enters the Gemini era: AI conversation summaries, drafting/polishing, AI Inbox, natural-language search.
GLM-4.7 for coding: Reddit users report 85–90% usable code, stable long-file performance, ~1/5 the API cost of Claude Sonnet 4.5.
Qwen-Image locally: Run Qwen-Image-2512 and Edit-2511 at ~14GB VRAM via ComfyUI/quantized GGUF builds.
WSJ: Anthropic raising ~$10B at ~$350B valuation, reportedly one of the largest private AI funding rounds ever.
Google AI Studio sponsors TailwindCSS after open-source funding controversy.
Funding: Autonomous raises $15M (Garry Tan led); Protege AI raises $30M (a16z led).
NVIDIA skips new GeForce GPUs at CES for the first time in five years; RTX 50 Super rumors officially denied.

---

📌 Source: Easy AI Daily (zhichai.net) 🤖 Compiled by: AI assistant