Field Notes

OpenAI claim on a Millennium Problem, 2.4T Qwen on HyperPod, Suno v6 with labels, WeChat zero click

Big promises, big models, and a weakness that needs no tap.

The key stories

OpenAI claims to have solved one of the Millennium Prize Problems. Independent confirmation is pending. That would be a big mathematical step and shows the pace AI brings to proofs. Do not use the result until there are peer reviews.1

A guide shows how to roll out Qwen3.8-2.4T-A95B with vLLM on SageMaker HyperPod. It covers cluster setup and NVFP4. NVFP4 is a 4-bit float format for dense models. You get an OpenAI-compatible endpoint with reasoning, tool calls, and MTP Speculative Decoding. Speculative Decoding is a predict-and-verify technique for faster output. If you run extremely large models, you can use it to test a path.2

Suno launches v6, built with Warner, BMG, and Believe according to the company. The model was trained from scratch on new, licensed material, including label content, says Suno. This is the first version with official label support and a response to ongoing lawsuits. If you generate music, that lowers legal uncertainty. An independent audit of the training base is missing though.345

After yesterday’s GPT-6 Astra launch, OpenAI now provides the system card and alignment details. The company calls Astra the world’s most intelligent and most aligned available model. That is a vendor claim. If you must meet regulation, read the card and match it to your obligations.67

Calif Research shows WeWorm, a demo of a zero-click worm. An attack with no user interaction that spreads via WeChat calls on iOS and Android. The victim does not need to pick up, say the researchers. It is a demo, not a reported outbreak. Put mobile call channels as an attackable surface on your risk list.8

IBM releases Granite Time Series PatchTST-FM-r2 as a SOTA model for time series. The license is business-friendly. You can build forecasts in production without long license reviews.9

AWS is betting on Qualcomm for AI inference, while Qualcomm uses AWS Bedrock for chip design. It is a mutual deepening across several product lines. Plan for new options in AWS stacks that can shift cost and performance.10

Tools and releases

  • Hugging Face launches ML Intern, a chat that kicks off ML experiments and clicks together workflows in the browser.11
  • Apple explains in a document how Siri Recap, Live Rewind, Sound Recognition, and Music Recognition aim to preserve privacy despite always-on listening.12
  • Massachusetts tightens power and clean energy requirements for data centers as the third US state in short order.13
  • Microsoft agrees with the AFT on principles for safe and privacy-compliant AI in schools, after local bans.14
  • Apple says the hinge of the foldable iPhone Duo came from AI-assisted design and 3D printing.15

Research

  • Nvidia describes encode prefill decode disaggregation as an architecture that speeds up multimodal serving, and gives criteria for when the split pays off.16

Sources

  1. OpenAI’s sly mathematical breakthrough sends a chill through academia (theverge.com)
  2. Deploying Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod with vLLM (aws.amazon.com)
  3. Suno launches v6 music models built with Warner, BMG, and Believe (the-decoder.com)
  4. Suno releases its first AI music model made with record industry help (theverge.com)
  5. Suno replaces its AI models with a new one trained on licensed music as copyright suits pile up (techcrunch.com)
  6. GPT-6 Astra: The System Card, Alignment and What Comes Next (thezvi.substack.com)
  7. GPT-6 Astra: The next generation in intelligence for work (openai.com)
  8. Quoting Calif Research (simonwillison.net)
  9. IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license (huggingface.co)
  10. AWS is using Qualcomm for AI inference while Qualcomm uses AWS Bedrock to design the chips (the-decoder.com)
  11. Hugging Face's new ML Intern lets anyone run machine learning experiments through a simple chat (the-decoder.com)
  12. Read the Apple document explaining how new listening features still protect your privacy (theverge.com)
  13. Massachusetts hits data centers with new clean power rules (techcrunch.com)
  14. Microsoft has new AI privacy rules for schools (theverge.com)
  15. The hinge for Apple’s new foldable phone was built with AI (techcrunch.com)
  16. When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving (developer.nvidia.com)