Today in 12
11 sources · 6 first-party announcements · 1 new model · 8 briefs
01 Model releases and updates
02 Product releases and updates
03 Industry
04 Research
05 Tips and perspectives
06 Briefs
- How Hugging Face Trained Qwen3.8-27B Adapters for Three Agent HarnessesHugging Face BlogFollow-up
- ChatGPT Space Launches for Plus UsersGreg Brockman · X
- Huawei Cloud’s Ascend-Trained CloudRobo-WM1.2 Tops Two Robotics BenchmarksQbitAI
- AWS Launches SageMaker HyperPod Global Inference GatewayAWS Machine Learning Blog
- How AI Is Helping Scientists Find Leads in Neurodegenerative Disease ResearchOpenAI · X
- OpenAI Expands ChatGPT Agent Publishing Eligibility Beyond the U.S.OpenAI · X
- OpenAI Ties Reasoning Items to Conversation History in the Responses APIOpenAI Developers · XFollow-up
- Apple Introduces Depth Pro 2 for Sharp, Temporally Consistent Metric DepthApple Machine Learning Research
Product releases and updates

Gemini Adds GIF Creation with Gemini Omni 1.2 Flash
Google says Gemini can now create GIFs with its video model, Gemini Omni 1.2 Flash. The post gives looping photo effects and funny memes as examples, and says the model is designed to produce smooth motion and consistent quality across every frame.

AWS Adds Marketplace Purchases to Amazon Bedrock AgentCore Payments
AWS adds AWS Billing as a payment option in Amazon Bedrock AgentCore payments, allowing agents to buy per-request from x402-enabled AWS Marketplace sellers in USD and charge purchases to the developer’s AWS account. Developers can set session spending limits and approved sellers, while transactions appear in AWS CloudTrail and AWS Cost Explorer. The launch also includes an optional AWS Marketplace connector in AgentCore Gateway for discovering products and endpoints.

Amazon SageMaker Inference Reviews 2026 Launches in Part 2
Amazon SageMaker AI’s second 2026 launch review covers inference updates for managed endpoints and HyperPod, continuing from 13 launches in part 1. The post describes tools for benchmarking configurations, sharing GPU capacity, serving larger models across nodes, routing requests across clusters and Regions, and monitoring endpoint performance. It also presents a hypothetical support team using these capabilities across four product lines and two Regions.
Industry

How OpenAI Is Combating Adversarial Distillation
OpenAI details its defenses against attempts to extract protected reasoning and use it to train models without comparable safeguards. In an internal adaptive black-box evaluation aggregated across six attacker models, the average success rate for extracting reasoning from an earlier request fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. OpenAI also describes request-level protections, campaign investigations, and enforcement.

AI-Assisted and Human Proofs Claim to Solve the Riemann Hypothesis
Mathematician Levent Alpöge says he found a Lean-compilable solution to the Riemann Hypothesis after working with Claude. TechCrunch reports that mathematicians are examining the result and that Alpöge has committed to making the argument understandable. Separately, Stanford postdoc Alessandro Masullo published a human-written proof of the General Riemann Hypothesis after twelve years of work, introduced by Stanford professor Ravi Vakil.

Meta Reconsiders Its 2030 Net-Zero Goal as AI Emissions Rise
Meta is re-evaluating the timeline and interim milestones for its 2030 net-zero emissions goal, citing AI infrastructure demand that has grown faster than almost anyone projected. Its sustainability report says emissions rose from 15.6 million metric tonnes of carbon dioxide equivalent in 2024 to more than 21.7 million in 2025, with much of the increase attributed to capital goods such as data center servers.
Research
Tencent Hy Releases Hy4 Open Weights and System Card
Tencent Hy says it is releasing Hy4’s open weights alongside a System Card covering safety, security, and alignment evaluations, with closer attention to frontier risks in cyber and bio. The company says the evaluations include reviews by Tencent’s Xuanwu, Zhuque, and Keen Labs, as well as independent testing by FangcunLeap and alice_dot_io.
MiniMax Cyber Finds Practical Attacks on Cryptographic Schemes
MiniMax says its research team used MiniMax Cyber to turn theoretical cryptanalysis into practical attacks on NIST candidates, peer-reviewed schemes, and public cryptography challenges. Reported results include breaking full-round collision resistance in Waterfall-256 and Blender-256, recovering keys from several constructions, and solving 55 previously unsolved McEliece challenge instances. The work focuses on academic designs, past competition candidates, and public challenge parameters; MiniMax says it does not affect deployed cryptographic standards.
A Three-Month Study Tracked 936 AI Agents Producing 936 Papers
A report published by alphaXiv with MIT, Stanford, and other institutions tracked 936 AI agents across 35 labs from August to November as they produced 936 papers. In blind reviews of 202 AI and human papers, two Gemini models rated the AI papers 2 points higher for experimental rigor and 0.4 points lower for novelty on a 10-point scale. The authors caution that the overall AI score advantage came entirely from the earliest lab, and the public dataset excludes raw data and model weights.
Tips and perspectives
Google Duplex Adds Business Calling to Tasks in Gemini
Google launched Duplex in Tasks in the Gemini app this week, letting the personal AI collaborator call businesses and handle requests for users. The calling stack uses Gemini for multimodal understanding, planning, and dialogue, and accepts free-text task descriptions instead of relying on fixed business categories. Calls disclose that the caller is an automated system from Google, and businesses can opt out through Google Business Profile settings or during a call.

How the Claude Code Team Built AI-Assisted Code Review
Baoyu summarizes a presentation by Claude Code product manager Morgan Lunt on Anthropic’s internal code review system. Lunt says the share of pull requests with substantive issues found rose from 16% to 54% after automated review, while merged pull requests doubled and rollbacks fell by 40%. He recommends building an evaluation set from a company’s own incident-linked pull requests and using review rules to focus agents on important findings.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-11-20
