AI News Brief Roundup: 2026-11-13 Afternoon

AI News Brief: Afternoon roundup, 2026-11-13

Summary

This window was dominated by the agent ecosystem moving from research into products. ChatGPT Wallet and Google's Gemini Decisions API both arrived alongside tooling from LangChain, Databricks, and AWS, all aimed at letting companies build, run, and govern agents. Several papers pushed the same theme from the other side, asking how reliable these systems are when they revise decisions, when they get safety budgets, and when they must remember long histories. Nvidia's abrupt CEO departure was the biggest non-technical story and a reminder that the industry's leadership is also in flux. Together, the items show where AI tools are heading and which questions of trust and oversight still lack clear answers.

1. Researchers describe a hybrid CNN-Transformer that detects breathing phases from audio recorded on commodity smartphones, reporting 76.94% frame-level accuracy on an unseen test set. An ONNX pipeline runs on-device with latency under 20 ms, so the approach avoids contact sensors and cloud processing.
2. This essay argues that AI's fragility, including hallucinations, bias, and weak explainability, comes less from how much data a system processes than from whether it understands what that data means. It frames data meaning as the basis for trustworthy, explainable AI.
3. Google Labs began testing Tabletop, an experiment in which group members share an online space and ask an AI agent to build the planning tools they need. Roman Nurik, design and engineering lead at Google Labs, used coordinating a family trip or potluck as the example.
4. InertiaMPDD is a research preprint for multimodal depression screening that uses speech, visual, and language signals, modeling how emotional states persist over time. Reported results reach 98.41% F1 on D-Vlog, though the standalone descriptors discriminate weakly on AVEC2014.
5. Perplexity released Decider v1.3 and Flash v1.3, open-weight models that return probabilities instead of text. Version 1.3 adds Locate for span extraction and reports 91.6% mean accuracy across 63 query-classification tasks.
6. This controlled audit finds that LLMs revising answers after disagreement often abandon correct answers, and no tested model met the selectivity needed to improve accuracy under expert disagreement. Accuracy fell by up to 16.1 points across 1,989 decisions in medicine, science, and law.
7. Databricks published an enterprise guide describing three ways to use OpenAI tools on governed company data: ChatGPT analytics through Genie, centrally controlled Codex deployments, and custom agents hosted as Databricks Apps. It builds on a partnership first announced in September 2025.
8. Natural PC adds a lateral circuit to predictive coding networks so they can perform local natural gradient learning. The authors report matching backpropagation trained with AdamW on ImageNet reconstruction tasks at roughly an order of magnitude lower compute and two orders of magnitude fewer samples.
9. LangChain raised $300 million in a Series C round at a $3 billion valuation, with existing investor CapitalG leading. The funding supports LangSmith, a platform for operating, improving, and governing agents regardless of the framework used to build them.
10. This paper treats safety in vision-language models as a constraint with explicit budgets for harmful assistance and over-refusal, rather than a tuned reward weight. On a Qwen3-VL-2B policy, held-out harm fell from 8.9% under standard GRPO to 4.8%, with MMBench unchanged.
11. AstroMemBench is a benchmark of 2,240 instances across nine spacecraft-operation tasks, testing whether embodied agents can use operational history. Results show Mem0 outperforms RAG by 12.25 points after a session restart, while memory systems show no clear advantage over RAG within a session.
12. OpenAI's October 2026 recap covers Amazon Bedrock, Amazon Bedrock AgentCore, and Strands updates, including new models for different tasks, features to keep agents available and grounded, and clearer cost reporting for builders.
13. OpenAI introduced ChatGPT Wallet, which lets agents in ChatGPT complete purchases only with user approval and under existing card protections. The feature works in ChatGPT Work, Codex, and dots.
14. Simon Willison tested Google's Gemini Decision 1 on video questions and found its yes/no answers come back as calibrated probabilities, such as 0.92 for "robot" on a clip with a robot. He notes the pricing matches OpenAI's GPT-6 Luna decision model at $0.10 per million input tokens.
15. Nearly 70 percent of US government agencies already use generative AI, and public-sector unions are winning protections against AI surveillance and automated firings. The piece cites an Oakland strike vote that secured retraining instead of layoffs and California's ban on so-called bossware.
16. This study tests a sign language encoder trained only on American Sign Language, frozen and evaluated on six other sign languages with no target-language training. One-shot retrieval ranged from 36.4% to 80.7%, and the best configuration reached 68.0%.
17. TechCrunch reports that ChatGPT Wallet lets AI agents shop and check out on users' behalf, starting with a limited rollout to US Pro subscribers. The feature extends ChatGPT from recommendations into completing transactions.
18. This paper proposes a two-branch EEG decoder that combines a graph branch grouping 61 electrodes into anatomical regions with a convolutional branch for movement-versus-rest classification. It reaches a mean F1 of 0.74 across held-out subjects and highlights the left parietal region (P1 and P3) as most connected.
19. Tools such as LangSmith, Langfuse, Opik, and Braintrust added gateways in front of model providers in 2026. The article explains the motive: a tracing tool can only record a model call, while a gateway can stop or control one.
20. Folio is a context manager for long-horizon agents that records, folds, and recovers history through linked views, rather than relying on model-written summaries. On 75 LOCA-bench tasks, it solved 66 against 49 to 59 for Claude Code, Codex, and SoL-Pi, at 2.4 times lower cost than its base runtime.
21. Nvidia CEO Jensen Huang's abrupt departure followed his admitted affair with a Chinese executive who was accused of being a spy, and reports of relationships with other employees. SiliconANGLE argues the episode raises questions for the wider AI industry.
22. Singularity-Bench asks whether agents can build the systems that train and run AI, using six end-to-end tasks scored against human baselines. Opus 5.5 with Claude Code matched or beat the human baseline on all six tasks, while the other agents averaged 0.58.
23. OpenAI introduced ChatGPT Wallet on November 13, 2026, letting users approve agent-assisted purchases without exposing card details to businesses or agents. Built with American Express, Mastercard, and Visa, it rolls out first to eligible US ChatGPT Pro subscribers through dots.
24. Google launched the Gemini Decisions API, powered by Gemini Decision 1, which returns calibrated probabilities for up to 1,000 yes/no, choice, and rubric questions. It accepts text, images, audio, video, and PDFs at $0.10 per million input tokens.
25. AWS released aws-emulator, an official Rust-based emulator for running tests against a local clone of its services without a paid API key. Swami Sivasubramanian described its build in a guest post on Werner Vogels' blog, and it currently requires Docker on macOS and Windows.

Summaries written with AI from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Acoustic Respiratory Phase Detection on Mobile Devices
2. AI That Understands – Why Data Meaning Is the Key to Trustworthy and Explainable AI
3. Google tests Tabletop, letting groups build planning tools with an AI agent
4. InertiaMPDD: Multimodal Personalized Depression Detection via Emotional Inertia-Inspired Dynamics Modeling
5. Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate
6. When Expert Disagreement Hurts: Prestige and Selectivity in LLM Decision Revision
7. Databricks Details Three Enterprise Adoption Paths for OpenAI Tools
8. Natural Gradient Learning in Predictive Coding through Lateral Circuits
9. LangChain Secures $300M Series C at $3B Valuation for Agent Governance
10. Safety as a Budget: Calibrating the Control Channel of Constrained GRPO for Vision-Language Models
11. AstroMemBench: A Benchmark for Memory Systems in Embodied Spacecraft Agents
12. ICYMI: What landed for AI builders in October 2026
13. Introducing ChatGPT Wallet
14. Gemini Decisions API
15. Nearly 70 percent of US agencies use generative AI, and unions are finally catching up with guardrails
16. Studies on Sign Language Retrieval in Low-Resource Dictionary Settings
17. ChatGPT’s new Wallet lets AI agents shop and check out for you
18. A Dual-Branch Graph and Convolutional Network for Interpretable Movement EEG Classification
19. Why LLM Tracing Tools Are Now Building AI Gateways
20. Folio: Recoverable Multi-Resolution Context Folding for Long-Horizon Language Agents
21. Nvidia CEO Jensen Huang’s abrupt departure shakes AI industry’s foundations
22. Singularity-Bench: Can Agents Build Agents?
23. OpenAI Launches ChatGPT Wallet for Agent-Assisted Purchases
24. Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens
25. aws-emulator