AI News Brief: Morning roundup, 2026-11-17
Summary
The most consequential stories in this window concerned how AI is governed, deployed and paid for. Nvidia's board created a standing committee for national-security and export oversight during a leadership dispute, while Appian, Quorum Cyber and Sarvam all moved to make agentic and hosted AI easier to control and operate. A second thread is whether AI can be trusted and steered, from Meta's editable interest layer to research on verification and evaluation, and the window also carried a sharp warning about AI-generated child abuse material. Much of the rest is academic work, including a randomized trial showing that instructor-designed AI improved student exam results, which matters to anyone deciding how AI should enter their own organisation or classroom.
- UK compound semiconductor wafer supplier IQE raised its full-year 2026 outlook in a November 17 trading update, forecasting revenue growth above 40%. It attributed the gain to optical-communications demand from AI and data-centre infrastructure converting into sales, and expects adjusted EBITDA in the mid-teens of millions of pounds.
- An arXiv paper simulated 3,200 generative agents built from PIAAC adult-skills survey respondents to rate Canada's 2026 AI for All strategy. The agents backed nearly every policy but were far less sure they would personally benefit, and that gap widened among adults with less digital experience, which the authors call "conditional endorsement."
- Appian expanded its AI-agent control plane and unveiled Ask Appian at Appian Europe in London on November 17. The control plane manages Appian and third-party agents in one view, tracking activity, cost and permitted models, and Raiffeisen Bank International reported an 80% cut in token use and processing time in testing.
- Meta's CLEAR system, deployed on Instagram and Facebook for over a year, makes a ranked set of inferred user interests readable and editable in natural language. Users rated about 88.5% of displayed interests as accurate, and accepted additions and suppressions are stored as authoritative preferences.
- Sarvam launched Token Factory, serving its own and third-party open-weight models from GPU clusters in India through one OpenAI-compatible API. Sarvam says the service already handles more than 450 billion DeepSeek V4.1 Flash tokens a day.
- Researchers introduced RI-BNO, a Bayesian neural operator that learns its own inadequacy from PDE residual error estimates and uses half as many full training simulations as the vanilla model. In a thermal-break design example, online optimization time fell from 178 hours with a finite element model to 18 minutes.
- Quorum Cyber completed its acquisition of Ontinue, combining Ontinue's Agentic Security Operations Center with Quorum Cyber's resilience, threat intelligence and incident response services. Both companies focus on Microsoft-centred security, and the integration has now begun.
- DecayFormer is a compact hybrid state-space and attention model that forecasts battery state-of-health trajectories from early charge-discharge cycles. Using one fixed configuration across 22 aging-condition folds with no per-fold tuning, it reports mean test MAPE of 2.25% on Li-ion and 0.95% on Na-ion cells.
- Nvidia's board established the National Security and Export Control Oversight Committee to oversee national-security policy, export-control compliance and global trade risk at board level. Suzanne Nora Johnson chairs it, with Stephen C. Neal and John O. Dabiri as members.
- Agent-Guided Process Verification breaks a math solution into ordered key claims, checks them one by one and stops at the first error, then has a correction agent re-derive the rest. Across five datasets and four models it reached 63.8% average accuracy, versus 50.3% for the strongest baseline.
- The Theorem Evolution Graph maps how theorems modify, strengthen or extend earlier results, and uses it to guide LLM agents toward new conjectures from a seed paper's citation network. LLM judges and an expert mathematician preferred its conjectures over those from an unconstrained baseline, mainly for significance and ambition.
- This paper introduces assume/guarantee contracts for training heterogeneous multi-agent teams with reinforcement learning, so subtasks can be assigned to agents with different capabilities. The authors report 100% final evaluation success on three benchmarks, with comparable baselines needing roughly 7 to 11 times more training-loop time in CPU pilots.
- CONTEXT-PCa is a study protocol testing whether adding 43 clinical variables to a fixed image-only MRI model improves detection of clinically significant prostate cancer. The cohort includes 6,022 cases from University College London Hospitals, with five external cohorts for testing generalisability.
- An open letter with more than 165,000 signatures, including Elon Musk, asks Nvidia's board to bring back Jensen Huang while an independent investigation establishes the facts. The letter follows reports that the former chief executive was already plotting a return just four days after resigning.
- Partial-Order Likelihood Policy Optimization learns from Pareto dominance among responses that satisfy different sets of verifiers, rather than collapsing outcomes into a scalar reward. On code generation and instruction following, it improved learning efficiency and raised rates of complete requirement satisfaction.
- MarkTechPost's write-up of the same launch describes Sarvam Token Factory as open to everyone, serving eight open models, including Sarvam 105B, GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Gemma 4 31B, through one OpenAI-compatible API from Indian GPU clusters.
- A pre-registered randomized trial across 7,200 students at 19 universities found that AI tools designed by frontline instructors improved performance on proctored, in-person exams. Course engagement and satisfaction showed no change.
- According to TechCrunch, Apple has launched a new app for a study on how people use AI to answer everyday medical questions. The article is brief and gives few details beyond that.
- Liquid AI released d1.2 and d1.2-ultra, decision models that read text, images and audio and return calibrated probabilities without generating output tokens. The ultra version scores 75 on the Artificial Analysis Decision Index v2.1.
- A report from the Edinburgh-based Childlight Global Child Safety Institute found that reports of AI-related child sexual abuse material rose more than 8,400% in two years, from about 4,700 in 2023 to over 400,000 in 2025. The increase includes deepfakes made with "nudify" apps, and the report calls for governments to treat child sexual exploitation as a preventable public health emergency.
- A Latent Space roundup reported that Perplexity reached a $1 billion revenue run rate, up four times in 2026. The same newsletter noted that Computer gained Teammates, Perplexity shipped a Monitor API, and a new research lab launched.
- Datasette 1.0a43 adds streaming support for output renderers, including the built-in CSV and TSV exports, and fixes a security issue that could expose the title and SQL of private stored queries. It also adds CSS multi-page view transitions, so navigation inside the application feels more like a single-page app.
- This survey reviews how LLMs are used across the AI infrastructure lifecycle, from accelerator design and compilation to deployment, operations and energy management. Its cross-study synthesis finds consistent gains in some areas, but reported results for workload orchestration depend on whether they came from simulators or real clusters.
- ROBOTERA announced on November 17 that its humanoid robots now sort parcels at 11 China Post hubs across eight Guangdong cities. The project was implemented jointly with the China Post Group Guangdong Branch and extends humanoid logistics work across the province.
- This paper proposes a receiver-side framework that uses a masked language model and a channel-posterior estimate to reconstruct corrupted tokens in one inference pass, without retransmission or extra training. In noisy-channel experiments it improved cosine similarity by up to 67.2% over context-only reconstruction.
Summaries written with AI from the linked source articles.
Sources in this roundup
| cs.AI updates on arXiv.org | 12 article(s) | |
| Unite.AI | 6 article(s) | |
| AI (artificial intelligence) | The Guardian | 2 article(s) | |
| MarkTechPost | 2 article(s) | |
| AI News & Artificial Intelligence | TechCrunch | 1 article(s) | |
| Latent Space | 1 article(s) | |
| Simon Willison's Weblog | 1 article(s) |
Most-mentioned keywords
| models | 4 mention(s) | |
| agent | 3 mention(s) | |
| learning | 3 mention(s) | |
| open | 3 mention(s) | |
| sarvam | 3 mention(s) | |
| token | 3 mention(s) | |
| appian | 2 mention(s) | |
| context | 2 mention(s) |
Sources
- IQE Raises Outlook as AI Optical Opportunities Generate Revenue
- Conditional Endorsement: Simulating How Adults with Different Skills Evaluate Canada's National AI Strategy
- Appian Expands AI Agent Controls and Unveils Ask Appian
- CLEAR: Legible and Steerable Recommendation via Grounded Natural-Language Interests
- Sarvam Opens India-Hosted Token Factory for Open-Weight Models
- Learning inadequacy of Bayesian neural operators for PDE-constrained inference and design
- Quorum Cyber Closes Ontinue Acquisition to Expand Agentic AI Defense
- A Hybrid State Space-Attention Architecture for Forecasting the Trajectories of Battery Degradation
- NVIDIA Adds Standing Committee for Security and Trade Oversight
- Agent-Guided Process Verification for Mathematical Self-Correction
- Guiding AI Agents to Generate Conjectures by Encoding How Theorems Build on One Another
- Assume/Guarantee Contracts for the Compositional Multi-Agent Reinforcement Learning of Heterogeneous Teams
- CONTEXT-PCa: Evaluating the role of clinical context in MRI-based AI for detecting clinically significant prostate cancer (study protocol)
- Elon Musk joins 165,000 others calling for Nvidia to bring back Jensen Huang in open letter
- Policy Optimization With Multiple Verifiers via Pareto Rankings
- Sarvam Launches Sarvam Token Factory: Bringing India-Hosted Inference to 8 Open Models
- Instructor-designed AI improves learning in a randomized trial across 19 universities
- Apple launches new app to study how people use AI for health
- Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index
- Reports of AI-related child sexual abuse material rise 8,400% in two years, data shows
- [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026
- datasette 1.0a43
- Large Language Models for AI Infrastructure
- ROBOTERA Deploys Humanoid Robots at 11 Postal Hubs in Guangdong
- Dual-Stage Channel-Posterior Integration for MLM-Aided Token Communications
