AI Daily · 2026-10-22

Today's standout releases push foundation models toward zero-shot and domain-specific deployment: NVIDIA's Kumo Relational predicts directly on relational databases without task training, claiming to beat supervised baselines, while Google's AI-assisted ultrasound study lets novice operators estimate gestational age with low-cost probes. On the agent infrastructure side, LangSmith LLM Gateway entered public beta with layered spend and rate limits, cross-model fallback, and PII redaction. IBM shipped Granite time-series ensemble and Granite Speech 5.0 updates, NVIDIA expanded Cosmos world-model tooling, Modal launched Dojo for LLM training, and VocabX showed vocabulary specialization can cut inference cost by 77%. Security and policy were also in focus: OpenAI detailed consistent safety standards and updated usage policies, while PHP rushed patches for two RCE flaws after a disclosure timeline leak.

North America · First-hand

OpenAI

⭐⭐ [Other] Broad access through consistent standards

OpenAI published a post describing its work to align safety policies across the services that provide access to its models, requiring hosting and reseller partners to maintain safeguards consistent with OpenAI's own, investigate potential violations, and coordinate on enforcement so restrictions stay effective across services. The company says it evaluates a model's capabilities and risks before deployment, trains it to follow safety requirements, and tests whether safeguards hold when people deliberately try to bypass them, then monitors for abuse after deployment. OpenAI is updating its Usage Policies effective November 20, 2026, to clarify rules for safe model development and sensitive uses, and to require services providing direct access to its models, such as model routers or AI gateways, to get its review and approval first. OpenAI also says it is working with customers on SafetyIDs, which let it distinguish individual end users in an investigation without learning their identities, and cites Private Safety Processing as an example of its work with customers and partners.

Google

⭐⭐⭐ [Research] Empowering frontline maternal care with AI-assisted ultrasound

Google Research and Google for Health describe a prospective multi-center study, published in JAMA Network Open, on AI-based gestational age assessment using 「blind sweep」 ultrasound, which lets frontline workers without specialized sonography training scan with low-cost portable probes. The post cites WHO estimates of more than 250,000 preventable pregnancy-related deaths each year and notes that two-thirds of people globally lack access to diagnostic imaging such as ultrasound. The AI system was originally trained on ultrasound data from North Carolina and Zambia using Sonosite, GE, and Butterfly Networks commercial probes; in this study it was minimally tuned on only one target population and then prospectively tested at Northwestern Medicine in Chicago and Jacaranda Health in Nairobi. The system ran live on mobile devices with novice operators and unfamiliar Clarius C3 handheld transabdominal probes. The authors describe the work as a step toward expanding maternal care globally.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

NVIDIA released Kumo Relational, a pre-trained relational foundation model that makes predictions directly on enterprise databases: given a predictive query it gathers up to 10,000 historical examples as context and predicts outcomes such as churn, demand or fraud in a single forward pass, with no task-specific training, no feature engineering and no table flattening. NVIDIA says it is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks: across 41 tasks on 15 databases it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines, and on the SAP SALT enterprise benchmark it surpasses AutoGluon ensembles by 8% in mean reciprocal rank. Specific numbers cited include nearly 4 AUROC points on average over gradient-boosted trees on expert-engineered features on RelBench V1 classification tasks, and on SAP SALT 7 MRR points over XGBoost on engineered features out of the box and 13 MRR points after fine-tuning. It scales to databases with 500B+ rows, is available through NVIDIA's open-source, GPU-native structured-data-models library or as an NVIDIA NIM, and is built to be called by agents via its Predictive Query Language (PQL).

⭐⭐ [Model Release] Ensemble forecasting with IBM Granite time-series models

IBM Research published a Granite Timeseries Ensemble recipe (granite-timeseries-ensemble-r1) on Hugging Face that combines four pre-trained Granite time-series models behind one extensible forecasting interface: Granite PatchTST-FM-r1 (patched Transformer), PatchTST-FM-r2 (patched Conformer), FlowState-r1.1 (state-space model) and TTM-r3 (Tiny Time Mixers, an MLP mixer). The ensemble requires no retraining on the user's data and no fitted ensemble weights; the default is a Probabilistic Ensemble using uniform probability pooling, with an adaptive Inter Quartile Range (IQR) variant available, and the same interface can be extended with compatible forecasters. Outputs include point forecasts plus quantile forecasts (0.1–0.9 by default, where 0.1 and 0.9 define a nominal 80% interval). The recipe repository stores the ensemble configuration rather than merged weights and downloads the four member checkpoints at runtime, so the first run takes longer; it requires granite-tsfm>=0.3.11, and both the recipe and its default members are permissively licensed.

⭐⭐ [Model Release] Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

IBM released three updates to its Granite Speech collection on Hugging Face: a new model, granite-speech-5.0-2b, plus revised versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. granite-speech-5.0-2b has the same parameter count as 4.1-2b and transcribes English, French, German, Spanish, Portuguese and Japanese, translating to and from English in those languages plus English to Italian and Mandarin; it was trained on 390,000 hours of public audio along with synthetic data for Japanese ASR, keyword biasing and translation, and adds language-ID prompting, keyword lists up to 2,000 terms, and on-request punctuation and capitalization. It ships under Apache 2.0 with support in Transformers, vLLM, llama.cpp and mlx-audio, and the post's chart places it #2 among top open-weight ASR models at 5.06% average English WER. The Turbo CTC update adds multilingual transcription across seven languages, native punctuation and capitalization, and robustness to reverberant audio, while 4.1-2b-plus improves speaker attribution, timestamps, keyword biasing and long-speech performance.

Product Update

⭐⭐⭐ [Product Update] LangSmith LLM Gateway: Runtime Controls for Agents

LangChain announced that LangSmith LLM Gateway is now in public beta. The gateway sits between agents and the models they call as a centralized governance layer, giving one entry point for both closed and open weight models and avoiding provider lock-in. Runtime controls include time-bound spend caps and rate limits at the organization, workspace, API key, and user levels (a capped agent receives a 402 response), usage tracking, and model fallback rules across models, providers, and hosts. For multi-tenant and reseller setups, a custom request header identifies each end customer so separate spend caps and rate limits can be enforced while all calls route through a single API key. It also detects, redacts, and replaces PII and secrets in requests before they reach model providers.

⭐⭐ [Product Update] NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models

NVIDIA has expanded its Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints, aimed at helping developers build world models faster so physical AI systems can understand and simulate the real world. NVIDIA says the Cosmos platform ranks in the top five on six different open model leaderboards. The post positions the platform as a route for world-model builders to frontier open capabilities.

⭐⭐ [Product Update] Get started with LLM training on Modal

Modal announced Modal Dojo, a training library it says provides the building blocks, tuned defaults, and observability needed for production-grade LLM training, available now. The library builds on open-source frameworks Miles and Slime, and Modal says it has patched and upstreamed some PRs itself; users can tune recipes, build complex environments, patch the underlying frameworks, and reshape the built-in observability dashboard. The example shown starts training in under 100 lines: pick a model such as Qwen3 8B/27B, load a dataset from Hugging Face, define a custom reward function, set a recipe, and call launch(), demonstrated on GSM8K. Modal says the library has been validated on models from half a billion to multi-trillion parameters with support for bringing your own, and that auto-optimized recipes are provided for all supported models. The post argues that teams with an agent harness in production plus evals and traces collected from it are most of the way to training on their own data, and that owning the full stack is how companies compete on task-intelligence per dollar.

Research

⭐⭐ [Research] How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications

The article previews the forthcoming VocabX paper, which specializes a small language model's vocabulary around the cost of each token rather than token frequency. The authors tested it on a production Portuguese essay grader built on Qwen3-4B; the original general Qwen3-4B vocabulary had 151,669 entries, causing domain-frequent text to be split into many tokens and requiring scores over the full vocabulary at every generated token. VocabX rebuilt the vocabulary while keeping the same base model, data, and training recipe, and measured a 77% reduction in inference cost. The article also explains that inference has prefill and decode phases, with token-by-token generation making writing more expensive than reading. The paper is still under review, so the reported numbers may change.

Other

⭐⭐ [Other] Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium.

This is Part 1 of a series on profiling TorchNeuron, describing how to systematically analyze inference performance on AWS Trainium and identify where time is spent, starting from scratch with no assumed profiling experience. It explains that TorchNeuron is Trainium's open-source native PyTorch backend, supporting eager execution, torch.compile and standard PyTorch APIs, and integrating with NKI for custom operators. The walkthrough configures torch.profiler via NeuronConfig (data streams, target NeuronCores, output directory, and correlate_device_time to surface device time in the key_averages() table) and uses NeuronProfiler's export_trace callback to write trace.json. It then covers reading key_averages() to compare Neuron device time with Self CPU time per operation, and using the browser-based Neuron Explorer to inspect the system trace, device timeline and per-engine breakdown, with paired .neff and .ntff files produced when device data is captured. The article begins its analysis with RMSNorm.

⭐⭐ [Other] Today's PHP Security Release, and How We're Strengthening PHP Security

PHP has shipped security updates across the entire 8.x line fixing two remote code execution holes, and exploits have been circulating since the previous night after a leak in the coordinated disclosure process, making an upgrade urgent for anyone running PHP. The PHP Foundation explained what went wrong: around 15:00 UTC on October 20 the two CVE IDs began circulating on forums and social media with a wrong release time and a wrong CVSS score of 10.0 (the actual score is 9.2), though no technical details leaked; the project moved its pre-announcement up to 19:30 UTC and shifted the release from October 22 at 14:00 UTC to October 21 at 14:00 UTC. After the leak and pre-announcement, around 20:00 UTC, Calif, one of the security firms that reported the issues, mistakenly published a video showing its proof of concept working against several targets and revealing a few limited additional details; copies of the deleted video spread on social media and accelerated development of various PoCs, increasing circulation on October 21. The team that found the vulnerabilities published a timeline: on August 19, 2026 Wiz Research discovered the first PHP vulnerability using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol; on August 20 unauthenticated RCE was confirmed against Drupal and on August 21 against WordPress; on August 24 the vulnerabilities were reported to the PHP security team; on September 21 PHP acknowledged the reports and began developing a coordinated fix; and on October 21 patched releases became available and public disclosure followed.