Shopify Analytics introduces an annotation feature that allows users to add annotations with titles, descriptions, dates or date ranges, and types to time-series reports. These cover scenarios such as campaign launches, influencer mentions, supplier changes, promotions, or operational optimizations. Annotations are displayed alongside context provided by Shopify and installed apps, are shared with employees who have Reports permissions, and support searching, filtering, editing, and deleting.
Google DeepMind has released the Gemini 4 Argon model, targeting programming, enterprise knowledge work, and cyber defense. It supports up to 1M token outputs via the new API feature, Long Decode Continuation.
NVIDIA's technical blog published a tutorial on using the NeMo framework to fine-tune the Nemotron 3.5 ASR model for Saudi Arabic dialects (Najdi and Hijazi), including lightweight corpus filtering that retains 82.5% of the original SADA samples.
The Ethereum Foundation and the Open Anonymity Project jointly launched zkAPI, a privacy-focused API payment system based on zero-knowledge proofs. The system allows users to deposit ETH or USDC into an Ethereum contract and then authorize usage quotas through Groth16 proofs. Neither the server nor AI providers can link requests to the source of payment. The system is live on the Ethereum mainnet, offering a local client, browser SDK, and the OA Chat demo application.
DeskForge is a controllable desktop environment that generates large-scale supervised data by combining real applications, including 1.2M annotated desktop observations and 159.7M element instances. Fine-tuning four vision-language models based on DeskForge-1M improves accuracy by 11.51 and 10.11 percentage points on ScreenSpot-Pro and OSWorld-G respectively, and increases the number of completed long-horizon tasks on WebArena-Infinity and OpenApps. The code, dataset, and fine-tuned models have been open-sourced.
Shopify is introducing multiple fulfillment feature updates for the 2026 peak season: dynamic shipping rate calculation based on actual box dimensions, market-based regional shipping settings, mixed shipping and pickup within the same order, display of split shipment information on the checkout page (full rollout in July 2027), batch processing of up to 250 orders, scan-to-verify inspection, automated shipping label generation via Flow, direct purchase of UPS return labels from the admin dashboard for US merchants, and automatic upload of proof-of-delivery photos to the order page for supported carriers.
OpenRUA is a zero-abstraction-layer robot control interface that provides only ROS 2 terminal access to off-the-shelf coding agents (Claude Code + Claude Opus 5), without custom workflows or dedicated interfaces.
This article proposes a modular edge computing architecture based on Cloudflare Workers, using lightweight gateway Workers to orchestrate multiple single-responsibility functional Workers (such as image optimization, failover), achieving zero-network-overhead calls via service bindings; it compares with Akamai implementations, revealing how differences in platform primitives impact code structure, and provides invariant design principles for cross-CDN scenarios, rolling release strategies, and a three-layer testing system.
Physicist Matthew Schwartz published a long article in Anthropic Research, proposing the concept of 'Claude-shaped science': rather than forcing LLMs to mimic human scientists, it identifies and focuses on problems naturally suited to them—so-called 'Claude-shaped' problems—which are quantitative scientific tasks requiring broad knowledge, strong coding abilities, high-precision numerical computation, and cross-domain pattern recognition.
NVIDIA Dynamo-Triton (formerly Triton Inference Server) has released end-to-end inference support for the Hierarchical Sequential Transduction Unit (HSTU) Generative Recommender (GR).
Google DeepMind has released Gemini 4 Argon, a frontier large language model designed for complex workflows, which is now available to trusted cyber defenders through the Fairwind program; it supports up to 1M output tokens.
NVIDIA has released general versions of the cuObject client and server libraries and is driving an xio-sig standard extension to include cuObject. Additionally, it launched the SCADA Server SDK, enabling storage vendors to build accelerated storage servers that respond to GPU-side requests. IBM has already implemented a prototype integration of Storage Scale based on this SDK.
An NVIDIA technical blog post published a tutorial introducing how to use NeMo Relay to trace the execution process of Hermes Agent. It covers setting up an isolated runtime, running terminal tool tasks and multi-tool meeting research tasks, generating trajectories in three formats (ATOF/ATIF/OpenTelemetry), visualizing and analyzing them in Arize Phoenix, and demonstrating how to evaluate the effects of agent harness modifications based on 108 controlled experiments.
During his internship, Rohan Kannan from Microsoft Research developed an end-to-end machine learning system that uses L1 point solar wind observations, AE/Dst index forecasts, geological conductivity, and substation location data to predict geomagnetically induced current (GIC) risks for 66935 substations in the continental United States with a lead time of 30–60 minutes.
Google has released its new flagship model, Gemini 4 Argon, targeting software engineering, legal and financial enterprise knowledge work, and cybersecurity defense; the output limit has been increased to 1 million tokens; on DeepSWE v1.1.
A team from Google DeepMind published a paper in Nature proposing the SynthIDBio watermarking technology, which can embed invisible and hard-to-remove watermarks into AI-designed protein sequences while maintaining their structural and functional integrity. The technology is based on the ProteinMPNN tool, introducing key-driven amino acid preferences during the side-chain selection stage, causing the watermark to be randomly distributed across the full-length sequence and detected via statistical methods. Experiments show that watermarked proteins still bind target proteins normally.
AWS has published an official tutorial on natural language querying of claims documents using Amazon Bedrock Knowledge Bases. It utilizes the AgenticRetrieveStream API to implement multi-turn conversations, metadata filtering, context grounding verification, and answers with source citations. The tutorial includes full-process code examples covering S3 data preparation, knowledge base creation, streaming retrieval calls, citation rendering, permission control, and cleanup.
Bessemer Venture Partners and Bain & Company jointly released the "2026 Healthcare AI ROI Scorecard." Based on a survey of 226 healthcare enterprise executives covering 65 use cases, the report indicates that returns on AI investment in healthcare have been realized ahead of expectations: administrative use cases (such as revenue cycle management) achieve an average ROI of 4.0x, with 67% having deployed semi-autonomous or fully autonomous agents, whereas clinical use cases yield only a 2.9x ROI, with just 4% adopting autonomous agents. The primary drivers of this value divergence are lack of trust, unclear liability attribution, reimbursement mechanisms, and regulatory pathways, rather than insufficient model capabilities. The report also introduces an ROI scorecard that positions the maturity of each use case using two axes: adoption (horizontal) and realized ROI (vertical), noting that healthcare AI has shifted from the proof-of-concept stage to large-scale deployment.
Amazon Bedrock AgentCore introduces a new compute option called Runtime Instances, supporting persistence, multi-day execution, GPU acceleration, and co-location of multiple agents within a single instance.
Google DeepMind released SynthID Bio, a watermarking technology designed specifically for synthetic biology. It embeds imperceptible signatures into amino acid sequences and AlphaFold 3-predicted 3D coordinates, enabling verifiable watermarks at the physical protein level while preserving the binding affinity of VEGF-A, SARS-CoV-2 RBD, and PD-L1 proteins as well as natural sequence diversity. The methodology paper, code, in vitro experimental data, and model weights have been open-sourced.
NVIDIA and CoreWeave announced at the CoreWeave Fully Connected event that the Vera Rubin NVL72 system and Vera CPU are now available on CoreWeave Cloud.
Cohere has introduced Rubric-Calibrated Preferences nDCG@10 (RCP-nDCG@10), a new retrieval evaluation method based on calibrated AI judges and explicit relevance rubrics. It aims to mitigate the problem of missing relevant documents due to traditional nDCG's reliance on limited human annotations (qrels). In a blind test involving 46 participants, this method outperformed traditional nDCG with a 77% win rate and has been used to optimize Cohere's fifth-generation Embed and Rerank models.
DeepSeek announced that its open-source operator development tool TileLang natively supports the Huawei Ascend 950 NPU, and simultaneously released DeepGEMM, DeepEP, FlashMLA, and TileKernels.