#Data/Training

AIAnthropic says Claude Science helped produce the first complete all-sky ultraviolet map by merging data from several space telescope missions. About one third of the map was estimated by the model in regions lacking ultraviolet observations, and the map labels each pixel as measured or predicted with an uncertainty estimate.

AIStandard Bots, which calls itself America's largest AI-native industrial robot manufacturer, runs a shared base model that customers adapt through demonstrations and fine-tuning, with its largest model in the low billions of parameters. CEO Evan Beard and Head of AI Leif Jentoft say the company controls the arm, end effector, control system and AI, and runs inference on-premises on edge GPUs because most factories lack reliable internet.

AILi Auto says all new models launched from 2026 will use its self-developed batteries, which are already in the new L8, L6 and i8 and will reach the new MEGA, i9 and 2026 i6 in the fourth quarter. Li Auto president Ma Donghui says the move aims to connect data across cell production, pack assembly, vehicle integration and after-sales service. The company says it has about 300 battery R&D staff and nearly 4,000 patent applications and granted patents.

AIXiaomi's MiMo-V2.6 technical report argues that scaling reinforcement learning, not more pretraining data, is the main lever for frontier capability, along batch size, environment diversity, and grader strength. The article summarizes the report's methods, including groupwise agentic grading, a frozen-router fix for expert load collapse, and reward hacking defenses. It reports RL post-training costs of $2.6 million for MiMo-V2.6-Pro and $0.9 million for MiMo-V2.6-Flash.

Why it matters: The piece walks through the report's three-way RL scaling method, batch size, environments, and grading, with concrete failure modes and stabilization fixes useful to agentic RL practitioners.

AIMicrosoft releases Microsoft-Decision-1, a decision-scoring model post-trained from Alibaba's Qwen3.5-9B that returns calibrated probabilities for fixed answer options instead of generated text. The model is available through Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with free output, and has a 32,768-token context window. Microsoft reports 83.5% average accuracy across 36 benchmarks and 85 ms p50 latency, though the results are vendor-run and the latency comparison with competitors is disputed.

AIA paper decoded 315,320 reasoning blocks scraped from public repositories and recovered 367 PII artifacts and 182 credentials. The authors say reasoning traces can reveal hazardous information even when the model's visible output refuses a malicious request. They also warn that attackers could hide prompt injections in encrypted blocks to poison public agentic rollouts.

AIBoston Dynamics says its new humanoid hand has four fingers, 13 degrees of freedom, and direct actuation, and it is designed for mass manufacturing. The company dropped the pinky finger and reduced the gripper size, and built the hand for high-fidelity simulation to enable sim-to-real reinforcement learning. Alberto Rodriguez, director of robot behavior for Atlas, says the earlier hands could already lift more than 100 lb., so the new design focuses on reliability and tool use.

AIMistral AI has released LIDstral-Arabic, a fast classifier that identifies Modern Standard Arabic, Arabic dialects, and non-Arabic languages written in Arabic script across 51 classes. On Moroccan Darija, it scores 88.67% F1, versus 72.65% for LahjatBERT ALDi CL and 71.28% for GlotLID v3, across 84,870 evaluation examples. The model runs on CPU and is available from a private Hugging Face repository under Apache 2.0.

AIAi2's AI Infrastructure team replaced its priority-based scheduler for GPU clusters with GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The team says the change moved debates over how much GPU time each research project deserves from case-by-case operational decisions into a transparent budgeting process. The clusters range from 88 to 1024 GPUs across NVIDIA H100, B200, and B300 hardware, and serve about 150 internal researchers.

AIAi2's AI Infrastructure team replaced its priority-based GPU cluster scheduler with a system using GPU time budgets, hierarchical fair-share allocation, and a time-slicing contract. The team says the change turns decisions about how much GPU time each research project receives into a transparent administrative budgeting process. Its clusters, which range from 88 to 1024 GPUs including H100, B200, and B300 units, serve about 150 researchers facing demand two to three times available capacity.

AIBaseten explains how to decide which layers of a model can run in 4-bit NVFP4 without losing needed information. The post compares architecture-based heuristics, isolated-layer sensitivity scoring, and SaturationQuant, which accounts for other quantized layers. It also covers calibration with representative data and block-level scales of 16 values.

Why it matters: The post explains how to choose which layers run at NVFP4 precision using heuristics, sensitivity scoring, and saturation-aware scoring, with clear calibration steps.

AIPerplexity says its Computer tool created Alexandria, a free, source-backed encyclopedia with 66,574 entries, for $25,000 in credits (2.5 million). Computer scanned and synthesized 348,980 sources to build it, and will maintain the site at alexandria.pplx.app.

AICredo's latest full-DSP module, Cardinal 802, uses a 4×200G design aimed at both 800G and 1.6T, after the company expanded its ZeroFlap optical module line from 800G to 1.6T over the past year. The company also added Kfir200 silicon photonics PIC from its DustPhotonics acquisition and a PILOT diagnostics platform to the module.

AIAnthropic's Claude Science mapped the entire sky in ultraviolet light for the first time. AI agents downloaded data from multiple space missions, calibrated it, merged it, and used inpainting to fill gaps left by NASA's GALEX mission. In tests, predictions averaged about a ten percent deviation from actual measurements.

AIByteDance Seed researchers found that language models compressing their KV cache in fixed-size chunks retrieve the same information unevenly depending on token position. In a 128K-token needle-in-a-haystack test, base DeepSeek-V4 checkpoints differed by up to 40.2 percentage points by phase, and post-training narrowed but did not eliminate the gaps. The authors urge evaluating such models across positional phases, since high average accuracy can hide systematic failures.

AIHuawei's OceanStor M900 and Nvidia's CMX move reusable inference KV cache into a shared storage tier, but no agreed SSD specification exists for it. A storage executive said endurance requirements for one design rose from 3 to 9 drive writes per day and could change again. Industry sources expect convergence to take 6 to 12 months, with another year for development and validation.

AICaterpillar and CoreWeave are working to cut the time needed to train construction machines from months to hours, according to Caterpillar's Brandon Hootman. Caterpillar's ecosystem holds about 18 petabytes of federated machine, dealer and customer data, and a single machine can generate terabytes of LiDAR, camera and control data in a day. The partners use AI models, working with Nvidia, to annotate incoming field data so it can feed simulation and training within the same workday.

AIKore.ai launches Autoloop, an optimization engine that keeps adjusting the AI agents customers build on its Kore.ai Agent Platform to hit goals such as task completion, business-rule adherence and cost, including after deployment. Autoloop writes the first version of each agent from a company's operating procedures, and uses Kore.ai's StateTrace layer to flag where production agents miss goals. The engine is available now to all customers on the Artemis edition of the Kore.ai Agent Platform.

AIOpenAI has published 722 math papers generated by an unreleased AI model, posted to GitHub, spanning about 20 mathematical subfields. The article reports partial progress on the Riemann hypothesis, proving the quasi-Riemann hypothesis without a complete answer, and new results on matrix multiplication, integer multiplication, and Navier–Stokes equations. OpenAI plans to release Lean proofs for more of the papers and to fund review programs for AI-generated math.

AIDatabricks has released a public Beta of a Workday Data Connect federation connector for Unity Catalog, letting teams query Workday HR and finance data in place without copying it. Workday administrators share approved tables through Workday Data Cloud, and Databricks administrators create an OAuth connection and foreign catalog to govern access. The data is read-only, and Workday remains the system of record.

AIDatabricks has released Funke, a Python and PySpark library and deployable pipeline that parses HL7v2 healthcare messages into native Spark types while preserving the full message hierarchy. It succeeds Smolder, the Scala data source Databricks open-sourced in 2021, and ingests through Auto Loader into Unity Catalog bronze and silver tables. Users can query segments, fields, components, and subcomponents directly with DataFrame or Spark SQL expressions.

AIDatabricks introduces database branching in Lakebase Postgres, letting each coding agent work in its own isolated database branch created in under a second regardless of size. Branches use copy-on-write storage, consuming extra space only as they diverge, and scale to zero when idle so unused branches incur no compute cost. Schema changes are tracked in code and promoted to the parent branch through migrations rather than merged back, and ephemeral branches are created per pull request for testing.

AICoreWeave Forge is a platform that runs, observes, curates, improves, and evaluates AI agents and models in a repeating loop, according to CoreWeave vice president of product marketing Susanne Seitinger. The platform adds Agent Lens, Registry, RL Rollouts, model distillation, and the generally available AI Research and Iteration Agent, or ARIA. CoreWeave also introduced a partner network of tested, co-engineered integrations.

AIHospitals, academic centers, medtech firms, and pharma companies all face the same obstacle: imaging data is locked in clinical systems and hard to share. The EXAM study across 20 institutions showed federated learning, which shares model weights rather than patient data, improved AUC by 16% on average. Collaboration remains difficult due to scanner and protocol heterogeneity, privacy governance, and the lack of a common data substrate.

AIAn astrophysicist worked with Claude Science to create the first complete ultraviolet map of the sky, covering regions never observed in UV. Claude located existing datasets, combined them, and filled gaps with statistical inference, taking a few days rather than weeks of human work. The map is presented as a teaching tool and an example of low-priority scientific work that AI now makes feasible.

AIGoogle Research is presenting ContinuousBench, a standardized benchmark for measuring knowledge transfer and data contamination in differentially private (DP) synthetic data generation. Alex Bie is giving an encore presentation at the Google booth #107 at COLM 2026 today, at 2:00 PM PT.

AIThe Alzheimer's Translation Challenge, led by Primamente and AlzData with NVIDIA, Hugging Face, Nebius, Talisman Therapeutics, Ultima Genomics, Cellanome, Prime Intellect, and Boltz, is now open for registration. The competition starts in spring 2027, according to the post.

AIParticipants will use a dataset to build AI models, evaluated through a series of evals ranging from general benchmarks to more complex tasks. Top teams will be shortlisted and have their experimental hypotheses tested in Prima Mente's wet lab.

AIThe Alzheimer's Translation Challenge is built on a new atlas of 150M cells, covering neurons, astrocytes, and microglia across different genetic backgrounds under combinatorial perturbations with multi-modal readouts. The data will be made available through the AD workbench and Prima Mente's modeling platform.

AIThe NVIDIA KGMON team placed second in the KDD Cup 2026 Data Agents competition with a system built around a smaller, clearer, and easier-to-verify agent harness. The competition required agents to answer natural-language questions over heterogeneous sources, including databases, CSV and JSON files, prose documents, PDFs, and briefing videos.