iCreatAI News
Back to main site
Skip to content
Only First party#Data / training
Nov 4Wed
  1. Hugging Face Blog58

    Palette: teaching a small model to build decks in your design language

    Palette fine-tunes a 20B open-weight model with LoRA to build slide decks in two roles—a designer that writes a structured JSON brief and a coder that writes pptxgenjs code per slide—trained on roughly 1,000 synthetic decks (~10,000 slides) generated by a large open teacher and filtered by a checker. On PresentBench it ranks first on Economics and Education, third on Talk, and second across the three-domain mean (59.7 vs. NotebookLM's 60.8). A check-and-repair harness that keeps a rewrite only when the checker confirms improvement adds 3.5 points on average across 131 decks.

  2. Google DeepMind73

    Agentic cheminformatics with Google Antigravity

    Google describes Cheminformatics Skills for Google Antigravity that pair a language model for workflow orchestration and scientific reasoning with deterministic RDKit-based tools and ChEMBL/PubChem connectors. In a lung cancer walkthrough, an agent retrieved thousands of compounds, ran matched molecular pair analysis, built potency models, and proposed rigidified macrocycles with synthesis routes in about an hour. The authors note that sparse data and assay noise limit predictive precision, so computational designs remain testable hypotheses requiring experimental validation.

    Why it matters: It offers a concrete, technically detailed example of agent-assisted drug discovery, including both demonstrated capabilities and stated limitations.

  3. NVIDIA Newsroom38

    AI Is Changing How Semiconductor Fabs Work — and Engineers Are Moving Faster Because of It

    NVIDIA describes an "AI-defined fab" in which accelerated computing, physics-based AI, AI agents and digital twins help semiconductor engineers shorten simulation cycles, predict process behavior and test changes virtually before physical trials. The article cites up to 10x acceleration for COMSOL multiphysics simulations with cuDSS, up to 17x for Siemens Calibre lithography workloads with cuLitho, and Lam Research predicting plasma behavior in milliseconds versus days for full simulation. It also covers Advantest using NeMo Switchyard agents to cut model-selection response times from hours to minutes and Applied Materials building fab digital twins on Omniverse libraries.

Nov 3Tue
  1. Apple Machine Learning Research30

    Attention to Mamba: A Recipe for Cross-Architecture Distillation

    The authors propose a two-stage cross-architecture distillation recipe: first distilling a Transformer into a linearized version of Attention using an adaptation of the kernel trick, then distilling that linearized version into an adapted Mamba model with no Attention block. The distilled Mamba model retains most of the original Pythia-1B Transformer's performance, reaching a perplexity of 14.01 versus the teacher's 13.86, with results largely carrying over to downstream tasks though some gaps appear on tasks relying more on in-context recall. They conduct ablations at 1B scale with 10B tokens, scaling analysis on model sizes and distillation tokens, and a sensitivity analysis on token allocation between stages.

  2. NVIDIA Technical Blog25

    Keep Distributed Training Running with Automated Workload Recovery

    As distributed AI training scales, failures across hardware and software dependencies increasingly disrupt training progress, making fast, reliable recovery essential. The material notes that a stalled training rank can trigger an NCCL timeout, a GPU node can stop responding to the scheduler, or a straggler can slow the entire job, and that each disruption is manageable on its own.

  3. Allen Institute for AI Blog34

    How Olmo 3 earned the highest ethical rating from Nextcloud

    Nextcloud awarded Olmo 3 its highest Ethical AI rating, green, which requires open-source training and inference software, a self-hostable trained model, and publicly available training data. Nextcloud bundles Olmo 3 with its llm2 local LLM backend, citing the 7B model's ability to run on consumer hardware without high-end GPUs and its fit for customers seeking independence from big tech. Nextcloud engineer Marcel Klehr said fully open models make AI transparent, explainable, auditable and controllable, and that the company is working on expanding its AI features.

  4. PyTorch Blog28

    Announcing the Special Interest Group (SIG) for Ray on Accelerators

    The Ray community announced a Special Interest Group for hardware accelerators, aiming to bring together hardware vendors, Ray contributors, and application developers around a vendor-neutral roadmap for accelerator support. Its initial roadmap covers development and testing infrastructure, hardware-aware scheduling, and networking/transport, including Ray Direct Transport for direct device-to-device data movement. Ray currently supports more than a dozen kinds of accelerators, from NVIDIA and AMD GPUs to TPUs, Trainium, and Gaudi.

  5. AWS Machine Learning Blog36

    Build efficient agents with knowledge distillation on Amazon SageMaker AI

    This post explains how to use sequence-level knowledge distillation (SeqKD) to fine-tune a smaller shopping agent on Amazon SageMaker AI, using a frontier teacher (GLM-5.3 on Amazon Bedrock) to generate multi-step tool-use trajectories that are filtered by ShoppingBench's programmatic scorer and used for serverless supervised fine-tuning of a Qwen3.5-4B student. The distilled student reached an average Absolute Success Rate of 0.541 on 324 held-out ShoppingBench problems, versus 0.297 for the base model and 0.540 for the greedy teacher, while IFEval general instruction-following dropped roughly 6 to 8 points. The post also covers per-step trajectory encoding, LoRA training configuration, a low-cost model-evaluation pre-check, and serving the base model and student adapter as two inference components on one real-time endpoint.

Nov 2Mon
  1. Apple Machine Learning Research48

    Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps

    The authors propose an online reinforcement learning framework that trains UI navigation agents directly on live iOS applications, splitting the agent into a Planning agent and a UI Grounding model and using an LLM-as-a-Judge for dense step-level plus terminal rewards, optimized with PPO and step-level advantage estimation. They introduce GiGPPO, a neighbor-conditioned value estimation method that improves critic stability using reward signals from visually similar states across concurrent trajectories, backed by infrastructure orchestrating hundreds of concurrent virtual iOS devices. On UINavBench, iOSWorld, and AndroidWorld, their 8B RL agents substantially outperform SFT and terminal-only RL baselines, achieve nearly a sevenfold improvement over the SFT baseline in out-of-distribution generalization to Android, and outperform Gemini-3.5-Flash while tying Gemini-3.1-Pro on real iOS applications.

  2. Sakana AI Blog47

    [Event Report] Sakana AI Symposium: Jürgen Schmidhuber on the Past and Future of AI, and Language Models, World Models, and Physical AI from Japan

    On October 26, 2026, Sakana AI held a symposium at Hitotsubashi Hall in Tokyo featuring a keynote by Chief Scientific Advisor Jürgen Schmidhuber and a conversation with CEO David Ha, drawing over 400 attendees from across Japan. Sakana AI announced it is moving into pre-training its own models, unveiling the "Sakana Sawara" (provisional name) language model series, with the small first model "Sawara Small" (provisional name) planned for release in spring 2027 if development proceeds smoothly. The report presents edited transcripts of the talks, including Takuya Akiba on LLM development at the RSI Lab, Robert Tjarko Lange on open-ended autonomous discovery systems, and Stefania Druga and Sebastian Risi on world models and Physical AI.

Nov 1Sun
  1. Databricks Blog50

    Lakebase branching for CI/CD: lessons from 100 PR databases

    Snorkel AI replaced ~20-minute Aurora database restores for each PR environment with Databricks Lakebase copy-on-write branches, provisioning a writable production-like database in under 60 seconds. A DMS-replicated main branch stays within about 2 minutes of production Aurora, and each PR branch is created via API, given a TTL, and deleted when the PR closes. The design targets about 100 concurrent PR branches on a ~2.5 TB dataset, with per-branch compute scaling to zero and service principals using short-lived OAuth tokens for CI.

Oct 31Sat
  1. Wiz Research78

    “Black October” for AI Developer Platforms

    At least nine AI developer platforms disclosed incidents during October, with Fireworks, OpenCode and Warp reporting indications of an “agent swarm” attack and OpenCode and CommandCode attributing theirs to attempted “distillation.” Wiz describes downstream abuse of enterprise credentials tied to one vendor breach, including two keys used for roughly 800k and 1.7 million Bedrock calls respectively, likely over $100k in usage at list price. Wiz also links the keys to resale operations at llm-routing.com and xingjiabiapi.org, and publishes IOCs including 23.106.46.239 and 38.246.114.118.

    Why it matters: Provides Wiz's technical account of LLMJacking follow-on activity and IOCs from the October platform breaches.

  2. AWS Machine Learning Blog36

    Managing Amazon Bedrock costs end to end: A playbook for FinOps and platform teams

    This AWS post lays out a five-stage, looped operating model for Amazon Bedrock cost management: attribute every dollar to an owner via IAM principal and session tags in CUR 2.0, make every request observable through model invocation logging and stamped request metadata, enforce budgets at one of three tiers (monitor-only, IAM deny on breach, or synchronous gateway blocking), optimize platform defaults such as intelligent prompt routing and prompt caching, and operate a scorecard with a review cadence. It notes that agentic workloads issue many invocations per task and scale token consumption with autonomy and context size rather than user interactions, and it addresses the shared-role attribution problem in LLM gateways. FinOps leads the attribution and operate stages, platform engineering leads observability, enforcement, and optimization, and the review is joint.

  3. Databricks Blog42

    Databricks and SAP unleash end-to-end data & AI capabilities to level up enterprise AI transformation

    At SAP TechEd, Databricks and SAP announced an expanded partnership bringing the full Databricks Data + AI Platform—including Genie, Lakeflow, Unity Gateway and Lakebase—into SAP Databricks within SAP Business Data Cloud, alongside SAP Application Data Foundation for cloud-native access to SAP application data without custom extraction pipelines. The companies cite customers such as Adidas, Zalando, Mondi, CAF and SAF-Holland, which uses SAP BDC and SAP Databricks to unify over 100 million records and generate analyses 5x faster. SAP Databricks and SAP BDC Connect for Databricks are available today.

  4. Baseten Blog57

    The fastest delta weight syncs in the world

    Baseten rebuilt its delta weight sync path so the GPU weights themselves serve as the parent checkpoint, eliminating the ~770 GB CPU copy and cutting a GLM-5.3 sync from 36 seconds with a 6-second pause to 15 seconds with a 0.8-second pause. The new design uses a native decoder that keeps only nonzero XOR bytes, O(1) Adler-32 checksum corrections, and a separate CUDA-free helper process, and it frees host memory for KV cache offloading, raising cache hit rate from 2% to 92% and output throughput per GPU by 3.2x at concurrency 48 on AgentX. It also enables Kimi K3 RL rollout on two GB300 nodes, syncing an 8.4 GB delta in 75 seconds with a 4.1-second pause.

  6. Factory Blog38

    Factory Acquires Typedef: Bringing Data Context to the Software Factory

    Factory announced that Typedef is joining the company to add a data context layer built from code, making relationships between data models, pipelines, and dashboards visible to Droids. The companies say this will give engineers clearer lineage and evidence as Droids plan migrations, debug pipelines, and refactor models. Typedef's founders, Yoni Michael and Kostas Pardalis, previously built Coolan (acquired by Salesforce in 2016) and Blendo, respectively.

Oct 30Fri
  1. Wiz Research22

    Key Takeaways: AI Threat Readiness Survey Report 2026

    A survey of 100 senior IT and security decision-makers found that 55% of organizations actively deploy autonomous agentic systems and 31% run more than 25 distinct AI systems in production, yet only 17% maintain a comprehensive AI asset inventory and 52% discovered unauthorized AI applications in their perimeters over the past year. Leaders cited governance and cross-functional ownership complexity (26%) as their top AI security bottleneck, while defensive spending remains fragmented, with 43% stretching legacy cloud tools, 38% funding dedicated AI security budgets, and 37% relying on homegrown scripts. The report outlines a five-priority AI Threat Readiness Framework spanning production-grade controls for AI stacks, graph-based asset context, behavioral least privilege, AI-driven exposure scanning, and code-to-cloud response.

  2. Microsoft Research40

    From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence

    Microsoft describes its Industry Evaluation Data Generation Platform, which builds fully synthetic enterprise environments grounded in public sources such as O*NET occupational data, job postings and SEC filings to evaluate Microsoft Copilot without using customer data. The platform combines synthetic tenants across pharmaceutical, capital markets and IT/software industries with an automated six-stage query generation pipeline and quality frameworks covering structural fidelity, contextual relevance, behavioral realism and content safety. Microsoft reports a 39% improvement on its composite synthetic data quality metric versus its initial approach, some industry environments scoring above 0.9 out of 1.0, and estimated savings of about $100,000 per evaluation dataset and $300,000 in total project savings.

  3. Prime Intellect Blog34

    ResearchOCT 29TH, 2026Prime Mega MoE: Fast Fused MoE Kernels for Training

    Prime Mega MoE extends DeepSeek's Mega MoE kernel, which previously had only a forward implementation, with a fast overlapping backward pass so it can be used for training. It reports up to 3.11× faster MoE kernels than the default PyTorch-based implementation and up to 1.34× faster end-to-end training than the Prime-RL baseline (34% higher throughput). Supported data types are bfloat16, MXFP8 (Blackwell only) and blockwise FP8 (Hopper only).

  4. Sakana AI Blog40

    Intelligence Report: Analyzing the US Midterm Elections Through SNS AI Analysis

    Sakana AI, which works on defense and intelligence alongside finance, describes an analysis of about 710,000 reactions to 2,042 X posts from both candidates and both parties' state organizations in the Texas Senate race, using its narrative analysis system and its Sakana Marlin product to cross-check X observations against polling, campaign finance, prediction markets and news. It says much of the "support" for Paxton before the Republican runoff was intra-party mobilization and fell from about 50% to under 20% after the runoff, and that shifts in evaluations of the candidates appeared about two months before major raters such as the Cook Political Report changed their judgments. The report argues such findings require human-AI collaboration, with AI deriving hypotheses from data at a scale humans cannot read and humans examining them and making final judgments.

  5. Amazon Science38

    Speeding up language models for travel applications

    Researchers combined knowledge distillation, which trains a compact student model on a large teacher's outputs, with Medusa, a framework that predicts multiple tokens in parallel, and applied the pipeline with Booking.com to three travel features. The final model matched the quality of a 70-billion-parameter teacher while running 10 to 20 times as fast, with reported gains including 2.9% higher engagement versus GPT-3.5 for AI Trip Planner and 15-times-faster Smart Filters responses for 99% of queries. The work received the Industry Track best-paper award at ACL.

  6. AWS Machine Learning Blog36

    Lend and borrow resources across teams with AWS Batch’s quota management for SageMaker Training jobs

    AWS Batch now offers quota management, letting administrators divide compute capacity into quota shares per team or project with RESERVE, LEND, or LEND_AND_BORROW sharing strategies, plus cross-share and optional in-share preemption. The post walks through configuring a Scheduling Policy, Service Environment, and job queue, then submitting SageMaker Training jobs to quota shares via the SageMaker Python SDK v3 aws_batch module.

Oct 29Thu
  1. Hugging Face Blog44

    Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready

    Project ProtoStar is a new initiative bringing together universities, industry, and the open science community to release publicly funded scientific datasets, models, and code in AI-ready, machine-actionable formats on open infrastructure. Founding partners include Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University, and The George Washington University through the Open Forum for AI, the Astera Institute, Hugging Face as infrastructure partner, and ORCA as community partner. The first open challenge, hosted by Carnegie Mellon University from November 2–13, 2026, asks participants to use historical weather data from 210 U.S. cities through 2018 to predict 2019–2026 climate conditions and propose a climate adaptation policy idea.

  2. Databricks Blog28

    How Procore is powering construction analytics with OpenSharing

    Procore rebuilt its Procore Analytics product on a Databricks medallion architecture and OpenSharing, replacing a SQL Server backend that capped refreshes at one to three per day and relied on costly, brittle recursive API calls. The new design shares hundreds of curated, Unity Catalog-governed tables per customer at schema, view, table, and partition level, with Change Data Feed for incremental syncs and automated per-user, per-project entitlements. Procore says the catalog expanded to hundreds of tables in less than half the time the first generation took, and customers can now refresh on their own cadence.

  3. Databricks Blog28

    How Procore is powering construction analytics with OpenSharing

    Procore rebuilt its Procore Analytics product on a Databricks medallion architecture and OpenSharing, replacing a SQL Server backend that capped refreshes at one to three per day and relied on costly, brittle recursive API calls. The new design shares hundreds of curated, Unity Catalog-governed tables per customer at schema, view, table, and partition level, with Change Data Feed for incremental syncs and automated per-user, per-project entitlements. Procore says the catalog expanded to hundreds of tables in less than half the time the first generation took, and customers can now refresh on their own cadence.

  4. AWS Machine Learning Blog36

    Distributed training with Ray Train DLCs and EFA on EKS

    This AWS post shows how to fine-tune Qwen2.5-1.5B with PyTorch FSDP across eight NVIDIA L4 GPUs on two EFA-enabled g6.12xlarge instances on Amazon EKS, using the published Ray Train DLC (public.ecr.aws/deep-learning-containers/ray:train-ml-cuda-v1.1) without building any image. KubeRay creates the Ray head and worker pods from one RayCluster manifest, while hardware, transport and per-job Python packages are configured from outside the image via resource requests, environment variables and Ray runtime_env. The sample repository includes idempotent deploy scripts that locate g6.12xlarge capacity, provision the cluster and node group, install the GPU and EFA device plugins and KubeRay, and submit the job from inside the head pod, with the script failing if NCCL did not use EFA.

  5. Allen Institute for AI Blog38

    How researchers are adapting olmOCR 2 for historical and archaeological archives

    Ai2 describes how researchers adapted its open olmOCR 2 model: historian Maurice Casey built ArtificeOCR to run olmOCR 2 locally and make his Tropy photo library of archival sources searchable, while TU Wien and Austrian Academy of Sciences researchers fine-tuned olmOCR 2 with LoRA on 57 annotated Carnuntum pottery records, lowering transcription error on 450 test records from 15.58% to 1.11% through their CENTURIA benchmark.

Oct 28Wed
  1. Artificial Analysis Articles60

    Announcing Artificial Analysis Intelligence Index v5

    Artificial Analysis released Intelligence Index v5, its biggest upgrade since v4, adding FrontierCode 2.0 and Terminal-Bench-Science and upgrading HLE-Diamond, AA-Omniscience v1.1 and CritPt v2, with 60% of the Index now based on private test sets or held-out answers. Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55), while the open weights gap widens from 3 points to 12. It also launched the Progress Index v1, which tracks model capability over time on one continuous scale.

    Why it matters: It details a major benchmark revision and new leaderboard results, useful for readers tracking how frontier and open-weight models are measured.

  2. Claude Blog38

    Manage costs and demonstrate ROI for Claude Code on Google Cloud

    This post outlines a control, track, and prove framework for managing Claude Code spending on Google Cloud, using native tools such as per-project billing, budget alerts, spend caps, BigQuery exports, quotas, and the self-hosted Claude apps gateway for per-user and per-group spend limits and model allowlists. It describes tracking via Claude Code's OpenTelemetry metrics and events exported to Cloud Monitoring or BigQuery, and proving ROI through usage, satisfaction, and impact metrics such as PR cycle time and review turnaround. It cites Delivery Hero, Shopify, and Fifth Dimension as examples of measured impact.

  3. Databricks Blog33

    Unlocking Investment Insights Without Migrating Your Data: Lakehouse Federation

    Databricks Lakehouse Federation lets users create federated connections to external databases such as SQL Server, PostgreSQL, MySQL, and Snowflake and query them directly from Databricks with standard SQL, keeping data in place. The post walks through a wealth management example that joins on-premises portfolio holdings with FactSet data subscribed from the Databricks Marketplace in a single query, then powers a dashboard and a Genie natural-language agent without pipelines, materialized views, or persisted objects. It notes the same pattern applies to other industries, such as healthcare, manufacturing, and retail.

  4. Databricks Blog38

    Vital Skills: Upskilling Genie Code with Healthcare and Life Sciences skills

    Databricks introduces Vital Skills, packaged playbooks built on the open Agent Skills standard that give Genie Code healthcare and life sciences domain knowledge, such as cohort building, HIPAA de-identification, RNA-seq analysis, and open-weight model deployment. On BiomniBench, Genie Code's average score on the 5 in-scope tasks rose from 71.9 to 76.9 across three trials, while scores on 18 out-of-scope tasks held steady. The skills are open source and installable in Genie Code, Genie One, or Genie Code CLI.

  5. Google Research58

    The sleep framework: A new way for LLMs to retain knowledge in continual learning

    Researchers introduce a "sleep framework" for continual learning in which LLMs alternate between an active wake phase that stores temporary information in short-term memory and an offline sleep phase that consolidates it into long-term parameters via Mixture-of-Experts parameter expansion, knowledge distillation, and a "dreaming" self-improvement loop. The authors report that this approach avoids catastrophic forgetting in continual translation, scales to 10M tokens on BABILong, reaches 48.9% on single-task and 46.2% on CPT factual-knowledge evaluations, and scores 79.2%, 69.0%, and 46.1% on AIME 24, AIME 25, and HMMT 25.

  6. Databricks Blog38

    From spot fixes to measured gains: improving document extraction accuracy with ai_extract and MLflow Prompt Registry

    A Databricks engineering post describes extracting roughly 40 structured fields from partner order documents across dozens of PDF layouts, where a schema fix that halved errors on the revision field raised total errors across all fields by 9%. The authors recommend putting field rules in schema descriptions, document-level logic in global instructions, precision mode (ai_extract 2.1) for complex structures, and citations to diagnose errors, while versioning prompts and schemas in MLflow Prompt Registry with aliases such as @prod and @candidate. They evaluate candidates against a frozen human-reviewed test set, tracking accuracy, coverage and run-to-run variation, and report that restructuring the prompt and schema plus precision mode cut overall errors by 32% and revision errors by 64%.

  7. Allen Institute for AI Blog58

    Can we trust LLMs to judge whether a research idea is novel?

    Ai2 released "Old Ideas, Novel Problems," a study with a benchmark, data, code, and an automated pipeline for testing LLM novelty judges. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, retrieval and extra reasoning did not consistently help, and two purpose-built novelty evaluators scored worse than a simple prompted Claude Opus 4.6 baseline while costing far more.

  8. AWS Machine Learning Blog50

    Exowatt beats frontier models on accuracy with SageMaker serverless model customization

    Exowatt used Amazon SageMaker AI serverless model customization to fine-tune Qwen3-1.7B for edge alarm triage (0.911 macro-F1 vs 0.698 for Claude Sonnet 4.6) and Qwen3-14B for cloud energy-market analysis (0.975 macro-F1 vs 0.625 for Claude Opus 4.7, at about 49× lower cost per 1,000 answers). The pipeline spans seven AWS serverless services, with the edge model quantized to GGUF and run offline via llama.cpp on existing CPUs, and the cloud model served through Amazon Bedrock Custom Model Import. Exowatt moved from experimentation to production in nine days and plans to reuse the pipeline across fleet tasks.

  9. Hugging Face Blog58

    Can we trust LLMs to judge whether a research idea is novel?

    Ai2 released "Old Ideas, Novel Problems," a study with a benchmark, data, code, and an automated pipeline for testing LLM novelty judges, built from ICLR 2026 peer reviews to form Human-Only and Human+Generated evaluation sets. Across six judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, retrieval and extra reasoning did not consistently help, and two purpose-built novelty systems scored below a simple prompted Claude Opus 4.6 baseline.

  11. Meta Engineering36

    Python Typing Survey 2026: AI Tools and the Shifting Checker Landscape

    The 2026 Python Typing Survey, with 1,794 responses, found type hint use stable across three years (85.5% in 2026) while checker selection shifted sharply, with 49.9% of respondents selecting two or more checkers and speed cited by 56.5% as the top reason for choosing one. Missing type information in third-party libraries was the leading difficulty at 65.1%, and AI tools wrote type hints for 42.5% of 1,517 respondents, with 24.4% saying an AI coding tool invoked their checker.