iCreatAI News
Back to main site
Skip to content
#Data / training
Oct 24Sat
  1. AWS Machine Learning Blog20

    Visualize points and routes with GeoJSON in Amazon Quick Sight layer maps

    Amazon Quick Sight layer maps now support Point and LineString geometry types from GeoJSON files, letting users plot individual locations and routes joined to a Quick Sight dataset by a shared key. The post explains the required GeoJSON structure and dataset join, compares layer map points with the native point map, and covers advanced techniques for controlling the viewport, layering geometries, and updating GeoJSON files via Amazon S3 and the Quick Sight APIs.

Oct 23Fri
  1. AWS Machine Learning Blog18

    How U.S. Hunger connects leaders and data using Amazon Quick

    U.S. Hunger, which reports delivering more than 250 million meals, uses Amazon Quick to let CEO Rick Whitted explore Full Cart datasets directly and to give 17 colleagues across leadership, technology, finance, and compliance shared Spaces for KPI definitions, financial planning, and research. Data from its Impact Manager platform is refreshed from SQL Server into SPICE nightly at 2 AM ET, and the organization plans a Q4 2026 enterprise KPI framework routing NetSuite, Salesforce, and other systems through Boomi into Amazon Redshift.

  2. AWS Machine Learning Blog34

    Train Qwen3 on AWS Trainium with torchtitan

    This post walks through training Qwen3 three ways on a single trn2.48xlarge using torchtitan-neuron: via Hugging Face Transformers v5 as a modeling backend, via torchtitan's native Qwen3 by changing config.model_spec, and via NKI mixture-of-experts kernels from the torchtitan-neuron recipes, all using the same torchrun launch pattern. It also shows how to load and save published safetensors checkpoints, attach a custom NKI rope_hf kernel through torchtitan's override registry, and scale with data and tensor parallelism across 64 logical NeuronCores.

  3. TechCrunch · AI58

    Left alone with nothing to do, AI agents started playing

    MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no specific task, and report in an arXiv paper that the agents began inventing their own challenges, behavior the researchers say bears the hallmarks of play. The Eko agents that spent those 30 hours playing generally beat identical agents that had never been on the island when later given specific challenges, such as getting as many objects as possible onto the tallest rock spire. Six of the 13 agents also learned they could throw a block farther by letting go of it mid-jump, something the researchers hadn't meant to make possible.

  4. Hugging Face Blog62

    Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

    Microsoft and Hugging Face describe using ThinkingBox's executable terminal-database-state checks as binary rewards to fine-tune Qwen3.8-27B with GRPO for 50 steps on an offline-selected pool of 187 tasks, raising pass@1 from 51.70% to 60.78% and tasks passing all 20 attempts from 38 to 107 on the 507-task ThinkingBox-Bench, while pass@20 dipped slightly from 453 to 446. The post details offline policy-relative difficulty selection via Jeffreys posteriors, the reward loop, cost-per-dependable-task falling from $24.36 to $8.00, and releases the training stack in microsoft/thinkingbox-training.

    Why it matters: It offers a concrete, reproducible recipe for converting stateful agent grading into RL rewards, plus an honest account of the repeatability-versus-coverage tradeoff.

  5. The Verge · AI50

    Meta’s AI could still scrape news sites, legal experts say

    Meta revised its business terms to remove language saying it would "train" its AI on content from sites linked to business accounts, after backlash from the publishing industry. Legal experts told Nieman Lab the changes are insufficient, since the terms still let Meta use scraped content to "develop" any "business product," a term they say is undefined and broad enough to include training. Meta spokesperson Daniel Roberts said content owners can use robots.txt to signal preferences and the update doesn't grant rights that supersede them.

  6. The Verge · AI60

    Sure, let’s smoosh SpaceX and Tesla together, who cares

    SpaceX and Tesla have formed special committees to discuss a merger, with Musk saying on SpaceX's earnings call that "there's a lot of merit to combining" the companies to cooperate on superintelligence software. The piece weighs the Mars narrative against financial motives, noting Musk's over-50% voting control at SpaceX versus roughly 20% at Tesla, and that a stock-for-stock deal at $2 trillion or more would trigger payouts and accelerated vesting under his Tesla compensation package. It also covers SpaceX's Q3 SI solutions revenue of $7 billion, Tesla's second consecutive quarter of declining profits, and the Terafab and Macrohard projects, which the S-1 says are in very early stages with financial terms and IP rights not finalized.

  7. AWS Machine Learning Blog58

    Announcing new ways to customize models with SageMaker AI, without managing GPUs

    AWS announced the general availability of training sessions in Amazon SageMaker AI serverless model customization, letting customers write and coordinate their own post-training loop, including custom loss functions and reward signals, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage. The sessions support techniques such as SFT, DPO, RLVR, RLAIF, and multi-turn RL, with LoRA or full fine-tuning, and are available with support for gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B. Customers can start via the SageMaker Python SDK and cookbook or in SageMaker Studio, with data kept in their own S3 buckets and IAM, KMS, and VPC controls.

  8. AWS Machine Learning Blog58

    Write your own training loop — without managing GPUs — with Amazon SageMaker training session APIs

    Amazon SageMaker AI introduces training session APIs that let you write a custom training loop in Python on your local machine while SageMaker runs GPU-bound operations such as forward_backward(), optim_step(), and sample() on managed accelerators with your base model and LoRA adapter. The APIs are organized around ServiceClient, TrainingClient, and SamplingClient in the SageMaker Python SDK, and the post walks through a minimal GRPO loop on GSM8K (held-out accuracy rising from 56% to 75% in 10 steps) plus a Code RL recipe on competitive programming problems (held-out accuracy from 40% to 50.5%). A companion cookbook provides ready-to-run SFT and RL recipes.

  9. Latent Space62

    Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO

    In this interview, Modal CEO Erik Bernhardsson explains why fixed-term GPU capacity contracts look like rectangles while real workloads grow, spike, and fall, and describes Modal's approach of pooling demand across thousands of customers to use capacity elastically. He also discusses Modal's sandboxes, which he says have grown to roughly a third of revenue, and the infrastructure work—such as checkpointing and GPU memory snapshots—that makes fast startup possible. Modal reports crossing $1 billion in annualized run rate, raising a $750 million Series D at a $16 billion valuation, and spinning up nearly three billion sandboxes.

  10. AWS Machine Learning Blog32

    Stop guessing auto scaling targets for TTS on Amazon SageMaker AI

    This AWS post deploys Kokoro-82M, an 82-million-parameter non-autoregressive TTS model, as a custom container on an Amazon SageMaker AI real-time endpoint via AWS CDK stacks, then replaces a guessed auto scaling target with values derived from two live measurements: Q_max (the highest per-instance concurrency that still meets a 3,000 ms p95 time-to-first-audio SLO, measured at 50) and T_total (end-to-end scale-out time, measured at 260.2 s and planned against 320.2 s). Scaling uses the native ConcurrentRequestsPerModel metric, and a five-command tool chain (drift, qmax, ttotal, plan, validate) computes the scale-out threshold (37.5), scale-in threshold (25.0) and asymmetric cooldowns by arithmetic. The authors note the methodology generalizes to other latency-sensitive, concurrency-bound real-time SageMaker AI endpoints, while the specific numbers hold only for Kokoro-82M on ml.g5.xlarge.

  11. The Decoder52

    Only 13 percent of companies see real returns from AI, Gartner survey finds

    A Gartner survey of 2,002 CIOs and technology leaders, presented at the Gartner IT Symposium in Orlando, found that only 13 percent of respondents report major or transformative returns from AI. While 58 percent of CIOs are tasked with delivering cost savings through AI, 51 percent expect AI to raise total operating costs for their tech stack, and 44 percent of CEOs struggle to demonstrate AI-driven financial returns to stakeholders. Gartner analyst Alicia Mullery sees a growing gap between the "innovation race" vendors are running and the "outcomes race" companies need to win, while Bloomberg CEO Vlad Kliatchko says it is too early to measure AI's economic payoff.

Oct 22Thu
  1. Google DeepMind58

    Beyond tree cover: mapping the world’s forest types

    Google DeepMind and Google Research, with the World Resources Institute and IIASA, released Forest Typology (ForTy) 2020, a global 10-meter-resolution map classifying land into six FAO-based categories: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land. An accompanying early time series tracks annual changes from 2017 to 2025, including where natural forests were replaced by plantations or farms. Validated against more than 9,000 expert-reviewed sites, ForTy identified the correct forest type 90% of the time and distinguished forest from non-forest land 95% of the time, and the map and time series are freely available on Google Earth Engine under an open license.

  2. Latent Space20

    You are the Team — Anekha Sokhal, Moshi

    In this AI Engineer New York podcast episode, Moshi CEO Anekha Sokhal tells swyx how she moved from mathematics and energy derivatives trading at JPMorgan to founding Moshi, a recommendation engine that connects market information to what matters for a trader's book. She describes Moshi as what might happen if Bloomberg, Claude and Netflix had a baby, and discusses trusted sources, document extraction, building versus buying models, and the pressure on AI engineers told "you are the team."

  3. The Decoder74

    OpenAI agents analyze mouse brain data in 32 minutes, a job that would take humans months

    OpenAI, PerturbAI and Scripps Research deployed thousands of AI agents running on GPT-Rosalind to analyze a whole-brain CRISPR atlas of millions of mouse cells, completing the analysis in 32 minutes versus an estimated several months to a year by hand. Each agent handled a single gene perturbation, analyzed cell responses and searched the literature for explanations, surfacing unexpected differences between two disease-related genes and a possible link between cholesterol regulation and neurodegeneration. In blind comparisons, human experts preferred the distributed agents' reports over a single agent's 75 percent of the time, and the resulting "Reasoning Atlas" is on GitHub with a paper on bioRxiv.

  4. MIT Technology Review · AI30

    AI’s hidden storage bill: What inference leaves behind

    This WD-sponsored article argues that AI infrastructure planning overlooks data storage costs, since inference runs continuously and leaves back-end data behind—citing an internal measurement where an 8-second 720p video clip returned as a 3.97MB file left more than 500MB of associated data. It cites McKinsey's projection that inference demand will grow at a 35% CAGR through 2030 and a WD-sponsored IDC survey of 763 IT and business decision makers in which nearly 95% reported data volume increases from generative AI over the past 12 months. It recommends planning storage around the full AI data life cycle, with HDDs as a cost-effective high-capacity foundation for retained data.

  5. Hugging Face Blog38

    Ensemble forecasting with IBM Granite time-series models

    IBM introduces Granite Timeseries Ensemble, a recipe that combines four pre-trained Granite models (PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1, and TTM-r3) through one extensible forecasting interface to return median and quantile forecasts without retraining or fitted ensemble weights. The default Probabilistic Ensemble pools member quantiles with equal influence, while an adaptive IQR Ensemble weights members by forecast spread; on 97 GIFT-Eval configurations both ensembles beat all four individual members on aggregate MASE and average MASE Rank, with the Probabilistic Ensemble reaching 0.683 MASE versus 0.685 for the best single member. The authors note the MASE margin is small, that PatchTST-FM-r2 retains the lowest aggregate CRPS/MWQL, and that the recipe forecasts channels independently and loads four checkpoints, increasing inference time and memory.

  6. NVIDIA Blog40

    Pioneers Across Europe Deploy NVIDIA Earth-2 Open Models to Power Real-World Weather and Climate Prediction

    NVIDIA Earth-2 offers developers and researchers a family of open AI models and tools, running on NVIDIA computing, for global weather prediction, detailed regional forecasts and environmental research. StormScope EU, a regional short-term forecasting model trained on European weather and climate data, uses EUMETSAT satellite observations to predict cloud conditions across 16 Meteosat channels at 10-minute intervals, and is being shown in a live interactive demo at GTC Berlin built with Earthmover. Organizations including AXA, Vaisala Xweather, the University of Manchester, TotalEnergies and Space42 are adapting Earth-2 models such as FourCastNet 3, CorrDiff and PhysicsNeMo for cyclone risk, lightning-enhanced forecasting, U.K. air quality, solar forecasting and radar-based surface-height estimation.

  7. Hugging Face Blog75

    NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

    NVIDIA introduces Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise databases via in-context learning in a single forward pass, requiring no task-specific training, feature engineering, or table flattening. 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 benchmark it surpasses AutoGluon ensembles by 8% in mean reciprocal rank (MRR). It scales to databases with 500B+ rows and is available through NVIDIA's open-source structured-data-models and kumo-engine libraries or as an NVIDIA NIM.

    Why it matters: Useful for ML teams weighing whether to replace production XGBoost pipelines with a pre-trained relational model, though the reported gains come from NVIDIA's own paper and benchmarks.

  8. Hugging Face Blog58

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

    Dharma-AI describes VocabX, a method that rebuilds a small model's tokenizer around the cost of each token rather than its frequency, combining sub-word merges with multi-word "macro" tokens under a shared cost criterion. In a production Portuguese essay grader built on Qwen3-4B, the approach cut average response length from 350 to 101 tokens and measured inference cost by 77.0% with the same base model and training recipe, while the bench score rose from 0.703 to 0.810. The article previews results from a VocabX paper still under review, noting the numbers may change before publication.

Oct 21Wed
  1. Ars Technica · AI58

    The $15 billion ad-powered AI startup your doctor may already be using

    OpenEvidence, a $15 billion AI startup, is being adopted by doctors and is backed by investors including Thrive Capital and the Mayo Clinic, with founder Nadler saying the end goal is literature-based drug discovery. Anthropic's Claude Science creator Alexander Tarashansky has joined the company, which faces competition from OpenAI and Anthropic in healthcare AI. Studies have shown mixed results, with a Mayo Clinic study finding accurate evidence-based recommendations in five sampled cases while a Nature paper found it underperformed Gemini and GPT on clinical benchmarks.

  2. Mistral AI58

    Mistral Advances AI for Engineering with NVIDIA

    Mistral and NVIDIA are collaborating to integrate Mistral Large 4 and Mistral's data generation and physics modeling pipelines with NVIDIA CUDA-X libraries, Warp and Omniverse to accelerate engineering simulation and enable closed-loop workflows connecting simulation, AI model development and design. The partnership will develop public benchmarks and open datasets for evaluating physics models and agentic AI in CAE and EDA, assessing agents' tool use, multistep workflows, technical correctness, execution time and computational cost. The work builds on the companies' prior co-development of Mistral NeMo and their collaboration through the NVIDIA Nemotron Coalition.

  3. NVIDIA Blog73

    NVIDIA Open Models Advance Europe’s Sovereign Intelligence Across Industries

    NVIDIA is expanding its family of open models across language, robotics, biology and quantum computing, including the new Nemotron 3.5 Super VL and Nemotron 3.5 Bolt models, Nemotron Retriever models, the Kumo family for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo for healthcare and science, and Ising for quantum computing. The company says the open models give organizations a foundation they can customize with local data and deploy on infrastructure they control, with Nemotron built using data and expertise from the NVIDIA Nemotron Coalition including European partners Black Forest Labs, H-Company and Mistral. Availability details include Nemotron 3.5 Super VL on Hugging Face, ModelScope, OpenRouter and build.nvidia.com, with Nemotron 3.5 Bolt E4B 4GB expected Friday, Oct. 23.

    Why it matters: The piece is a broad product roundup that shows how NVIDIA is positioning open models as a sovereignty-friendly foundation for European enterprises and researchers.

Oct 20Tue
  1. MIT Technology Review · AI20

    The new economics of enterprise AI

    The article argues that as enterprise AI shifts from isolated experiments to a foundational production layer, the key metric becomes effective cost per useful outcome rather than the price of a token or GPU-hour, with utilization central to the economics. It compares three operating models — a bespoke in-house stack, on-demand hyperscaler consumption, and an integrated private AI platform — and stresses that owning GPUs only improves economics with sufficient demand and operational discipline. It also describes HPE Private Cloud AI as a production-ready foundation with a 90-day forward-deployed engineering engagement, citing HPE-commissioned independent testing in which an integrated platform at high utilization delivered a lower effective cost per token than repeatedly renting equivalent capacity.

  2. MIT Technology Review · AI22

    AI adoption by industry shows what other sectors can learn

    TripleTen survey data shows tech leads AI adoption, with finance second (48% of workers asked to use AI chatbots), while healthcare (18%) and education (25%) lag due to compliance-heavy and human-facing operations. Across industries, 79% of workers report being asked to learn at least one new skill, and 88% of employee-led AI innovations became standard company practice when managers learned them from an employee. Out-of-pocket training spending varies widely, from $1,758 on average in tech to $883 in education, with nearly half of healthcare and education workers spending nothing.

  3. MIT Technology Review · AI27

    A visit to Samsung’s massive new semiconductor hub

    Samsung is building a large semiconductor manufacturing complex in Pyeongtaek, South Korea, where three fabs already run around the clock and a fourth is coming online; construction of a fifth was halted during a 2024 drop in memory chip demand but restarted when the AI boom sent demand soaring. The site covers an area about the size of New York City's Central Park, and safety officer Choi Hye-rin, 35, who earns 140,000 won ($97) a day, says the rate is higher than at other construction sites and the work is relatively easy. Samsung predicts memory chip demand will keep surging through 2028, and the build-out is its attempt to reclaim its lead after watching domestic rival SK Hynix pull ahead in high-bandwidth memory for Nvidia's AI accelerators.

Oct 17Sat
  1. Google Research71

    Grounding AI agents in biological reality: Building blocks for autonomous wet labs

    Google researchers describe three projects aimed at connecting AI agents to real wet-lab experimentation: the VERID framework, a semantic digital twin of a lab workcell that tracks 3D state, runs fast virtual dry runs, blocks unsafe commands and logs synchronized video and actions; closed-loop aptamer design over three design-test cycles that raised on-target hit rates to 33.5% for myoglobin and 42.1% for α-thrombin; and agent-guided epitope excision for de novo binder design against FAP, which accelerated each folding cycle more than eightfold and, in a human–agent comparison at equal compute, nearly doubled the in-silico hit rate of a scientist working alone.

    Why it matters: It offers a concrete, example-rich account of how agents are being tied to physical lab hardware and closed experimental loops, including failure modes and safety safeguards.

Oct 16Fri
  1. Google DeepMind58

    Accelerating the search for antimicrobial peptides with Google Antigravity

    Google describes how its agentic development platform, Google Antigravity, lets biologists discover, structurally evaluate and prioritize antimicrobial peptide candidates in a single session, using ProtNLM to scan millions of uncharacterized UniProt proteins. A technical walkthrough shows a researcher and agent identifying a novel antimicrobial peptide candidate with predicted activity against respiratory tract infection pathogens, now undergoing wet-lab validation. The post also introduces the Biomolecular Structure Viewer, a 3D visualization tool embedded in Antigravity that shares visual state between researcher and agent.

Oct 15Thu
  1. Mistral AI67

    Introducing Timestral

    Mistral releases Timestral, its first family of time-series foundation models in four sizes from 5M to 209M parameters, with the largest ranking first on GIFT-Eval, TIME and fev-bench among pre-trained zero-shot models. Timestral produces probabilistic multivariate forecasts, supports past-only and known-future covariates, and runs on CPU without a GPU. It combines synthetic-first pre-training, covariate-aware attention and decoupled horizon forecasting, and can be fine-tuned on proprietary data through Mistral Forge.

    Why it matters: Useful for teams weighing time-series foundation models, since it reports benchmark standings, a size-versus-accuracy trade-off and CPU latency figures alongside the adaptation options.

  2. Google Research58

    Research@ 2026: 10 ways we're amplifying human ingenuity

    At its Research@ event in Mountain View, Google Research outlined 10 updates, including expanded early access to Google Earth AI's experimental geospatial capabilities, a multi-agent orchestrator for reproducing prior methods, and public access to its Deep Review verification system. Other highlights include the Synaptics Astra SR200 Series chip with Google's co-designed Coral NPU, upcoming integration of TimesFM and TabFM into Gemini for business, and a Lancet study in which 98 patients consulted AMIE ahead of urgent care visits. Google.org is also providing $10 million in funding for crisis resilience and public health work.

Oct 14Wed
  1. GitHub Blog · AI & ML34

    GitHub’s model for secret detection: Finding more secrets with less noise

    GitHub describes a two-stage secret scanning pipeline that uses regex candidate generation followed by a fine-tuned ModernBERT classifier that reads each candidate alongside its file path, variable names, and surrounding code. Compared with its previous LLM-based pipeline, GitHub reports 11.5% higher repository owner-weighted acceptance, 86% lower median post-push scan execution time, and more than 98% lower serving cost, with offline evaluation showing 13% higher precision and 19% higher recall. The efficiency gains made it practical to extend generic secret detection to push protection.

Oct 13Tue
  1. Microsoft Research52

    Simulating student learning for better AI tutoring

    Microsoft Research introduces StudentSim, a research system that trains personalized AI student simulators from publicly available, de-identified learner records to evaluate and improve AI tutors before testing with real learners. The work also presents StudentSimEval, a benchmark spanning chess, second-language English writing, and mathematics that measures behavioral fidelity and guidance responsiveness, with StudentSim outperforming baselines such as GPT-5.4 and Maia2. In a chess tutor reinforcement learning study, a tutor trained with StudentSim feedback reached 90.5% accuracy versus 75.7% for the no-RL baseline and 71.6% for the GPT-5.4 reward condition, as rated by blind human chess experts.

Oct 8Thu
  1. Microsoft Research71

    Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses

    Microsoft Research Asia introduces Harnessed Agentic RL, a paradigm in which the same agent harness used in deployment participates directly in reinforcement learning, and open-sources Agent Lightning v1.0, a roughly 3,500-line framework with an API Gateway, Rollout Controller, and Customized Trainer built on verl. Agents run as standard Kubernetes jobs without commercial sandbox services, and Collocated Async RL shares GPUs between rollout and updates, achieving about a 2x end-to-end speedup over synchronous RL. An end-to-end coding agent pipeline raised Qwen3.5-9B from 41.8% to 56.4% Pass@1 on SWE-bench Verified using about 6,000 training samples.

    Why it matters: It offers a concrete, reproducible recipe for training deployed agents in place, with code size, infrastructure, and benchmark numbers stated explicitly.

Oct 6Tue
  1. Google Research58

    Unlocking Earth AI’s planetary geospatial foundation models for global public health

    Google researchers describe using the Population Dynamics Foundation Model (PDFM), part of Google Earth AI, to turn privacy-preserving search trends, mobility, built-environment and environmental data into monthly location embeddings that can be plugged into existing epidemiological models without task-specific fine-tuning. Independent partner evaluations across five public health challenges — MMR vaccination at the US–Canada border, US cardiovascular disease nowcasting, dengue forecasting in Mexico, postpartum depression risk prediction, and cholera emergence in the DRC — reported matched or improved performance over conventional inputs. The authors note current limitations such as static snapshots and say PDFM embeddings are available in Preview as Population Dynamics Insights, with no-cost access for select academic and public health research.

Oct 1Thu
  1. Microsoft Research40

    Forecasting space weather risks on power grids

    A Microsoft Research intern developed a machine learning pipeline that uses forecast-time solar-wind data, AE and Dst index forecasts, local geology and grid data to produce location-specific geomagnetically induced current risk estimates for 66,935 substations in the continental United States, 30 to 60 minutes ahead. During the 2020-2026 evaluation period the pipeline detected nearly 80% of major space-weather events, with detection rates of 76.5% for events ≥10 nT/min, 81.2% for ≥20 nT/min and 64.1% for ≥50 nT/min, and it produced estimates for all substations in about 333 milliseconds. The authors note that further validation with utilities and operational data would be needed before use in grid operations.

Sep 28Mon
  1. Mistral AI46

    Hallo, Deutschland!

    Mistral has opened a hub in Munich housing research teams focused on Physics AI and Industrial AI alongside applied engineers serving enterprise partners. The company says it will build one gigawatt of European compute capacity by 2030 and points to its open-weight architecture, which lets customers run models on their own infrastructure under European law. Following its May 2026 acquisition of Emmi AI, Mistral says it is working with BMW on crash simulations and engineering AI and with Siemens Energy on industrial AI applications, and has formed a research partnership with the Technical University Munich.

Sep 10Thu
  1. Mistral AI40

    Cloudera and Mistral Partner to Bring Specialized, Sovereign Intelligence to Enterprise Data

    Mistral and Cloudera announced a partnership to integrate Mistral's models with Cloudera's hybrid data platform, enabling enterprises to run inference across private and public cloud, on-prem, and fully air-gapped environments. The collaboration also lets enterprises train custom models on proprietary data within controlled environments while retaining ownership of both data and resulting intelligence. Cloudera cites 30 exabytes of customer-managed data on its platform, and the partnership is positioned around demand for sovereign AI that keeps data, intelligence, compute, and operations under customer control.

Jul 26Sun
  1. Berkeley AI Research38

    Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction

    The post introduces ABBEL, a framework that replaces full interaction history with periodically updated natural-language "belief states" and supervises their contents through belief grading, including an autoencoding-inspired reconstruction objective. On CollabBench collaborative coding, reconstruction-based belief grading roughly halved the gap to full-context models and trained in 50 steps instead of 100, while using fewer peak tokens; on Combination Lock and multi-objective QA, domain-knowledge grading and a peak belief length penalty further improved efficiency. The authors argue that isolating and supervising summaries addresses the performance cost of recursive summarization in long-horizon tasks.

Jul 7Tue
  1. Berkeley AI Research38

    Intelligence is Free, Now What? Data Systems for, of, and by Agents

    A Berkeley EPIC Data Lab perspective argues that as inference costs fall toward near-zero, data systems face three challenges: redesigning for agent workloads (e.g., exploiting the 80-90% duplicate sub-queries in agentic speculation), building a substrate for agent swarms (structured memory, concurrent shared-state edits, coordination), and letting agents synthesize custom data systems that can be trusted. It surveys work including Bespoke OLAP, GenDB, and custom key-value store synthesis, and sketches a future where agents and data systems co-evolve and blur together.

May 8Fri
  1. Berkeley AI Research38

    Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling

    This post surveys parallel reasoning methods for LLMs, contrasting fixed approaches like self-consistency, Best-of-N, and tree search with adaptive parallel reasoning (APR), where the model itself decides when to parallelize, how many threads to spawn, and how to coordinate them. It describes fork-join inference designs, including engine-modifying KV cache stitching (Multiverse) versus client-side orchestration (ThreadWeaver), and covers training via SFT demonstrations and RL rewards based on critical-path efficiency gated by correctness. It closes with evaluation differences across Multiverse, ThreadWeaver, NPR, and Parallel-R1, plus open questions on stability and whether parallelization mainly helps as a training-time exploration scaffold.