Damian Bogunowicz

Rigid tools and fragmented workflows slow #engineering teams down. In this recorded session, discover how #Dyad —built on Julia—enables modern, scalable model design. We’ll walk through acausal #modeling, reusable components, and integrated #simulation using an RLC circuit example. Learn how this smarter, unified approach applies across domains and simplifies complexity. https://lnkd.in/eQxnNyuD #JuliaLang #SystemSimulation #AcausalModeling #ReusableModels #EngineeringDesign #ModelBasedDesign #TechnicalComputing #SimulationSoftware #DigitalTwinTechnology

8

LLM-based agents are showing remarkable progress in real-world software engineering tasks, but the field faces two key challenges:    1️⃣ A lack of large-scale, interactive SWE datasets for training  2️⃣ Benchmark contamination that skews progress measurement At #NeurIPS2025, Ibragim Badertdinov, Lead Research Engineer, Nebius AI R&D, will present SWE-rebench: an automated pipeline for task collection and decontaminated evaluation of SWE agents. The research introduces a new, continuously updated benchmark designed to ensure contamination-free evaluation, along with a large-scale dataset of 21,000+ interactive Python SWE tasks. Don’t miss how Nebius is building the foundation for the next generation of SWE agents with reproducible, scalable and decontaminated evaluation.    🗓️ Thu, Dec 4 | 11:00 AM – 2:00 PM PST | San Diego, CA  📍 Exhibit Hall C,D,E #106 #NeurIPS #AIResearch #SoftwareEngineering #SWErebench #LLM #Agents 

11

The JEPA paper family, 𝘳𝘦𝘷𝘰𝘭𝘶𝘵𝘪𝘰𝘯𝘪𝘻𝘪𝘯𝘨 𝘈𝘐 𝘧𝘳𝘰𝘮 𝘵𝘰𝘬𝘦𝘯 𝘨𝘶𝘦𝘴𝘴𝘪𝘯𝘨 𝘵𝘰 𝘭𝘢𝘵𝘦𝘯𝘵 𝘳𝘦𝘢𝘴𝘰𝘯𝘪𝘯𝘨, really caught my attention and highlights a game-changing shift away from traditional LLMs’ purely generative paradigm. JEPA (Joint Embedding Predictive Architecture) originated in Yann LeCun’s 2022 position paper https://lnkd.in/dMUK5NbB and has since evolved through variants targeting self-supervised learning in vision, video, and multimodal domains. These works span foundational theory to specialized applications like LeJEPA and V-JEPA, all emphasizing latent-space prediction over token generation. By modeling semantic structure in latent space rather than superficial tokens or pixels, JEPA unlocks several key advantages:  -> Latent predictions, not raw reconstruction:     - Unlike SOTA LLMs glued to token-by-token generation (and prone to compounding errors), JEPA predicts compact embeddings of future or missing content—capturing meaning without being distracted by noisy surface forms.  -> Dramatic efficiency gains:     - Operating in a compact latent space (instead of huge token vocabularies) cuts parameter counts and compute while improving generalization—challenging today’s bloated LLMs on reasoning and multimodal tasks.  -> Superior multimodal fusion:     - Joint vision–language embeddings align modalities natively, excelling in cross-modal reasoning and transfer where token-based LLMs often struggle with disjointed vision-text handling.  -> Predictive stability over generative fragility:     - JEPA’s non-generative objective (predict structure, not raw pixels/text) leads to more stable training and richer world models—sidestepping classic LLM failure modes like brittle long-chain hallucinations. Recent benchmarks show VL-JEPA outperforms SOTA on visual reasoning, retrieval, captioning, and video prediction with leaner architectures. The future may be shaped by latent-first AI that truly understands before it generates. 2022 : Path Towards Autonomous... - Theory/world models https://lnkd.in/dMUK5NbB 2023: I-JEPA - Image-SSL arxiv.org/abs/2301.0824 2024-25: V-JEPA(2), ACT-JEPA - Video/Action arxiv.org/abs/2506.09985, arxiv.org/abs/2501.14622 2025-26: LeJEPA, VJEPA - Scal./Prob. arxiv.org/abs/2511.08544, arxiv.org/abs/2601.14354 #AI #JEPA #MultimodalAI #WorldModels #YannLeCu

6

Legacy airflow assumptions no longer apply to today’s dense, high-performance computing environments. Cooling needs to be agile, accurate, and modeled for variability. CoolSim adapts with your design, not against it. Don’t rely on outdated playbooks. 🌐 https://lnkd.in/eXu34f-a #coolingsmarter #datacenterdesign #cfdsimulation #equipmentefficiency

1

With millions of real-world scenarios to test, how do you validate autonomy at scale? Neural Sim ⚙️🤖.  Instead of slow, fragmented on-road testing, Neural Sim reconstructs thousands of drive logs into dynamic, photo-realistic 3D environments and sensor-level scenarios that mirror real-world behaviors, enabling large-scale evaluation of self-driving systems. 🔗 Read how Neural Sim accelerates safe, scalable autonomy development for self-driving programs → https://lnkd.in/gNHDtbC6 #Autonomy #Simulation #AI #Safety #NeuralSim

256

4 Comments

Learn why a balanced testing strategy, from fast unit tests to comprehensive E2E tests, is crucial for production-ready projects. Vincent Margot, Ph.D. shares a hands-on look at Python testing with pytest, including parametrization and mocking.

3

What if your AI models could work together like a highly specialized, dynamic team? 🤯 Our latest article dives into the Swarm Agentic Workflow in LangGraph, revealing how to build collaborative AI systems that go beyond linear chains. Prepare to rethink multi-agent possibilities! https://lnkd.in/damBAWm3 #LangGraph #AIagents #MultiAgentSystems #SwarmAI #GenerativeAI

1

Meet HCMBench, the open-source toolkit for evaluating hallucination correction models. Hallucinations in RAG systems erode trust—HCMBench helps fix that. With modular pipelines, diverse datasets (like RAGTruth, FAVABENCH, and more), and multi-level metrics (HHEM, MiniCheck, FACTSJudge), it’s the gold standard for testing how well models correct LLM outputs. Whether you're building or benchmarking HCMs, HCMBench gives you the power to rigorously measure and improve performance, sentence by sentence, claim by claim. Explore the toolkit! https://bit.ly/4328iOQ #OpenSource #RAGaaS #LLM #HallucinationCorrection #AITrust #Vectara #HCMBench

20

ROCm 7 isn’t just open. It’s built for speed and simplicity. With tight PyTorch integration and support for top inference frameworks, developers can get up and running fast. Hear more from Steen Graham.

62

1 Comment

We’re excited to introduce TorchSpec, a torch-native framework for scalable speculative decoding training developed by the TorchSpec and Mooncake teams. By streaming hidden states from inference engines to training workers via Mooncake, TorchSpec enables fully disaggregated pipelines where inference and training scale independently. 🔗 Read our latest blog from TorchSpec & Mooncake teams: https://lnkd.in/gUrnS4pQ LightSeek Foundation #PyTorch #TorchSpec #Mooncake #OpenSourceAI

85

1 Comment

Today at CES, we’re excited to announce the official open-source release of OpenTau (τ) by Tensor. At Tensor, we’re pushing the frontier of large foundation models for Physical AI. A Vision-Language-Action (VLA) model is a multimodal foundation model that integrates vision, language, and action. VLAs are emerging as a leading approach for Embodied AI—powering applications across autonomous driving, robot manipulation, and navigation. OpenTau (τ) is Tensor’s open-source training toolchain for frontier VLA models—built to make training reproducible, accessible, and scalable. We believe open research is how the community moves faster, together. OpenTau unlocks state-of-the-art AI training capabilities for everyone—technologies that were previously out of reach—including: ✅ Co-training on an adjustable mixture of heterogeneous datasets ✅ Discrete actions for fast VLM convergence ✅ Knowledge insulation between the VLM backbone and the action expert ✅ VLM dropout to reduce overfitting ✅ A reinforcement learning pipeline specific for VLA And more… We’re inviting researchers, developers, and builders to star the repo, fork it, and start experimenting. Check it out on GitHub: 👉 https://lnkd.in/e5FCphqD 📍 If you’re at CES in Las Vegas, come say hello: LVCC West Hall — Booth #5701, Entrance W3 Or see our car at the Fontainebleau — 4th Floor, The Foundry #DeveloperCommunity #GitHub #OpenSourceSoftware #CES2026 #DeepLearning #TensorAuto #Tensor #TensorAI #TensorRobocar

65

MiniMax M2.5 is now available on Together AI, bringing structured planning and SOTA coding to real-world agent workflows. Excited to partner with Together AI to power what’s next!

129

2 Comments

We’ve been told for years now that in the world of Large Language Models, 'Scale is King.' The recipe seemed simple: more data, more compute, and more parameters. But what if we’re hitting the limit of brute force? What if the secret to smarter AI isn’t more data, but better geometry? Welcome to the show. Today, we’re tearing up the standard scaling law playbook to look at a radical new framework: Semantic Tube Prediction, or STP. Most models treat token sequences like a chaotic cloud of points. But STP operates on a different premise called the Geodesic Hypothesis. It suggests that high-quality reasoning doesn't just wander aimlessly—it follows locally linear paths along a smooth semantic manifold. By using a JEPA-style regularizer, STP essentially builds a 'tube' around these optimal trajectories, forcing the model’s internal hidden states to stay on track and tune out the statistical noise. The results? We're seeing models reach peak accuracy in math, coding, and logic with a fraction of the training data usually required. And the best part for the architects out there: it does this without the overhead of extra forward passes or complex scaffolding. Is the era of massive, inefficient pre-training coming to an end? Is the future of AI found in the curves of a geodesic path? Today, we’re going inside the 'tube' to find out. #LLMJEPA #JEPA #WorldModels #STP #SemanticTubePrediction https://lnkd.in/gjnfZn6y

1

1 Comment

New preprint: Propagation Fixpoint Depth at the 3-SAT Phase Transition Excited to share the third — and most meaningful — open research output from Entelec AI, an autonomous research system I've been building that independently generates hypotheses, designs experiments, runs analyses, and drafts manuscripts with human oversight and course corrections. This paper tackles a fundamental question in computational complexity: when a SAT solver hits the hardness threshold at α ≈ 4.267, what is it actually doing? We introduce propagation fixpoint depth — the fraction of variables a CDCL solver must explicitly decide before constraint propagation determines the rest — and find a striking pattern across 28,000+ experiments: d/n ≈ 0.26 is a structural invariant of threshold instances — uncorrelated with difficulty, solver-independent Fixpoint depth is threshold-specific: drops 27% from easy to critical regime while decisions explode 100× Structural branching heuristics exploiting this property reduce median decisions by 40–55% The pattern — solvers cycling through shallow contradictions at a stable structural depth — is what we call kinetic confinement: the dynamical counterpart of the static phase transition. The real story here is the research process itself. Entelec AI autonomously explored 22 competing mechanistic hypotheses, ran the experimental pipeline, and identified the fixpoint depth signal. Human judgment directed the research, validated claims against raw data, and made the final calls — but the heavy lifting was AI-driven. Full paper & code: https://lnkd.in/dWm4W8t2 Code: https://lnkd.in/dwW4p23N #AI #Research #SAT #ComputationalComplexity #AutonomousResearch #MachineLearning

5