Nikita Rokotyan

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

The Open Geospatial Consortium (OGC) has officially published the SpatioTemporal Asset Catalog (STAC) Community Standards, a widely adopted framework for standardizing #geospatial asset #metadata Originally developed for satellite imagery, STAC now supports a broad range of data resources, including aircraft and drone imagery, #LiDAR, point clouds, DEMs, SAR data, vector data, and composites. The specification provides a minimal core with flexible extensions, enabling both data providers and users to work with a standard format, reducing the need for proprietary solutions. STAC allows data to be organized as Catalogs, Collections, and Items, supporting both static browsing and more complex queries through the STAC API. Its design aligns with OGC APIs and widely used web standards, ensuring interoperability across platforms and applications. The STAC Community Standards are freely available and represent another step forward in making geospatial data more accessible, interoperable, and actionable. Read more: https://lnkd.in/gE3rEY6K

8

Exciting news from AgileRL: Introducing Evolvable IPPO! 🚀 We're thrilled to announce that Evolvable Independent Proximal Policy Optimisation (IPPO) is now available on both the AgileRL open-source framework and the Arena platform! What makes the IPPO algorithm special? IPPO is a multi-agent reinforcement learning algorithm primarily used in scenarios involving independent, cooperative, or competitive multi-agent systems. IPPO is particularly effective when each agent has its own distinct role or objective. Why this matters: - Train multiple agents to develop identical behaviours without the overhead of individual training - Dramatically faster environment exploration as agents share learning - Simplified approach to complex multi-agent reinforcement learning problems Real-world applications: - Gaming: Create smarter NPC teams that coordinate naturally in multiplayer environments - Warehouse robotics: Multiple robots learning to navigate and fulfil orders efficiently together - Traffic management: Coordinate multiple traffic lights to optimise flow across entire city grids - Supply chain: Optimise fleets of delivery vehicles that share the same objectives - And many more… Having Evolvable IPPO on Arena doesn’t just make training complex team-based AI projects more accessible for engineers at all levels; it also significantly reduces training time and dramatically increases model performance. This release represents our ongoing commitment to advancing cutting-edge reinforcement learning techniques while making them accessible to practitioners across industries. Try evolvable IPPO today: - Integrate it into your projects with our open-source framework - https://lnkd.in/ewCntTUu - Experiment with it on our Arena platform with zero setup required - https://arena.agilerl.com/ #agilerl #reinforcementlearning #machinelearning #ippo #multiagentrl #ai

5

Analyzing large-scale profiling data is easier thanks to new work from Alexandr Kalinin, Ph.D., Gregory Way (University of Colorado), Shantanu Singh, and others. To improve upon existing methods that make assumptions about distribution, linearity, or sample size of profiling data, they developed a statistical framework and open-source software for information-retrieval-based assessment of profile strength and similarity. Employing mean average precision as an evaluation metric, they demonstrated its utility on simulated and real-world datasets for image, protein, and mRNA profiles. Their versatile framework could help streamline hypothesis generation and improve hit prioritization from a range of high-dimensional biological profiling datasets. Read more in Nature Communications. 🔗: https://lnkd.in/eg7rqZ9D #BroadInstitute #Science #ScienceNews #Research #ScientificResearch

37

🎙️ New #ACMByteCast Alert! ACM Fellow Michael J. Freedman talks with Rashmi Mohan about building Coral CDN, co-founding Timescale, and turning research into real-world impact. 💡 From peer-to-peer systems to databases for real-time analytics—don’t miss this deep dive into scalable tech. Listen now here or wherever you get your podcast: https://buff.ly/R2fFGLJ #DistributedSystems #TechInnovation

6

🧠💬 Memory in LLMs A practical guide showing how to implement conversational memory in LLMs using LangGraph, demonstrated through a therapy chatbot. Features code examples for basic retention, trimming, and summarization approaches. Learn to build memory-aware apps 👉 https://lnkd.in/gybcrV5v

967

21 Comments

Stanford Center for Research on Foundational Model's Marin project has released the first fully open model in JAX. It’s an 'open lab' sharing the entire research process - including code, data, and logs, to enable reproducibility and further innovation. Check out the project: https://goo.gle/44AMeLY

41

2 Comments

We launched Tonic Datasets to power your AI training and evaluation. Now you can acquire high-fidelity, specialized synthetic data for training, testing, and performance evaluations. In our new blog, we explore: --> Why synthetic data, a tech-first solution, is the most scalable layer in today’s data landscape dominated by services-first companies like Scale, Mercor, and Turing.  --> How we deliver fidelity, flexibility, and compliance by design. --> Real-world examples from AI teams building with bespoke datasets. Read the full blog here: https://lnkd.in/ebAUE4a7   #syntheticdata #AI #machinelearning #LLM #dataset #dataprivacy #AITools

40

1 Comment