Karolis Urbonas

Tuning Spark, Ray, or Dask for ML isn’t one-size-fits-all. In a new Domino Blueprint, Principal Solution Architect Sameer Wadkar shows how API-driven scaling complements autoscaling to right-size clusters on demand—cutting idle spend and speeding bursts for GPU/I/O-heavy work. Practical steps, a lightweight Python client, and Helm install included. If you care about FinOps and MLOps, this is for you. Read the blueprint: https://gag.gl/WwWso7

1

China's dominance in open-source AI is raising concerns for U.S. competitiveness, according to Andreessen Horowitz partner Anjney Midha. DeepSeek's breakthrough R1 model, built at a fraction of the cost of U.S. frontier models while matching their performance, exemplifies this shift. The Chinese startup's latest innovation processes information as visual tokens rather than language tokens, making models 10 times more efficient. Meanwhile, Sequoia Capital announced $950 million in new early-stage funds, targeting Series A startups with a $750 million fund and seed companies with $200 million. The legendary VC firm is doubling down on early investments as AI valuations skyrocket, with recent bets on Clay, Harvey, and Sierra appreciating significantly. In the semiconductor space, AMD secured a major AI chip order from Oracle as it races to compete with Nvidia's dominance. The developments highlight the intensifying global competition in AI infrastructure and the critical importance of early-stage positioning in the rapidly evolving tech landscape.

1

Attending Snowflake BUILD 2025 was such an incredible experience. It’s amazing to see how Snowflake is re-imagining the future of AI and data. Here were my top 5 takeaways from the conference: 1️⃣Building the agentic future: AI agents moving from concept to production. 2️⃣Cortex AISQL: natural language queries, no prompt engineering needed. 3️⃣ Cortex Agents: from queries to multi-step reasoning and automation. 4️⃣ Managed MCP Server: bridging AI tools through standardized connectivity. 5️⃣ Gen2 Warehouses: a data foundation built for real-time, enterprise-scale AI. Watch the full breakdown in my video 🎥 And if you’d like to learn more about Snowflake Intelligence, you can check out the link here: https://lnkd.in/gnRkJPv2 #SnowflakePartner #SnowflakeBUILD

303

31 Comments

Proud to see Imandra featured as a leader in automated reasoning in The Wall Street Journal alongside Amazon's neurosymbolic AI efforts. As Grant noted in the piece: pure neural networks "can deliver even extraordinarily wrong answers with irrational confidence" - a problem that can't be solved by just adding more compute power. This is exactly why, at Imandra, we've spent years building the foundations for AI that can reason, validate and scale reasonably. Building AI agents or LLM frameworks? Reach out - we'd love to discuss how neurosymbolic approaches could strengthen your systems.

17

Had a great time chatting with Marko Rukavina from Tesseract Analytics about one of my favorite topics; dbt Labs Every team running dbt knows the struggle: endless tweaking, model tuning, trying to shave seconds off runs that somehow still cost more than they should. At Yuki, we’ve been obsessed with solving that pain. Instead of optimizing after the fact, we let Snowflake optimize itself on the fly. And I’m really proud of something our team, led by Alex Mirkin, just rolled out A new Gantt visualization that helps dbt users see exactly where their jobs bottleneck. It’s not just about speed; it’s about understanding how performance connects directly to cost. Big thanks to Marko for the conversation and for spotlighting what automation in data engineering really looks like today. 🎧 Watch the short clip below 👇 #Snowflake #dbt #FinOps #DataEngineering #Automation #CostOptimization #Yuki

23

1 Comment

What are the most underrated Databricks capabilities? Watch my conversation with Shiv Trisal (Global Industrials GTM @ Databricks) on the topic below. For the full, practical, no-fluff discussion on real-world data + AI: https://lnkd.in/gMPsFHbY

89

Modern workplace is a different place. Fresh thought is needed on training students so that they form modern AI ready workforce. Mohamed Imran K R shares his insights as subject matter expert

5

Are we focusing on the right aspects of generative AI? I ask myself this, and you should too While much of the discussion focuses on efficiency and automation, the most critical challenges for 2025 and beyond are emerging quietly amid the hype. After analyzing the current professional discourse, I’ve identified several key themes that are noticeably absent, summarized in the infographic titled "The Silent Shifts." These challenges extend beyond technical issues; they raise foundational questions about the future of work, collaboration, and human cognition: - The "Day-Two" Reality: As we build complex multi-agent ecosystems, are we prepared for the operational chaos that follows deployment? - The Social Gap: Can an agent genuinely be a "team member" without grasping the nuanced social dynamics and trust that characterize human collaboration? - The Cognitive Comfort Trap: In our eagerness to offload tasks, are we unintentionally outsourcing the very skills, if not intuition, strategy and creativity that make us irreplaceable? Take a look at the visual below. Which of these "silent shifts" do you believe will pose the biggest challenge for organizations in the coming years? Let's discuss in the comments.

2

A couple of months ago I took Tim Januschowski's Modern Forecasting in Practice course on Maven. Tim led the time series science organization at AWS AI, where his teams built SageMaker Forecast and other AWS forecasting services. One idea from the course kept pulling at me: foundation models that forecast diverse time series with zero domain-specific tuning. No feature engineering. No custom architecture. I kept thinking: this solves the problem every MMM practitioner runs into. We build causal models that explain sales well. Adstock effects, saturation curves, channel-level ROI. Then someone asks: what happens next quarter if we shift budget from TV to search? And the model can't answer, because it needs future employment rates and temperature data. Values that don't exist yet. Most teams handle this with naive estimates or by dropping controls entirely. Both options undermine the model they spent months calibrating. So I put together a proof of concept to test whether foundation models could close that gap. The idea: train an MMM on historical data, use Chronos2 to forecast the control variables, then feed those forecasts into the MMM for scenario planning. In our simulation, the added prediction error from using forecasted controls instead of actuals was minimal. The error propagation math also turned out to be straightforward, which means practitioners can estimate upfront whether their control forecasts are accurate enough to be useful. So it's not causal models versus foundation models. MMMs encode what causes what. Foundation models predict what hasn't happened yet. Combine them and you get a causally grounded scenario planning system with quantified uncertainty on the key assumptions. Full methodology, reproducible notebook, and results in our new blog post: https://lnkd.in/ee3gVHxW And if you want to go deep on modern forecasting, highly recommend Tim's course: https://lnkd.in/ezZQ4uhU

28

2 Comments

Introducing a new architecture of hybrid AI. ​ We (our team: Kamil Gadeev, Bahtiayr Usupov and me) are used to evaluating AI models by the number of parameters and the amount of data. But what if the next breakthrough lies in the plane of physics? ​Modern LLMs, despite their power, remain fragile. Their “thinking” processes are a black box, and hallucinations are a system failure. We offer a new perspective: coherence (internal order) in a neural network is not a given, but a precious, energy-consuming resource. ​In our new article, we present NFCS (Neural Field Control System) - a hybrid cognitive architecture built on the principle of “costly coherence”. ​What does this mean? Our system does not just process information, but actively fights to maintain internal order, solving the optimization problem: where, when and how much “energy” to spend on synchronizing its cognitive modules. ​This is not a static model. With its built-in “evolutionary pressure,” NFCS is capable of self-improvement and adaptation, generating new stable patterns (“Pulsars”) in response to internal contradictions. This is a step toward creating true “life in silico.” ​We propose a framework for creating not just powerful, but truly adaptive, resilient, and, critically, safe intelligent systems. ​We invite experts from the fields of AI, neuroscience, and theoretical physics to join the discussion. ​ ​#AI #CognitiveArchitecture #LLM #Neuroscience #TheoreticalPhysics #AISafety #MachineLearning #HybridAI

22

20 Comments

Building on top of large language models is fun, but getting consistent performance and reliability is extremely challenging. I love this post from Kimberly Patron about how Sierra uses health and performance-based traffic routing and request hedging to get meaningfully better performance and tail latency from high scale LLMs https://lnkd.in/g4qWRCb5

284

13 Comments