--- title: "Founder FOMO & the art of picking winners" newsletter: "MLOps Community" date: 2025-03-13 source: https://aaif.live/newsletters/mlopscommunity/2025-03-13-founder-fomo-the-art-of-picking-winners --- # Founder FOMO & the art of picking winners *Plus, Scaling ML, debugging RAG, and making AI more steerable.* *MLOps Community — Agentic AI Foundation, 2025-03-13* We're putting together a GPU buyer's guide reviewing pricing, capacity, and value props for different providers. So far, we've been comparing prices on Pump.co, Modal.com, Lambda Labs, and AWS (GPUs and Tranium/Inferencia). Reply back to this email with the GPU provider you want us to review next. ## Future of Software, Agents in the Enterprise, and Inception Stage Company Building Casablanca, The Great Gatsby, Titanic - classic stories of the one they let go. Eliot has a few of his own after ignoring his gut on certain founders and missing out on big opportunities. He breaks down what separates great founders, why strong product opinions matter, and how AI agents are shifting from unpredictable experiments to structured automation. * Reliability matters – Multi-agent systems are evolving from solo copilots to coordinated teams, improving accuracy. * Testing upfront – AI evaluation frameworks are ensuring agents perform as expected before deployment. * Smart automation wins – Founders who prioritize control over hype are setting the standard. Be sure to listen, so you know how to make room on the door for Jack when the time comes. Video [https://go.mlops.community/q1n4jt] || Spotify [https://go.mlops.community/lh3bo1] || Apple [https://go.mlops.community/1y8ztc] ## Kubernetes, AI Gateways, and the Future of MLOps Like a tree falling unseen making no noise, the modern version might be, if you make a brilliant app but no one sees it, does it exist? From the importance of evangelizing your work to keeping a "brag document," Alexa shared a lot in this chat, including early struggles with Airflow. Running Airflow manually meant constant crashes and scaling headaches, but now tools like Kubeflow, Argo Workflows, and KServe offer more flexibility. Airflow had its challenges: * Scaling issues: Frequent scheduler crashes under heavy workloads. * Operational overhead: Teams had to maintain forks and fine-tune configurations. * Few alternatives: Airflow was often used by default, even when not ideal. We also covered bridging infra and ML teams, open source, and platform usability. This episode exists, so be sure to hear it. Video [https://go.mlops.community/65heaz] || Spotify [https://go.mlops.community/dze4hj] || Apple [https://go.mlops.community/ajxe45] Airflow giving you headaches? Check out DataCamp [https://go.mlops.community/DataCamp]* for Airflow courses that can help. *This is an affiliate link - clicking helps support the newsletter. ## PODCASTThe Unbearable Lightness of Data Repairing the ozone layer, mapping the human genome, eradicating smallpox, parsing PDFs – some of humanity’s biggest challenges. And yet, one remains frustratingly unsolved. That’s just a small gripe Rohit has with reasoning models. The bigger issue? They still lack steerability. Instead of passively receiving outputs, users should be able to refine prompts mid-process and guide AI dynamically. Rohit envisions a more interactive approach, including: * Checkpointing thought processes – Pause AI mid-reasoning to adjust focus. * Guided refinement – Click into reasoning steps for deeper exploration. * Human-AI balance – AI suggests paths, users fine-tune direction. With these changes, reasoning models could feel less like black boxes and more like real collaborators. Now, let’s see if I can steer you into clicking below to listen. Video [https://go.mlops.community/94c059] || Spotify [https://go.mlops.community/8ftgn0] || Apple [https://go.mlops.community/zdll1k] ## Job of the Week [https://go.mlops.community/6a0tyd](https://go.mlops.community/6a0tyd) ## READING GROUPDeepSeek That, DeepSeek This I think, therefore I am... DeepSeek? This session of the Reading Group broke down DeepSeek R1, exploring its reinforcement learning approach to reasoning. Scaling alone isn’t enough, so DeepSeek optimized PPO with GRPO and distilled reasoning into smaller models. The team debated whether RL or distillation was more effective and how to evaluate reasoning quality. One challenge was verifying how models think, not just what they output: * Internal iteration lets models refine answers before responding. * Distillation vs. RL – which produces better generalization? The group also raised MLOps questions on tooling, collaboration, and reproducibility. I think, therefore I am… going to watch! Watch it here [https://go.mlops.community/yqwv9m] [Watch it here](https://go.mlops.community/yqwv9m) Bond might look suave in his suits, but real agents operate in rags. Agentic RAGs that is. The article demonstrates how SmolAgents and a multimodal vector database can be used to build a research paper retrieval system, offering: * Multimodal support – retrieveing text, images, videos, and metadata. * Graph-based metadata handling – enabling knowledge graph-style retrieval. * Efficient query processing – on-the-fly adjustments to data formats and indexing. The article provides a step-by-step guide to implementing this system, covering document extraction, chunking, embedding storage, and dynamic query refinement using SmolAgents. "Do you expect me to retrieve?" "No, Mr. Smol, I expect you to RAG!" Read it here [https://go.mlops.community/appvbh] Gaming, heat source, coffee tray - your laptop does a lot, but when it comes to scaling ML training, you’ll need more firepower. Scaling machine learning training requires distributing workloads across multiple devices. This article explores key strategies, including: * Parallelism Methods: Data parallelism (splitting data across devices), tensor parallelism (dividing tensors across GPUs), and pipeline parallelism (distributing model layers). * Infrastructure Choices: From local multi-GPU setups to Kubernetes-based orchestration (PyTorchJob, MPIJob) and vendor solutions like SageMaker and Databricks. It also covers fault tolerance strategies like checkpointing and elastic scaling to prevent failures from disrupting training. The guide provides practical steps for optimizing distributed ML workflows. Make good use of your laptop and click below to read. Read it here [https://go.mlops.community/3aehej] Part 3 of the AI blog that makes you hungry looks at keeping your data kitchen organized - or as The Bear would say, "clean as you go." Dataplex structures and manages data across lakes and zones, improving organization and governance in ETL workflows. It automates data discovery, tracks lineage, and integrates with BigQuery and Airflow. One of its key functions is organizing data into structured zones: * Logical segmentation: Separates raw, cleansed, and transformed data for clearer workflows. * Automated discovery: Identifies datasets and tracks freshness without manual effort. * Governance controls: Helps enforce policies for access and compliance. This makes it easier to track, manage, and document datasets. Grab a snack and click below to read. Read it here [https://go.mlops.community/qjy8rw] ## HERE TO HELP Before you go, here are three ways I can help - just hit reply: * Curated intros to other community members * What problems are you dealing with? Let me help you find the best solutions through my network * Looking to augment your staff for an MLOps or AI project? I got you covered Thanks for reading, catch you next time! Interested in partnering with us? Get in touch: partners@mlops.community Thanks for reading. See you in Slack [https://go.mlops.community/slack], YouTube [https://www.youtube.com/channel/UCG6qpjVnBTTT8wLGBygANOQ?view_as=subscriber], and podcast [https://home.mlops.community/public/content/] land. Oh yeah, and we are also on X [https://twitter.com/mlopscommunity] and LinkedIn [https://go.mlops.community/linkedin]. The MLOps Community newsletter is edited by Jessica Rudd [https://www.linkedin.com/in/jmrudd/]. --- Source: https://aaif.live/newsletters/mlopscommunity/2025-03-13-founder-fomo-the-art-of-picking-winners