--- title: "Forecasting at Scale: Lessons from Lyft" newsletter: "MLOps Community" date: 2025-04-17 source: https://aaif.live/newsletters/mlopscommunity/2025-04-17-forecasting-at-scale-lessons-from-lyft --- # Forecasting at Scale: Lessons from Lyft *Plus, blending GenAI with traditional ML, cutting AI's energy use, a Reading Group on Claude's paper on tasks performed with AI, Confessions, and Hidden Gems.* *MLOps Community — Agentic AI Foundation, 2025-04-17* [https://go.mlops.community/GwazeEmail](https://go.mlops.community/GwazeEmail) ## Real-Time Forecasting Faceoff: Time Series vs. DNNs If I knew how to model forecasts with unpredictable, erratic behavior, I’d have cleaned up on the markets these past few weeks. Potentially slightly easier than predicting Donald’s next public malfunction, Josh told me about real-time forecasting at Lyft and why traditional time series models still hold up. With thousands of geohashes updating every minute, autoregressive models beat DNNs on accuracy, retraining speed, and cost. They’re easier to adjust on the fly and more interpretable when things go sideways. Handling retraining is where this really matters: * Autoregressive models can refit in seconds using new data, making them responsive to sudden changes. * DNNs cost far more to retrain and often lag behind when the market shifts. Have a click to listen and become as good at forecasting as Marjorie Taylor Greene. Video [https://go.mlops.community/9ncjyb] || Spotify [https://go.mlops.community/tm1njf] || Apple [https://go.mlops.community/qt44o8] [https://go.mlops.community/qt44o8](https://go.mlops.community/qt44o8) [https://go.mlops.community/9ncjyb](https://go.mlops.community/9ncjyb) [https://go.mlops.community/tm1njf](https://go.mlops.community/tm1njf) ## Agents of Innovation: AI-Powered Product Ideation with Synthetic Consumer Testing When tech moves at different speeds, you get strange overlap - like using a cassette adapter to play a CD in your car. That’s kind of where we are now, blending modern GenAI with traditional ML tooling. Luca explained how his team combines the two to support complex forecasting and ease bottlenecks in analytics workflows. Bayesian models are still a core part of the stack, particularly when interpretability and uncertainty really matter, while LLM agents help speed up model building and communication. One setup runs through LangGraph, coordinating: * Modeling, forecasting, and scenario planning agents * A presentation agent that auto-generates stakeholder-ready decks * Quality-control agents that flag data issues early He’s also developing synthetic users to simulate realistic product feedback and guide early design decisions. Thankfully you won’t need your cassette adapter for this episode – just click below to listen. Video [https://go.mlops.community/xbaafa] || Spotify [https://go.mlops.community/ike6ka] || Apple [https://go.mlops.community/4yvsdu] [https://go.mlops.community/4yvsdu](https://go.mlops.community/4yvsdu) [https://go.mlops.community/xbaafa](https://go.mlops.community/xbaafa) [https://go.mlops.community/ike6ka](https://go.mlops.community/ike6ka) ## Wed Apr 23, 4:00–5:00 PM UTC ## UPCOMING MINI SUMMIT We’re teaming up with Snowflake for a free virtual session on how ML teams are adapting their stack for agents, Iceberg, and modern orchestration. You’ll hear how ZenML handles reproducible pipelines, how Featureform connects MCP to feature workflows, and how Snowflake is approaching enterprise-scale governance without blocking experimentation. Register for free [https://go.mlops.community/SnowflakeEvent] ## Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations Before it was used for VibeCoding, was AI being used for more than just VibeEmailing? This session discussed a paper that analyzed 4 million Claude chats to measure AI’s impact on work, focusing on real usage rather than survey data. They explored how tasks were classified using a hierarchical taxonomy and O*NET mappings, with most usage falling under augmentation, especially in tech roles. One part that sparked debate was how well jobs can be broken into tasks: * Task-level analysis may miss context, especially when tasks overlap or depend on each other * Classification of automation vs augmentation wasn’t always clear-cut Click below to listen, you won't need a paper by Claude to see the impact. Watch here [https://go.mlops.community/qo2l5f] [Watch here](https://go.mlops.community/qo2l5f) ## Job of the Week [https://go.mlops.community/nqnucc](https://go.mlops.community/nqnucc) [https://go.mlops.community/MLConfess](https://go.mlops.community/MLConfess) ## The Terawatt Time Bomb: Transformers, Trouble, and the Analog In-Memory Compute Fix AI's power demands are surging more than Brits putting the kettle on during ad breaks (seriously, Google 'TV pick up'). By 2027, data centers could use 500 TWh annually - more than double 2023 levels - driven not just by training but by inference at scale. Big Tech’s solution has mostly been to build more power-hungry infrastructure, but analog in-memory computing offers a different path. By merging compute and memory, it can cut energy use by up to 20x. EnCharge AI’s approach is particularly interesting: * On-chip analog compute: AI operations happen inside memory arrays, reducing data movement. * Practical and proven: Built on standard CMOS, with a working software stack and five validated chip generations. Paired with a shift to edge computing - processing data locally, not centrally - it could help scale AI without maxing out the grid. So, pop the kettle on and have a read. Read it here [https://go.mlops.community/zh05ey] [Read it here](https://go.mlops.community/zh05ey) ## Hidden Gems ## VIBES Great to feel the love [https://go.mlops.community/ogpga1]! Show your love here [https://go.mlops.community/TrainingDataTee]. ## 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! --- Source: https://aaif.live/newsletters/mlopscommunity/2025-04-17-forecasting-at-scale-lessons-from-lyft