Delphina

Thrilled to have the polymathic Noah Smith (who reaches 300k readers at his substack Noahpinion) on the main stage yesterday at National Association for Business Economics (NABE) TEC! Noah walked us through how improvements in  1/ solar power 2/ batteries  3/ AI  form a triad of extremely complementary innovations. You've probably heard that because of scientific breakthroughs and scale, costs of both solar power and batteries have fallen by an order of magnitude in the last decade, while battery density has doubled in the same time. What you might not have realized is that there’s an incredible feedback loop there: the biggest problem with solar power is it’s only available when it’s sunny. Except that’s not a problem with vastly better battery tech. And more scale in one drives scale in the other which further drives down cost. It’s hard to overstate how big a deal this is. Put that together with recent advances in AI, and we now we have * cheap * portable * clean energy * with a brain. We see real learning curves in all of these areas — it’s only going to get cheaper, more portable, and more intelligent. What can go right.. or go wrong? 🤖🤖🤖

🤖 Beyond the Algorithm: Chiara Farronato (HBS) on Collaborative Decision-Making in Data Science – High Signal Podcast 🌍 Excited to share a clip from High Signal featuring Chiara Farronato (Harvard Business School), where she explores the dynamics between managers and technical teams, sharing pivotal approaches for leaders navigating today’s data-driven landscape. 🎧 In this episode, Chiara sheds light on: 💡 Discretion in Data Science: Effective data science goes beyond algorithms; it’s about aligning on goals and using judgment to meet them collaboratively. Chiara highlights the importance of clear, goal-oriented communication between managers and data scientists. 🚗 Real-World Case from Uber: Chiara shares a fascinating “trip party” example where engineers, data scientists, and product managers come together to evaluate simulated trips—balancing efficiency and customer value, a core challenge in tech-enabled services. 🔄 Communicating Complex Concepts Simply: Using a relatable example from ride-sharing, Chiara explains “treatment and control interference” in an intuitive way, helping managers understand statistical challenges without needing deep technical jargon. Check out this clip, and find the link in the comments to this full conversation and other episodes featuring Michael Jordan, Andrew Gelman, and more coming soon 💫 Produced by Delphina, with Duncan Gilchrist and Jeremy Hermann, High Signal continues to bring you meaningful conversations that empower AI, data science, and tech professionals.

Calling data science leaders - join us tomorrow to discuss DS Leadership with @Uber Science Director @Ali Rauh! https://lnkd.in/gbtxrFj4

💫 Michael Jordan on the Future of AI, Economics, and Intelligent Infrastructure – High Signal Podcast Launch 🌍 I’m thrilled to announce the launch of High Signal, the podcast that brings you the best from the best in data science, AI, and machine learning. 🎧 Our first three episodes are live today! In Episode 1, I talk to Michael Jordan (UC Berkeley) about: 💡 The intersection of AI, economics, and planetary-scale decision-making systems 🔄 Managing uncertainty in intelligent infrastructure 🌐 Building systems that operate at global scales and drive real-world impact Also available now: 🎙️ Conversations with Andrew Gelman (Columbia University) and Chiara Farronato (Harvard Business School). And that’s just the start! Upcoming episodes will feature: 🏆 Guido Imbens (Nobel Laureate), Hilary Mason, Emily Sands, Ramesh Johari, and many more. High Signal is produced by Delphina, with Duncan Gilchrist and Jeremy Hermann. We’re bringing you deep, insightful conversations designed to help you advance your career and make an impact in AI, data science, and machine learning. 🎬 Watch the clip now, and check the comments for links to all three full episodes on our landing page.

Netflix is on AWS. In 2024, that seems obviously the right call. But how was it obvious for the N in FAANG — a company with virtually limitless engineering resources — when they made the move back in 2008? At last week’s Data Dialog, Savin Goyal, cofounder of Outerbounds (and formerly ML platform leader at Netflix) gave us a masterclass in platform decisions. Savin walked us through build vs buy in the framework of the economic theory of comparative advantage. The idea of comparative advantage is that even if you are literally better at *everything*, you still have limited time and resources, and so will ultimately do better by focusing on what you are best at *relative* to others. In other words, build the thing that translates into fundamental advantage. Buy the rest. For Netflix, nearly 20 years ago, that meant recognizing that having its own data centers didn’t make any sense. I’m sure that was an incredibly controversial and hard decision. How many hundreds of jobs, of careers, did that change? It made made me reflect -- in how many places in your own company (and life) are you being honest about what you are really best at, and focusing on that? ------ If this sounds interesting and you’re a data science leader: Data Dialogs are an off-the-record forum to get into the details of data tech, strategy, and people. Our next Dialog is coming up on Tuesday Oct 29 at 9am PT with Ali Rauh, Director of Science at Uber, and Hugo Bowne-Anderson where we’ll talk * OKRs, OECs, and OFs (what does all that even mean?) * Long term value and experimentation * The DS team structures that work (and don’t) Join us! Link in comments below.

What does it take to make devs truly productive in the non-deterministic, infra-heavy, cross-functional tangle that is AI/ML? Join us at this week’s Data Dialogues with Savin Goyal to find out. I first met Savin a few years back at a VC dinner and was blown away by his practicality and depth on questions like these. He is CTO and Co-founder of Outerbounds, and previously co-led machine learning infrastructure at Netflix and built Metaflow, a platform now used by hundreds of large companies to help scale their AI/ML systems. If you’re a data leader seeking to make your team more effective, don’t miss it. Thursday morning, 9am PT, hosted by Hugo Bowne-Anderson and yours truly. Apply at the link:  https://lnkd.in/gRP7ZBwu

One of our star engineers Thomas Barthelemy recently led a brown-bag on a topic everyone knows they need to do more of, but nobody feels like they’re doing right: automated testing. Thomas previously built the data infra and backend at Coursera and helped the company scale from 40 to over 1,000. We had an awesome discussion — articulating what’s wrong with most approaches and developing our own testing philosophy here at Delphina. Instead of being dogmatic, Thomas worked us through a principled ROI-based approach to testing.  A few (potentially controversial) takeaways: 1. The traditional “testing pyramid” is a lie. Long live the testing trophy, which emphasizes static and integration tests with strategically limited unit and end-to-end tests. 2. The right goal isn’t 100% test coverage. Some lines are just not that important to your business. 3. Tests that last a long time are vastly more valuable. A test does not guard against regressions if the next time the code is touched, the test needs to be deleted and rewritten. This stuff is so important to get right, but it’s not easy — especially with LLMs suddenly making things non-deterministic. Check out the article for a step-by-step breakdown and a peek behind the curtain at how we’re tackling automated testing at Delphina. #softwarengineering #datascience #artificialintelligence

Thrilled for Delphina to be featured by Felicis Ventures in their latest map of the Agent Economy, in the Data Science horizontal. When we were first pitching seed investors over a year ago, more than a few looked at us with a quizzical look when we explained we were building an agent for data science. Needless to say, it's fun to be part of the fast emerging world of agents. Looking at the market map, it's inspiring to see the number and variety of agent startups tackling a range of horizontals and verticals, both large and small; and it's hard not to get excited about the unlock for society as AI is able to take on mundane, formulaic work across these areas. Onward! 🚀

Uber's 5,000 production AI/ML models make a whopping 16 million predictions per second at peak. If peak lasted all day, that would come out to 1.4 TRILLION predictions a day! 🤯 We learned how they got there, how that's even possible, and much more when Distinguished Engineer Min Cai joined us last week at the Data Dialogs to discuss Uber’s AI/ML journey and the many learnings along the way. A few key insights: 1/ Uber’s models were originally tree-based (think xgboost) and have graduated over time to deep learning. Yet even today at Uber scale and maturity, with a tech team of thousands and close to $200B in annual bookings, 40% of Tier-1 models are trees. 2/ They make significant investment and see high ROI in hygiene around making sure they are reliably training, serving, and maintaining their models. Often not the most inspiring of work, but super important. 2/ On the AI side, Uber starts with the best commercial models from OpenAI and Anthropic before moving to fine-tuning open-source models for specialized applications (just like the rest of us!). The next Data Dialog is coming up next Friday with Victor Kostyuk, cofounder OfferFit, where we'll get real about the latest in reinforcement learning and marketing DS. In my experience, RL is one of the most fraught topics in ML — so conceptually exciting and incredibly high value in certain cases, yet so complex and easily overhyped. Can’t wait for this. If you’re a data science leader interested in these topics, grab the link in the comments to apply! #datascience #artificialintelligence #marketing #leadership

I’m thrilled to be doing a fireside chat with Noah Smith of Noahpinion (250k+ substack readers!) at the National Association for Business Economics (NABE) TEC conference in a month. Noah's got a lot to say about everything from techno-optimism and patriotism; to trade wars and immigration; even living in Japan and rabbits as pets. He's so prolific he recently won Tyler Cowen’s “writing every day award” — with 11 articles in the first two weeks of September alone. (Not exactly, No Opinion.) INPUT NEEDED: I need your help. What should I ask him about? What's something exciting / interesting / polarizing that Noah's said recently? Drop ideas below; I want to make sure we talk about the meatiest topics. While we're here, quick plug for NABE TEC — it’s THE conference for economists in data science and tech. In Seattle this year, with a studded speaker list including Elizabeth Stone (CTO of Netflix) and Melissa Dell (2020 John Bates Clark Winner). Check out the full list and sign up at the link in the comments. #datascience #economics #education #leadership