Saturdata | Podcast on RSS.com

Saturdata

by Saturdata Podcast

Careers Technology

Season 2

We're making up AI as we go | Saturdata with Joey Yudelson

What happens when you train an evil AI and it just lies really confidently? Joey Yudelson (https://www.linkedin.com/in/joseph-yudelson/), AI safety researcher at Redwood Research, joins Sam and Shifra to break down why 300 people standing between us and a catastrophic AI future might not be enough, and what data folks can actually do about it. We talk about: - Taxonomies of AI risk (silly vs. not silly, yes this is a real framework) - Why the evil AI wrote 30 paragraphs insisting its buggy code was perfect - Global AI regulation and who's actually doing a good job (hint: it's the EU) - How to use Claude agents like a multiplayer cheat code - Why you personally could make a dent in AI safety research Follow Saturdata, your favorite weekend data podcast Spotify: https://open.spotify.com/show/5QolhKm1jDZzVuHO0S9ZBo?si=910efb23833f4fc1 LinkedIn: https://www.linkedin.com/company/saturdata Instagram: @SaturdataPod #Saturdata #AISafety #DataScience #MachineLearning Chapters: 0:00 - Intro 0:57 - Joey's origin story: from high school Yudkowsky reader to full-time AI safety researcher 3:42 - A guided tour of the AI safety landscape 6:14 - Where Joey fits in the puzzle: model organisms and misalignment research 7:28 - The evil AI that wrote 30 paragraphs insisting its buggy code was perfect 10:55 - Deep in the lab vs. everyday AI user: how different are they really? 13:42 - The knowledge lag: why comedians are still calling AI "smart autocomplete" 17:24 - Taxonomies of risk: silly vs. not silly (yes, water use is on the table) 22:02 - Being a responsible AI user: what data folks can actually do 28:32 - How LLMs actually work, explained with a very talented dog named Jeeves 33:11 - Joey's lifelong vendetta against SQL (and how he gets away with it) 36:59 - Three rules for getting real value out of AI agents without losing your mind 42:48 - Why you personally could make a dent in AI safety (and the case for Talmudic AI research) 45:20 - Takeaways and outro

From spaghetti to clean code: pandas, Polars and DuckDB explained | Saturdata

Is your Python code held together with duct tape and prayers? Sam and Shifra untangle the spaghetti and walk you through what it actually means to write clean, maintainable data code, and which tools will get you there. From the humble origins of Pandas to the blazing speed of Polars and the SQL simplicity of DuckDB, this episode is your guide to leveling up without burning down your codebase. 🌊 Check out the deep dive here: https://youtu.be/htGazioOVvM We talk about: - What spaghetti code actually is (and why we've all written it) - The real limitations of Pandas at scale (single threading, row storage, and bloated data types) - One-line bolt-on fixes with PyArrow and NVIDIA RAPIDS cuDF - Why Polars feels like the dplyr of Python - What makes DuckDB the SQLite for analytics - Polars vs DuckDB: how to pick the right tool for your team - Future you is a different person, and other habits of engineers who sleep at night Follow Saturdata, your favorite weekend data podcast: Spotify: https://open.spotify.com/show/5QolhKm1jDZzVuHO0S9ZBo?si=910efb23833f4fc1 LinkedIn: https://www.linkedin.com/company/saturdata Instagram: @SaturdataPod #Saturdata #Pandas #Polars #DuckDB #DataEngineering Chapters: 0:00 - Intro 0:47 - The spaghetti code confession 3:23 - All the pasta shapes of bad code 5:14 - Trial by fire: how you actually learn to write good code 8:34 - The pandas origin story 12:19 - What's wrong with pandas (we still love you though) 17:48 - The PyArrow bolt-on: a one-line glow-up 21:51 - GPU-powered dataframes with RAPIDS cuDF 25:15 - Running out of RAM and spilling the tea 31:22 - Enter Polars: the polar bear to pandas' panda 42:14 - DuckDB: the cute duck who does SQL fast 50:24 - So which tool should you actually use? 56:46 - Future you is a different person: tips for writing better code 59:16 - Comment the why, not the what

Data storytelling: the good, the bad, and the pie chart | Saturdata

Your chart is full of information. So why does no one know what it means? Sam and Shifra break down everything you need to know about data visualization and storytelling, from picking the right Python library to presenting charts your exec will actually understand. Spoiler: the pie chart doesn't make it out alive. 🌊 We talk about: Matplotlib vs Seaborn vs Plotly (and when to use each) Declarative vs imperative frameworks and why it matters Chart types for EDA: scatterplots, heatmaps, box plots, and pair plots Color psychology, colorblindness, and pretty privilege for data Why pie charts get so much hate The right chart for the right people (Sam's trifecta) Chart titles, KISS, and presenting to executives Follow Saturdata, your favorite weekend data podcast! Chapters: 0:00 - Charts are communication, not decoration 3:03 - Your Python viz toolkit: Matplotlib, Seaborn, and Plotly 8:15 - Why Seaborn is the beginner's best friend 10:43 - Polars vs. Pandas: Know what your chart is actually for 15:20 - Declarative vs. imperative frameworks (and why it matters) 20:29 - The chart type lineup: Scatterplots, heatmaps, and box plots 28:36 - Pair plots: The boss of all plots 31:31 - Pretty privilege for data: Color, accessibility, and design language 40:22 - Why pie charts are everyone's villain (for good reason) 41:24 - The trifecta: Right people, right info, right chart 46:47 - KISS, chart titles, and presenting to executives 51:49 - Building dashboards people actually use

Statistics 101 at work | Saturdata

What if your A/B test needed 67 years to reach statistical significance? Sam found out the hard way. Join Sam and Shifra as they demystify statistical testing for the real world of data work, where the stakes are lower, the data is messier, and your stakeholders definitely do not know what a p-value is. We talk about: P-values, null hypotheses, and why 0.05 was basically made up Type 1 and type 2 errors through the lens of job interviews When A/B testing actually makes sense (hint: you need more than 10 visitors a day) T-tests, chi-square, ANOVA, and F1 scores explained without the jargon Why a suspiciously high model accuracy is actually a red flag The difference between statistical significance and practical significance Chapters: 0:00 - The 67-year A/B test 0:22 - Welcome to everyone's favorite hobby 1:37 - Knowing how to interpret tests (not run them) 2:27 - Is the analysis actually important to the business? 3:37 - P-values refresher: what they are and aren't telling you 6:07 - Why a raw p-value isn't enough 7:40 - Null vs. alternative hypotheses explained 10:16 - Type one and type two errors (a.k.a. the costly mix-ups) 15:06 - Lift: measuring if your marketing actually did anything 18:53 - When you already have all the data, statistics isn't the tool 20:57 - Sample size, statistical significance, and the 67-year problem revisited 24:04 - Common A/B test types: t-tests, chi-square, and ANOVAs 26:44 - F1 scores, confusion matrices, and picking the right metric 29:19 - Central limit theorem and the magic number 30 31:31 - We never prove things — we just reject the null 34:51 - Premortems and deciding if a project is even worth doing 35:52 - When n is too small vs. too big (and why both are a problem) 38:00 - Effect size: the stat that doesn't care how big your sample is 41:39 - Regression, slope, and explaining it to real humans 47:07 - Spend your time on the right things, not the fanciest model 52:33 - Wrap-up and big takeaways

Why your SQL costs more than you think | Saturdata

Think you know SQL? Sam and Shifra break down what separates a query writer from a true data thinker, from basic selects all the way to distributed systems, query plans, and the four pillars of production-ready code. Plus: why your data provider's incentives are working against you, how a 1,400-line monolith hid millions in overstated revenue, and the one approach that will save you from silent, soul-crushing data failures. 🌊 Check out the deep dive here: https://youtu.be/ayNKmcIELEo We talk about: - Three levels of SQL (and a secret fourth) - What they don't teach you in school: data fluency, granularity, and audit patterns - Why SELECT DISTINCT and ORDER BY are secretly expensive - WORM vs. WARO: taking cost at write time vs. read time - Idempotency, query plans, and writing SQL like a query engine Follow Saturdata, your favorite weekend data podcast: Spotify: https://open.spotify.com/show/5QolhKm1jDZzVuHO0S9ZBo?si=910efb23833f4fc1 LinkedIn: https://www.linkedin.com/company/saturdata Instagram: @SaturdataPod #Saturdata #SQL #DataEngineering #DataAnalytics #QueryOptimization Chapters: 0:00 - Intro: Squeal and Sasquatch 2:34 - The three levels of SQL: Basic, intermediate, and advanced 6:47 - When advanced SQL isn't really SQL anymore 9:12 - The 1,400-line monolith horror story 13:00 - The four "ilities": Readability, maintainability, observability, and explainability 15:34 - Write once, read many: Taking the cost at write time vs. read time 18:04 - SQL as analyst vs. SQL as architect 22:52 - Platform-dependent code and why your cloud provider is not your friend 25:31 - What school never taught you: Data fluency, granularity, and knowing your tables 35:09 - Why all of this matters: Defending your query like a lawyer 41:49 - What's holding people back: Select star, distinct, and the monolith trap 46:30 - Idempotency, query plans, and thinking like an engine

Data skills nobody taught you | Saturdata

Your SQL is great. But can you actually ship? Sam and Shifra kick off Season 1 (since Saturdata is zero-indexed) by covering the unsung skills that separate someone who writes queries from someone who builds things: terminal literacy, dependency management, Git, notebooks, and why UV might be Python's best friend right now. Plus, a deep dive into Marimo, the notebook tool that fixes everything you hate about Jupyter. 🌊 Check out the deep dive here: https://youtu.be/SWcLulIhVkg We talk about: - Terminal basics and why middle-school-level literacy is all you need - Virtual environments, dependency conflicts, and why UV is the move - Git fundamentals: staging, committing, and why humans still need to be in the loop - What's wrong with Jupyter notebooks (and who's fault it is) - Marimo: reactive notebooks, SQL cells, and dashboards all in one Chapters: 0:00 - Welcome back (and why season two is called season one) 0:47 - Auxiliary data skills: the glue holding everything together 2:08 - How this season works: yap episodes + deep dives 4:01 - What auxiliary skills actually are (bash, docker, makefiles, oh my) 5:43 - Terminal literacy: your middle school diploma starts here 10:12 - Finding your terminal on Mac, Windows, and the WSL escape hatch 11:52 - Essential commands every data person should know 15:46 - Notebooks: an educational tool that accidentally went to production 19:55 - The biggest problems with traditional Jupyter notebooks 23:37 - Virtual environments and dependency conflicts, explained 26:30 - UV: the one tool to rule them all 30:24 - Git: version control, branching, and why humans still need to be in the loop 38:15 - Marimo: the next-generation notebook that fixes everything 47:28 - Closing thoughts

AI regulation isn't just a tech problem, it's a people problem

It's not all firewalls and technical fixes. Here's why shaping the future of AI comes down to soft skills, cultural awareness, and actually showing up for the conversation. Call your rep, use your voice, because any regulation is better than none 🗣️ #shorts #saturdata #data #AIregulation #AIgovernance #TechPolicy #futureofAI

Stop coding, start directing: the AI shift you can't ignore

Claude Opus is probably a better coder than you, and Joey isn't sugarcoating it. Instead of writing code line by line, the real move could be writing design docs and letting a fleet of AI agents do the heavy lifting. The biggest mistake is assuming AI will always look the way it does right now #shorts #saturdata #data #AI #claudeai #futureofwork #techtrends

Your AI chatbot is basically a very well-trained dog named Jeeves 🐾

Ever wondered how ChatGPT actually works? Joey breaks it down in the most hilarious way possible, and honestly, we'll never think about reinforcement learning the same way again. Train it right and it fetches you a beer. Train it wrong and it wants to pee on your friend Dave's foot. The science checks out. 🍺🤣 #reels #saturdata #data #AI #machinelearning #ChatGPT #techexplained

How ChatGPT actually works: a dog explains | Saturdata with Joey Yudelson

What if you could explain ChatGPT using only a dog, some audiobooks, and a stick? Joey Yudelson joins Sam and Shifra to break down how large language models actually work, no PhD required. From next-word prediction to reinforcement learning, this one will make you feel like you actually get it. We talk about: - How LLMs learn to "speak human" by predicting the next word - What reinforcement learning actually does (and why your model needs a stick sometimes) - Why RLHF is basically dog training at a bazillion scale - Spurious correlations and how models learn the wrong lessons - What it really means when an AI "has a persona" Follow Saturdata, your favorite weekend data podcast: Spotify: https://open.spotify.com/show/5QolhKm1jDZzVuHO0S9ZBo?si=910efb23833f4fc1 LinkedIn: https://www.linkedin.com/company/saturdata Instagram: @SaturdataPod #Saturdata #MachineLearning #ChatGPT #LLM

1 of 5