EXPLAIN ANALYZE

How do you give an AI agent freedom while knowing it did the right work? Seth Rosen joins Jacob Matson and Mehdi Ouazza to explore checkpoints, structured evidence, and the data modeling behind reliable agents.

From GitHub’s internal analytics agent to changing data stacks, documentation, and AI-native organizations, they discuss where humans fit—and why traces, evals, and collaboration matter as much as the tools.

00:00 Meet Seth Rosen and ThruWire

01:19 The story behind the data meme

03:51 A room full of agents and the software factory

08:57 GitHub’s internal analytics agent

11:30 Checkpoints, receipts, and agent freedom

15:39 The disappearing boundaries of the data stack

23:33 Writing documentation for humans and agents

29:31 What makes a company AI-native?

36:31 Collaboration and the “slop cannon”

40:17 Agent traces, evals, and data teams

44:54 Structured receipts that steer agent work

50:53 Why this is a data modeling problem

52:07 Where to find Seth and closing thoughts

Show notes : https://motherduck.com/podcast

AWS bought DuckLabs—the company employing the core team behind DuckDB while DuckDB remains open source. What does the deal mean for AWS, MotherDuck, and the data stack building around DuckDB?

In this episode, @mehdio and @dumkydewilde unpack the acquisition, then connect Iceberg and Lance, SQL-first context for agents, Spotify's LLM evaluation funnel, AI-generated data lineage, CodexBar, WalGit, query optimizers, and the fight over whether SaaS products should ship their own agents or let users bring one.

All links on https://motherduck.com/podcast

00:00 Welcome to EXPLAIN ANALYZE

01:15 Why AWS bought DuckLabs

11:14 Analyze with Iceberg, serve with Lance

16:30 Why your agent should manage context through SQL

22:09 Better experiments with LLM evals

27:26 CodexBar makes AI usage visible

33:06 Architectural lessons for AI-powered data warehouses

38:10 WalGit: one Git server binary, backed by object storage

42:55 Query optimizers: the low-key MVP

47:13 Do users want your agent, or their own?

58:36 Wrap-up

#DuckDB #AWS #DuckLabs #DataEngineering #AIAgents #Iceberg #Lance #LLMEvals #MCP #DataPodcast

Agent code can look productive right up until a dependency changes, an eval misses the real failure mode, or an over-permissioned tool turns a routine task into a security incident. So what does it actually take to operate AI systems responsibly?

In this episode, Mehdi, Dumky de Wilde, and Maria Vechtomova connect MLOps and LLMOps to agent evals, MCP governance, regenerated software, security, and the engineering practices that still matter when outputs are non-deterministic.

All links and note : https://motherduck.com/podcast

Chapters:

00:00 Meet Maria Vechtomova

01:01 From MLOps to forward-deployed engineering

01:59 Principles first, Databricks second

06:18 What changes from MLOps to LLMOps

09:10 Deterministic tools for non-deterministic systems

09:50 Who maintains regenerated software?

12:38 Hiring for critical thinking with AI

17:38 MCP skills, files, extensions, and stateless servers

20:43 The missing governance layer for agent tools

24:15 Turning deployment pain into reusable practices

28:44 Testing and evaluating LLM systems

31:59 Why AGENTS.md beat skills in Vercel’s evals

40:13 When an agent accidentally hacks Hugging Face

45:00 Skills and the software supply chain

47:53 Consulting that leaves teams stronger

50:42 Wrap-up

​ @mehdio  and Dumky work through a round of data and AI news, starting with Analytics Tycoon, a browser game where you run a data team on a fixed budget and lose trust points every time a cron job breaks. Then the uncomfortable stuff: Python 3.15 is at release candidate while 3.10 is still the number two download and goes end of life in two months, your agent keeps writing 3.11 because it does not know what day it is, Stripe is reportedly circling OpenRouter for 10 billion dollars, and data modeling turns out to be very much alive once you try to replace it with a context layer.

Subscribe + all links: motherduck.com/podcast

Hosts: Mehdi Ouazza (@mehd-io) and Dumky de Wilde (@dumkydewilde)

CHAPTERS
00:00 Cold open
00:33 Analytics Tycoon: the data team as a strategy game
06:46 Python 3.15 is coming and 3.10 is about to die
14:07 If nobody blogs about it, will your model know?
15:20 Agent Craft and Agents View: maps for your agents
19:18 Stripe in talks to acquire OpenRouter
23:20 Google's next chapter and the shrinking AI moat
30:19 Is data modeling dead? Daniel Beach weighs in
35:10 Hard to eval is a product smell
39:57 Cloudflare OS and the return of internal tools
43:02 Supabase publishes public evals for AI agents
44:21 Superlogical: Hashimoto builds a multiplexer
48:47 That is a wrap

#ExplainAnalyze #MotherDuck #DataEngineering #DuckDB #Python #AIAgents #LLMEvals #DataModeling #OpenRouter #CloudflareOS #Tmux #DataPodcast

Is the safest bet in AI the hardware underneath it?

Show notes: https://motherduck.com/explain-analyze

Mehdi Ouazza and Jacob Matson unpack Meta's reported $10B compute lease talks with Anthropic, why AI labs are becoming infrastructure customers, and what it means for data and AI builders. Then they get into graph-shaped context for agents, removing AI slop from writing, reproducible development environments, AMD's Anthropic bet, Domo's acquisition and the future of standalone BI, lakehouse vs. warehouse, leadership bottlenecks, tldraw's isRecord saga, and loop engineering.

EXPLAIN ANALYZE is the MotherDuck podcast where the hosts bring data and AI stories the others have not seen, then work through them together.

00:00 Welcome to EXPLAIN ANALYZE
01:02 Meta's potential $10B compute lease with Anthropic
04:13 Graph context for agents
10:23 Removing AI slop from writing
18:33 Why “it worked on my machine” still happens
25:45 AMD commits up to $5B to Anthropic
27:26 Domo acquired: the end of standalone BI?
31:29 Where data engineering is heading
33:15 Lakehouse vs. warehouse
37:30 Leadership becomes the bottleneck
41:10 tldraw and the isRecord saga
46:07 Loop engineering: goals, stop criteria, and timers
50:03 Thanks for watching

AI is moving inside the database, and it keeps hitting problems databases already solved.

Stephanie Wang (staff engineer at MongoDB, founding engineer at MotherDuck) joins Mehdi and Jacob Matson to dig into Snowflake's Cortex AISQL and why AI inside SQL is really a query-optimization and cost problem, the compute crunch pushing RAM and inference prices up, RadAgents and Sakana's Fugu model router, open weights vs open development, and the "elephant and goldfish" method for designing before you let AI write the code.

All links and full show notes: motherduck.com/explain-analyze

00:00 Intro and guest intro (Stephanie Wang)

01:35 Cortex AISQL: AI operations native to SQL

10:14 SpaceX rents out idle compute for AI inference

12:11 Compute is the new database engine for AI

13:43 DRAM price hikes and the margin squeeze

18:40 Local AI models and sandboxing agents

22:39 RadAgents: multi-agent reasoning for radiology

27:34 Sakana Fugu and dynamic model routing

30:12 Open weights vs open development

33:43 The elephant and goldfish: design before you code

42:07 MPP: machine-to-machine payments for agents

46:56 Wrap-up

Yan Holtz, the data viz educator behind data-to-viz.com and the R/Python/D3 graph galleries, joins Mehdi and Alex for a full episode on data visualization in the age of AI.

AI has basically solved the "how do I write the code" problem, which makes the "what makes a good chart" problem matter more than ever - and it's quietly pushing all serious data viz toward JavaScript and D3, even for R and Python people. We get into picking the right chart before writing a line of code, whether dashboards are dying in favor of agents, why D3.js went from the hardest tool to the default, an agent that grades its own charts against the Tufte test, why Shiny and Streamlit may no longer be the right call, the token cost of data work, and whether a chat box is even the right tool to design a visualization.

Subscribe + all links: motherduck.com/explain-analyze

Yan's website : https://www.data-to-viz.com/
Yan's social : https://www.linkedin.com/in/yan-holtz-2477534a/

Chapters
00:00 Welcome + guest intro (Yan Holtz)
01:29 data-to-viz.com: choosing the right chart
07:58 BI is dead, long live BI: dashboards vs agents
17:09 ninejs: interactive ggplot2 charts in Python
20:01 Teaching an AI agent to make beautiful charts
29:20 D3.js: infinite flexibility, now AI-writable
36:24 Shiny: the right tool in the age of AI?
42:02 Token economics: data is token-hungry
46:33 Data visualization in the age of AI
49:37 Outro

#DataViz #DataVisualization #D3js #JavaScript #BI #DuckDB #AI #MotherDuck #ExplainAnalyze

Infrastructure keeps winning the funding while SaaS bleeds — but is it sustainable when AI agents spin up databases that die in days? Mehdi and Jacob Matson dig into Supabase's Series F, local AI on personal chips, DuckDB turning up everywhere (MariaDB, MondayDB HTAP), Anthropic's "data is not software" claim, and a wave of people building software just for themselves.

Show notes & links : https://motherduck.com/explain-analyze

➡️ Follow Us
LinkedIn: https://linkedin.com/company/motherduck
X/Twitter : https://twitter.com/motherduck
Blog: https://motherduck.com/blog/

00:00 — Intro: who's on today
00:48 — Supabase's Series F: SaaS vs infrastructure
05:14 — Local AI: NVIDIA's chips for personal computers
09:28 — MondayDB 3: HTAP for a trillion-table system
14:10 — A DuckDB storage engine for MariaDB
18:14 — Anthropic: "data is not software"
24:12 — Delba on feedback loops and less agent babysitting
29:20 — Tristan Handy: I built a (very small) agent swarm
32:54 — Home-cooked apps: software just for you
38:01 — DoomBench: can your database run Doom?
40:41 — A C documentary: languages in the AI era
44:28 — Wrap-up

Daniel Beach (Data Engineering Central) joins Mehdi and Dumky for a tour through the open lakehouse renaissance, AI-native data agents, and the slow grind of getting humans out of the orchestration loop.
We cover the catalog commits feature in Delta + Unity Catalog + DuckDB (multi-writer lakehouse without Spark), Cloudflare's new unified data platform and their AI agent Skipper, ADBC quietly replacing JDBC/ODBC, Aaron Francis launching SoloTerm, Addy Osmani on the orchestration tax, Mitchell Hashimoto getting a 40x speedup with an agent in a loop.

Subscribe + all links: motherduck.com/explain-analyze

👋 Hosts
Mehdi Ouazza —  @mehdio 
Dumky de Wilde
Guest: Daniel Beach (Data Engineering Central) —  @dataengineeringdan 

Check out Daniel Beach's work on https://dataengineeringcentral.substack.com/

➡️ Follow Us
LinkedIn: https://linkedin.com/company/motherduck
X/Twitter : https://twitter.com/motherduck
Blog: https://motherduck.com/blog/

00:00 Cold open · who's on today
01:43 Catalog commits · Delta + Unity Catalog + DuckDB
10:46 Cloudflare unified data platform · Skipper agent
17:53 Education in the age of AI · language as interface
24:39 ADBC · the next-gen database connection
30:17 SoloTerm · the terminal in the browser
38:56 The orchestration tax · Addy Osmani
40:38 Mitchell Hashimoto · an agent in a loop
45:50 Databricks Zerobus · streaming into the lakehouse
47:06 Outro

#DuckDB #DataEngineering #Lakehouse #DeltaLake #UnityCatalog #Cloudflare #AIagents #ApacheArrow #ADBC #MotherDuck

Mehdi (DevRel) and Dumky (Developer Experience) at MotherDuck run through ~10 links from the world of data and AI. This one keeps circling back to a single theme: as AI makes everything easier to produce, the hard part becomes telling the real signal from the slop, whether that's music, GitHub stars, or your data warehouse answers.

All links: motherduck.com/explain-analyze : subscribe there to get every new episode by email with full notes!

• 0:00 — Intro
• 0:43 — AI slop on Spotify?
• 6:00 — A skill file with 161k GitHub stars
• 11:30 — Up and Down the Ladder of Abstraction
• 17:00 — GitHub Actions are the new email scam
• 24:00 — The chat box is dying
• 30:00 — Slow tech: Kindle, TRMNL, Fitbit
• 37:00 — From Vibes to Evals
• 43:00 — The AI paradox (via Lenny's Podcast)
• 51:40 — Wrap up