The best database tuning setting sometimes comes from an obscure DBA Stack Exchange thread, not the manual. @ImmanuelTrummer had LLMs read both, shrink the search space, then let reinforcement learning finish the job. link below

Loss of control doesn’t have to look like a machine that refuses an order.

Often it’s weaker: agents with permissions, no identity, thin audit, slow revoke.

Capability ≠ reliability.

israelhayom.co.il/tech/tech-news… [hebrew]

OCSF is not just about putting security events into the same shape. The harder work is standardizing semantics and values so defenders spend less time cleaning data. link to episode below

Pulling a big result set from Snowflake into Tableau isn't network-bound. It's CPU-bound — burning cycles converting columns to rows and straight back to columns. @ianmcook of Columnar on ADBC, and the 10-100x you get by deleting that step. link below

Traditional databases were "sold on the golf course" — to purchasing departments, not developers. Hannes Mühleisen on how that blind spot created the gap between pandas and Spark that DuckDB walked right into.
Episode link below

Effective messaging seen on a trip.
This sign says -
Do not pee.
We will cut off your testicles.
And make fun with pictures online.
😂

Ep 28: Building the Open Lakehouse for the AI Era with @cto_datazip from @_olake

Hudi vs. Iceberg vs. Delta, sub-10-min CDC for fintech, Arrow-based ingestion, and what Iceberg decoupling from Parquet means for AI.

🎧 techontherocks.show/28

"What's your moat?" is the wrong question.

In AI, there are no permanent moats — only time-bound advantages and what you build on top of them.

The right question: how long does your moat last, and what does it buy you time to do next?

link below

From Session Replays to Autonomous Improvement

We talk to @getmilana 's @rohankatyal29 & @raghavsethi about shipping the first AI Product Engineer. An agent that turns session replays into shippable code.
Dashboards are dying. Agents own the OKRs now.
techontherocks.show/26

"Never bait and switch developers." @jamwt watched CRDTs promise magic on day one and deliver misery by month six. Convex starts with serializable transactions and typed, reactive code instead - and it turns out that's what AI agents write best too. episode link below

Hot take: Feature Stores failed because they were essentially Shadow IT. 📉

@hussainsultan from xorq joins to explain why we need "lock files" for data pipelines—making ML workflows reproducible across engines like Snowflake and DuckDB without the glue code.

link below

LLM apps are moving fast, and the risks are moving faster.

That’s why we’ve developed a guide for securing AI Applications.

In “Building Secure AI Applications,” we break down how the OWASP LLM Top 10 shows up in real systems and map each risk to controls teams can actually u’re building or securing LLM features, we include a full vendor-neutral reference architecture.

Download the Guide → dryrun.security/resources/owas…

“A terabyte is not big data anymore.”

New episode with the creator of pandas & Apache Arrow @wesmckinn on: Arrow vs Parquet, next-gen file formats, why DuckDB/DataFusion often beat “big data”, and how AI coding agents are changing open source infra.

link below