Brickster Newsletter
|
| 01 |
What should your AI agents remember?
A short guide to a16z's Databricks argument, plus three questions for your team.
|
|
| 02 |
This week in Databricks
AI connections, four dates to watch, Real-Time Mode and SAP.
|
|
| 03 |
From brickster.ai
Our iOS app for learning Spark, a Genie Ontology explainer and the official MCP migration guide.
|
|
|
|
01
|
🤖 What should your AI agents remember?
|
|
01 / THE LEAD
a16z says AI agents now query Databricks more than people do
a16z argues that reusable context and feedback can become a competitive advantage. Separately, Snowflake's CEO is asking for a clearer revenue breakdown. Here's the story in five steps.
|
|
| Shift |
→ |
Problem |
→ |
Pitch |
→ |
Catch |
→ |
Your move |
|
|
1. The shift
Your next data user may be an agent
|
|
The shift, according to a16z
🤖 Agents now generate more Databricks queries than people do
The essay does not provide a measurement window or counting method.
|
|
|
That claim comes from an Oct. 27 essay by Ben Horowitz and Michael McGuiness. In its Aug. 13 disclosure, Databricks reported an annualized revenue run rate above $7 billion, Q2 year-over-year growth above 80% and a $190 billion funding valuation.
|
|
2. The problem
Rivals can use the same AI model
|
|
So where does the advantage come from? The essay offers a thought experiment: two banks use the same model for credit decisions.
|
|
🏦 Bank 1
Prompt
↓
Answer
↓
Saves only the final decision
A new model arrives: less decision history to reuse
|
|
🏦 Bank 2
Prompt
↓
Answer
↓
Records evidence, human overrides and outcomes
↓
Tests and updates its policy
A new model arrives: the bank can reuse its accumulated context
|
|
|
Here, learning means retaining and using decision history and feedback to improve policy.
|
|
3. The pitch
Databricks wants to be Bank 2's memory
|
|
"AI does not have an intelligence problem," says Databricks CEO Ali Ghodsi. "It has a context problem." His proposed answer is a "System of Record for Agents." The essay maps that ambition to four questions and Databricks products:
|
|
What is true?
Lakehouse, Lakebase
|
What does it mean?
Genie Ontology
|
|
What is allowed?
Unity Gateway
|
Did it work?
Agent Bricks, Genie Agents
|
|
|
4. The catch
The headline metrics are company- and investor-reported
|
|
a16z led Databricks' Series A. In the reported CEO exchange, Snowflake's Sridhar Ramaswamy opened his response with four words: "Release all the numbers." Here is what the published evidence establishes:
|
| The claim |
What the evidence shows |
| Databricks' Genie accuracy claim |
Databricks reports 84.5% for Genie versus 52.4% for its strongest tested coding agent on 28 internal questions. These are different systems; the test does not isolate context's effect. |
| a16z: 17M+ agent-launched databases a day |
The essay discusses temporary branches and experiments but does not disclose how launches are counted, including how temporary branches are treated. |
| More than 80% growth |
Ramaswamy asked whether excluding pass-through inference revenue would reduce growth to 65%. That is a question, not a finding. In the exchange reported Oct. 29, Ghodsi did not provide the requested revenue breakdown. |
|
|
5. Your move
Three questions for your team this week
|
| 1 |
Are we Bank 1 or Bank 2?
Do we record evidence, overrides and outcomes? How do we use them to test and improve decisions?
|
| 2 |
Could we move this context to another platform?
Could we export our definitions, decision records and feedback in a usable format? Do our agreements allow us to reuse them?
|
| 3 |
What are we paying for?
At renewal, ask vendors to itemize platform and model-inference charges where those costs are bundled.
|
|
|
The strategic argument is useful. The headline metrics still need clearer definitions and independent validation. The model can change; the decision history should remain useful. Start with your own systems: what do they record, what can you export, and what has that evidence helped you improve?
|
|
|
|
02
|
📰 This week in Databricks
|
|
Access
Supported AI clients can use built-in OAuth integrations
Authorized users of ChatGPT, Codex, Claude, Claude Code and Cursor can connect to Databricks MCP servers through built-in OAuth apps, on by default. Unity Catalog permissions still apply, but turning the apps off later breaks existing connections. Review which integrations your organization wants enabled.
|
|
Dates
Four changes coming to Databricks
Dates in 2026; checked Nov. 2.
| Nov. 18 | New VARIANT tables created on DBR 17.3+ get shredding by default. DBR 15.4 cannot read or write those tables. |
|---|
| Nov. 30 | Genie One's Beta MCP endpoint retires. Move affected clients to Unity Gateway. |
|---|
| Dec. 2 | On supported compute, CREATE OR REPLACE VIEW preserves grants, tags and comments by default. |
|---|
| Dec. 8 | OpenSharing bearer tokens issued before Dec. 8, 2025 expire if their expiry is later than Dec. 8, 2026 or unlimited. Rotate affected credentials. |
|---|
Assign an owner to each change that affects your workspace. Check view-replacement workflows that remove or rename governed-tagged columns, and send replacement sharing credentials to affected recipients. Full list
|
|
Streaming
Real-Time Mode reaches GA in Spark Declarative Pipelines
Databricks reports that Real-Time Mode is up to 92% faster than Flink on feature-computation workloads in its own benchmark. It publishes test code, but the post names no benchmark customer or cost comparison. Test latency, cost and operational behavior on your own workload before retiring a streaming stack.
|
|
SAP
Genie and Lakebase are headed to SAP Databricks
The Oct. 30 post describes planned Genie and Lakebase additions without launch dates or availability stages. Confirm availability before including them in delivery plans.
|
|
|
|
📱 Our iOS app is out: Brickster Colony
Learn Apache Spark in short daily lessons and build a base on Mars as you go. Free and ad-free. Not yet available in the EU or mainland China.
Get it on the App Store →
|
|
|
|
|
🤖 What does Databricks document about conflicting metric definitions?
The Brickster Assistant searches our indexed archive and answers with source links.
|
|
|
Enjoying the Brickster Newsletter?
Forward this to a colleague who works with Databricks.
Forward this issue ↗
|
|
