Latest / The Augur Dispatch / Databricks's preview: prove the data exit before renewal

Transcript

Databricks's preview: prove the data exit before renewal · The Augur Dispatch

• 0:06
  SPEAKER_00: Before you cut a software bill, check whether the work can survive the cancellation.
• 0:11
  SPEAKER_00: Databricks has described a way to move control of a data table between catalogs without copying its files.
• 0:18
  SPEAKER_00: Its own implementation is still a private preview.
• 0:21
  SPEAKER_00: Augur's advice is to test a supported exit before treating it as negotiating leverage.
• 0:27
  SPEAKER_00: Keep production where it is until that test passes.
• 0:30
  SPEAKER_00: This is the Augur dispatch for Monday, October 5th, 2026.
• 0:35
  SPEAKER_00: Picture a renewal meeting.
• 0:37
  SPEAKER_00: Someone has found an agent that can replace the screens employees use, and the proposed savings look good enough to get everyone's attention.
• 0:45
  SPEAKER_00: Then, the data platform owner asks who will manage the underlying records after the old subscription ends.
• 0:51
  SPEAKER_00: The spreadsheet has no column for that conversation.
• 0:55
  SPEAKER_00: Nate Jones offers a useful way to examine the bill.
• 0:58
  SPEAKER_00: A software subscription bundles the data it stores, the screens people use, and the workflows that move work forward.
• 1:05
  SPEAKER_00: His argument is that agents can take on more of the looking up and doing, so each part must justify its price.
• 1:12
  SPEAKER_00: The direction is plausible, but replacing a screen doesn't remove the dependencies behind it.
• 1:17
  SPEAKER_00: Databricks' announcement is useful because it addresses one specific dependency, which catalog is responsible for the current state of a table.
• 1:27
  SPEAKER_00: A catalog is the record keeper that tells software where a table's latest information lives.
• 1:33
  SPEAKER_00: The data files can remain in your storage while the catalog coordinates changes to them.
• 1:38
  SPEAKER_00: Databricks describes two steps for transferring that responsibility.
• 1:42
  SPEAKER_00: Register tells a new catalog about an existing table.
• 1:46
  SPEAKER_00: Unregister tells the old catalog to let go without deleting the underlying files and returns the pointer the new catalog needs.
• 1:54
  SPEAKER_00: The order matters.
• 1:56
  SPEAKER_00: If two catalogs both believe they're in charge, changes made through one can be invisible to the other, leaving people with inconsistent results and potentially lost data.
• 2:06
  SPEAKER_00: Imagine a finance team reopening the same report through two different tools and seeing different answers.
• 2:12
  SPEAKER_00: The team may spend hours checking formulas when the disagreement is underneath the report.
• 2:18
  SPEAKER_00: Databricks says it contributed the unregister operation to the catalog standard for Apache Iceberg, which is an open format for storing the large tables businesses depend on.
• 2:28
  SPEAKER_00: Unity Catalog's implementation is available through a private preview requested from the account team.
• 2:34
  SPEAKER_00: There are important limits.
• 2:36
  SPEAKER_00: This is a route for supported tables in customer-owned storage, not a universal exit from every Databricks product.
• 2:43
  SPEAKER_00: It doesn't mean every table format becomes interchangeable.
• 2:46
  SPEAKER_00: Databricks also says production migration still requires safely stopping the programs that write data and pointing jobs at the new catalog.
• 2:55
  SPEAKER_00: Your team must check permissions and recovery too.
• 2:58
  SPEAKER_00: Moving the pointer can avoid copying the files.
• 3:01
  SPEAKER_00: It can't do the rest of the migration planning for you.
• 3:05
  SPEAKER_00: The workflow question has a different example at OpenRouter, the service that routes requests to different model suppliers.
• 3:12
  SPEAKER_00: It can run commands in an isolated computer environment, called a sandbox, on behalf of the model.
• 3:18
  SPEAKER_00: Its shell tool works through both the responses and messages request formats.
• 3:23
  SPEAKER_00: Its bash tool works through messages only.
• 3:26
  SPEAKER_00: Neither works through chat completions, and OpenRouter hosted execution uses the global service endpoint rather than its regional endpoints.
• 3:34
  SPEAKER_00: Both tools remain in beta, so their interfaces may change.
• 3:38
  SPEAKER_00: There's also a default worth checking.
• 3:40
  SPEAKER_00: Bash normally returns the requested command to your application, so you must explicitly choose OpenRouter's execution engine to have OpenRouter run it.
• 3:49
  SPEAKER_00: For a business leader, the distinction is where the work happens and who owns the security boundary.
• 3:55
  SPEAKER_00: A tool that sounds hosted isn't necessarily running there with your current settings.
• 4:00
  SPEAKER_00: This can help an agent application use different model suppliers.
• 4:04
  SPEAKER_00: It doesn't transfer your business software's approval rules or access controls, and different models may still complete the same task differently.
• 4:12
  SPEAKER_00: Leaving beta would reduce one source of uncertainty.
• 4:16
  SPEAKER_00: It wouldn't certify your workflow.
• 4:18
  SPEAKER_00: For the renewal decision, start with one important subscription and separate its data, screens, and workflows on paper.
• 4:26
  SPEAKER_00: Ask the vendor what a supported exit would require, including how the old system stops accepting changes after the handoff.
• 4:34
  SPEAKER_00: A documented answer gives your engineers something to test.
• 4:37
  SPEAKER_00: It isn't a substitute for the test.
• 4:40
  SPEAKER_00: And a no-copy transfer isn't the only acceptable exit.
• 4:44
  SPEAKER_00: Other products may have a workable export and migration process.
• 4:48
  SPEAKER_00: If you have a supported icebreaker setup and preview access, give the data platform owner one non-production table.
• 4:55
  SPEAKER_00: Compare the handoff with today's query results, permissions, job behavior, and recovery time.
• 5:00
  SPEAKER_00: Then review the evidence with procurement before the renewal decision.
• 5:04
  SPEAKER_00: The new catalog should manage rights, the old path should no longer keep writing, and the agreed data and access checks should pass.
• 5:12
  SPEAKER_00: Stop if results disagree, access changes unexpectedly, or recovery hasn't been proved.
• 5:19
  SPEAKER_00: Expand only after the procedure is repeatable and the migration plan has been reviewed.
• 5:24
  SPEAKER_00: Until then, keep the savings in the proposed column.
• 5:27
  SPEAKER_00: Procurement has enough imaginary discounts to manage already.
• 5:31
  SPEAKER_00: The advantage would come from knowing your dependencies and reducing switching effort.
• 5:35
  SPEAKER_00: Access to a preview that competitors can also request isn't a durable edge by itself.
• 5:41
  SPEAKER_00: For people running local models on Mac, a smaller change deserves a practical check.
• 5:46
  SPEAKER_00: Olama's 0.40 release candidate notes include a fix to make its text processing match the model publisher's tokenizer.
• 5:55
  SPEAKER_00: A tokenizer cuts text into the pieces a model reads.
• 5:58
  SPEAKER_00: If those rules differ from the publisher's, the model isn't receiving exactly the input its developers intended.
• 6:05
  SPEAKER_00: The checked third release candidate also makes the MLX runtime the default for supported model architectures on Apple chips.
• 6:13
  SPEAKER_00: It is marked as a pre-release, and the notes don't measure how the changes affect answers.
• 6:18
  SPEAKER_00: Picture a developer who depends on a local model to classify incoming documents.
• 6:22
  SPEAKER_00: Before adopting the update, rerun the same saved examples and inspect the mistakes, not just whether the application still opens.
• 6:30
  SPEAKER_00: Repeat that check when the stable release reaches your deployment.
• 6:34
  SPEAKER_00: The model serving item also needs a correction to the original brief.
• 6:38
  SPEAKER_00: GLM 5.3 Flash already has a first-party launch post and public model weights, meaning downloadable model files.
• 6:46
  SPEAKER_00: Z.ai has also published benchmark results.
• 6:49
  SPEAKER_00: Those are publisher results, not proof of better performance on your work.
• 6:54
  SPEAKER_00: SGLang version 0.5.21 adds specific support and optimization work, including for AMD hardware, rather than establishing the model's first serving path.
• 7:05
  SPEAKER_00: Check the exact checkpoint, license, hardware requirements, and workload results before switching.
• 7:11
  SPEAKER_00: This correction does not regrade Augur's earlier GLM forecast.
• 7:16
  SPEAKER_00: Anthropic's stock market plans require another correction.
• 7:20
  SPEAKER_00: The company announced on June 1st that it had confidentially submitted a draft registration statement for a proposed public offering.
• 7:27
  SPEAKER_00: The filing step is confirmed.
• 7:30
  SPEAKER_00: A first trading day is not confirmed by that announcement, which says the offering depends on market conditions and other factors.
• 7:37
  SPEAKER_00: Confidential submission also doesn't make the full prospectus public.
• 7:41
  SPEAKER_00: October and pre-Thanksgiving dates in commentary are expectations, not company commitments.
• 7:47
  SPEAKER_00: They aren't enough by themselves to establish an official delay.
• 7:51
  SPEAKER_00: For someone negotiating a Clawed contract, the useful questions concern service terms and operational dependence.
• 7:58
  SPEAKER_00: Keep initial commitments short where practical, and review public financial disclosures if they become available before a multi-year renewal.
• 8:06
  SPEAKER_00: Today's Augur call is unchanged.
• 8:08
  SPEAKER_00: Anthropics shares begin trading on a public stock exchange on or before November 25th, 2026.
• 8:15
  SPEAKER_00: The odds are 35%.
• 8:17
  SPEAKER_00: A hit requires an exchange notice, company announcement, or regulatory filing showing a first trading day by that date.
• 8:25
  SPEAKER_00: It's a miss if no anthropic shares have traded publicly by then.
• 8:29
  SPEAKER_00: A confidential filing alone doesn't satisfy it.
• 8:32
  SPEAKER_00: One policy item remains on watch.
• 8:35
  SPEAKER_00: TechCrunch reports that the new superintelligence force has 120 days to report on AI risks and opportunities.
• 8:43
  SPEAKER_00: Its coverage also quotes charter language about responding to threats and avoiding over-regulation.
• 8:49
  SPEAKER_00: That is more detailed than the original brief acknowledged.
• 8:52
  SPEAKER_00: It still doesn't establish a change in procurement rules or enforcement powers.
• 8:57
  SPEAKER_00: For today's decision, adopt the contract dependency review.
• 9:01
  SPEAKER_00: Pilot the catalog handoff only where supported and isolated from production.
• 9:05
  SPEAKER_00: Watch the unresolved release and listing milestones without treating them as guarantees.
• 9:11
  SPEAKER_00: Augur calls are dated predictions with explicit hit and miss criteria, so readers can check the judgment later.
• 9:17
  SPEAKER_00: The website tracks them.
• 9:19
  SPEAKER_00: The written brief with every source is at auger dispatch.ai.
• 9:24
  SPEAKER_00: Portions of the Augur Dispatch are AI generated.
• 9:27
  SPEAKER_00: That is the Augur Dispatch for Monday, October 5th, 2026.
• 9:32
  SPEAKER_00: Thank you for listening.