Deep Notes from Using Databricks AI Products

There are not many Databricks product write-ups on the Chinese internet. Since my company's databases and data-processing workloads are being migrated onto Databricks, I have had a fairly deep experience with Databricks' current AI product line. I have some thoughts, and I need to get them out.

Genie Space

Genie Space is a data agent that lets users interact in natural language, use SQL to retrieve data from databases, and generate simple visualizations. The concept itself looks good and is genuinely attractive, but the accuracy and stability of AI-driven data retrieval are still worrying, especially when the request is complex or spans multiple databases.

Genie Space provides customization features for instructions and patterns. But this is essentially building patterns on top of a technical black box. The opaque underlying instructions can conflict with the custom instructions you build on top, turning a setting that should strengthen the system into a source of friction.

For example, if you ask for sales from the past month, the phrase can mean many things: last calendar month, the past 30 days, or month-to-date. Even if the custom instructions define past month as last calendar month, the model may still randomly choose another definition if that conflicts with the hidden base instructions.

Of course, this kind of issue can be found and adjusted through testing. But because you never fully control the instructions, later testing and tuning become very difficult. When multiple ambiguous conditions are combined with cross-table queries, it becomes an engineering disaster.

This leads to two other problems.

First, Databricks' Git support is very limited. Much of it can only be operated through the Web UI, and Genie Space customization does not even support Git sync. Users can only use scripts to save the entire customization as JSON, sync that through Git, and then manually load and overwrite it. This is a disaster for team collaboration. CI/CD is basically nonexistent.

Second, Genie Space does not let you choose the base model or modify the interaction style. When the natural-language instruction is ambiguous, the simplest thing to do is ask the user to clarify. At the very least, it should point out the ambiguity. But because the underlying black box is constrained, this is hard to achieve.

Genie Code

Another Databricks AI product is Genie Code, an AI coding agent focused on data governance. You cannot choose the model, and you cannot see the context window. It feels like driving on a highway without a speedometer and without knowing how much horsepower the pedal gives you.

Here is one case from my own experience. Genie Code spent five minutes, through multiple rounds of edits and tests, before finally solving a very simple error: a ! pip install XXX cell had been set in a SQL environment instead of a Python environment.

To put it more seriously, Genie Code barely reaches the passing line in complex engineering environments. At one Databricks event I attended, I saw Databricks staff use Claude Code to build an MVP and run it on their own platform. To me, that detail highlighted the gap between Genie Code and mature AI coding tools.

There is no doubt that Genie Code is useful for building some MVPs, and its access to Databricks' internal knowledge base is very smooth. But given its stated target of becoming a coding agent for data governance and data analysis, its basic capabilities are still weaker than the best AI tools today, such as Claude Code and Codex. That is before we even talk about harness engineering and tool use in complex situations.

Outlook

For people who have not been exposed to AI-native workflows, Databricks' Genie Space and Genie Code are very attractive. But in genuinely deep AI-native workflows, they still feel a bit toy-like.

If Oracle plus Codex or Claude Code can fully replace Databricks' current AI products, and even work better, then where exactly is Databricks' moat in AI?

My personal view is that Databricks should heavily strengthen Genie Code until it reaches the level of GitHub Copilot, combine it with Databricks' database platform, and make it an AI tool that amplifies super individuals. That would narrow the gap with today's T0-class AI tools. It should keep Genie Code's cloud-side advantage while giving users some freedom to modify the tool itself.

At the platform level, Databricks could design a communication and coordination architecture specifically for Genie Code instances inside organizations. It could use its cloud advantage to manage and sync session history, sync project-progress information, and redesign a super-team architecture for AI-agent coordination. Reducing the friction of communication and coordination may become the real highlight of the tool.