# Databricks customers use AI more to read than to write

Markdown copy of https://realalpha.app/s/databricks-customers-use-ai-more-to-read-than-to-write/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Ali Ghodsi](https://realalpha.app/builders/ali-ghodsi/) · Ali Ghodsi is cofounder and CEO of Databricks. · Co-founded Databricks, which he has led as CEO since 2016, and helped create Apache Spark.
- Host: [Logan Bartlett](https://realalpha.app/people/logan-bartlett/) · Software Investor, Redpoint Ventures · Backs enterprise software companies like Ramp, Cribl, Cyera and Legora.
- Show: The Logan Bartlett Show
- Episode: [Databricks CEO on Super AGI and Unlocking AI’s Untapped Potential](https://realalpha.app/v/pAcF3GV4ygM/), Nov 22, 2024
- Moment: 19:55 to 22:35 of the episode (3 min), playing on https://realalpha.app/s/databricks-customers-use-ai-more-to-read-than-to-write/; the episode from there: https://realalpha.app/v/pAcF3GV4ygM/?t=1195

## Quote

> “it's probably at least a **10 to 1 ratio** between how much people are using the models to read stuff than to produce.”
> — [Ali Ghodsi](https://realalpha.app/builders/ali-ghodsi/) at [20:43](https://realalpha.app/v/pAcF3GV4ygM/?t=1243) · The Logan Bartlett Show

## Takeaways

1. Reading use can outweigh generation by at least 10 to 1.
2. Financial and insurance customers use AI to process large document sets.
3. Health records and SEC filings are other data-heavy applications.

## About this segment

Ali Ghodsi says usage on Databricks is at least 10:1 for reading versus generating. That points to extraction and analysis as major AI workloads, not just chat output.

From “Databricks CEO on Super AGI and Unlocking AI’s Untapped Potential”, Nov 22, 2024. [Watch the full episode](https://realalpha.app/v/pAcF3GV4ygM/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[20:00](https://realalpha.app/v/pAcF3GV4ygM/?t=1200) Every vertical in every industry, we see lots and lots of use cases, you know, like financial services. You know, there's so many financial advisors, FAs out there. You know, they basically now we generate automatically for one of our customers, you know, large bank automatically in advance. You know, what advice they should give to you? It's customized. They understand everything about you, your wealth situation and so on.

[20:19](https://realalpha.app/v/pAcF3GV4ygM/?t=1219) They can give you customized advice. You know, it affects, you know, thousands of thousands of these folks that are doing that. Right. In insurance, sifting through massive, massive amounts of documents to assess risk, right? It's like they can do it. And by the way, one thing that people underestimate is that the models are very fast at reading data. They're slow at producing text, right? They're much slower.

[20:44](https://realalpha.app/v/pAcF3GV4ygM/?t=1244) The other insight I would say is like, you know, that's surprising is, you know, it's probably at least a 10 to 1 ratio between how much people are using the models to read stuff than to produce. So they have, when they use Databricks, they use us 10 times or 100 times sometimes more or even more to read stuff and then produce very little text or very little data out. Whereas people are focused on that output that comes out, but it's usually much, much smaller.

[21:08](https://realalpha.app/v/pAcF3GV4ygM/?t=1268) So, you know, underwriting is another one that we're seeing. There's so many going through all the SEC filings and figuring out what's happening, you know, extracting signal. So figuring out alpha, we see it in all the financial sector. So there's so many use cases. In healthcare, electronic medical records, there's just so many of them. They're all handwritten. Being able to actually find structure in that, that's a treasure trove by itself.

[21:35](https://realalpha.app/v/pAcF3GV4ygM/?t=1295) So use cases are on that. So there's hundreds of hundreds of these you see all the time in every industry. They just haven't been productionized yet. They're not fully being used by the workforce. Do you think that read-to-write ratio, do you think that stays around that? Are there any trends that you sort of see as read going up in comparison to write? Or you think at some point the writing and the creation catches up with the reading and summarization?

[22:01](https://realalpha.app/v/pAcF3GV4ygM/?t=1321) I mean, hard to see that because, you know, how much are we going to read as humans? You know, how much could we, I mean, there's only so much and we read slowly, right? Whereas having the AI crunch a lot of information, you know, so that it has that context, makes a lot of sense. So if anything, that probably increases. And if we can do that faster. So that's where the trend is headed. We're seeing even more and more of that when we look at our data than when we started.

[22:25](https://realalpha.app/v/pAcF3GV4ygM/?t=1345) As you counsel customers, be it big enterprises or smaller ones, about figuring out