OpenAI Acquires Rockset

Very unexpected acquisition. I don't think that Rockset is a suitable infrastructure for RAG, a purpose-built inverted index would be far more efficient (both in terms of compute and storage), so I'm not sure how much of the technology would actually be useful for them.

I can think of two options

- Pure acqui-hire: virtually all of Rockset engineering leadership is ex-Meta, and OpenAI has been hiring several senior infra engineers from Meta, so these are all people that have worked together previously.

- OpenAI is building some product where customers can ingest large amounts of data, which could be managed by the Rockset infrastructure as source of truth, and then indexed by their RAG systems.

OpenAI has billions of dollars and nothing but GPUs to spend it on. This isn’t strategic per-se, it’s just rollup. Good place to be in for any data-adjacent product company.

Google and Amazon followed the same strategy for over a decade just buying anything that was possibly helpful.

Giuseppe! Long time no see. Rockset’s architecture changed somewhat since we last talked— not in fundamental ways, but in ways that would alleviate your concerns.

If you want to talk (not secret) technical details, you know where to find me :)

-Tudor.

I guess I stand corrected then :)

(Hi!)

EDIT: I forgot to say, with the recent hires and the Rockset team, OpenAI is building quite the infra dream team :)

I would speculate that OpenAI is in a phase where speed of delivery is make-or-break, and any bloat would be a distraction. I bet they're extremely deliberate in their acquisitions.

Does OpenAI use Rockset internally? I feel like I have some vague memory about that… in which case, the acquisition would make sense from a continuity of business perspective.

RAG doesn't have to involve vector search.

The (very thin) blog post said "Enhancing our retrieval infrastructure" - my guess is this is more about other forms of retrieval, like constructing and executing SQL queries and using the results to help answer questions.

Last time I heard of Rockset was at the Snowflake Summit where they positioned as a faster DWH.

Looking at the landing page now it seems they almost pivoted into semi/unstructed data.

To your point, I feel like nobody knows exactly how to do RAG really well (fast and accurate). I also doubt the Rockset team has it figured out but it seems like there is an opportunity to build a new kind of database/memory system and OpenAI believes the Rockset team can help.

they were using qdrant for RAG as of November 2023. Not sure if it's changed since then.

https://x.com/simonw/status/1722011967886688696

> RAG doesn't have to involve vector search.

This. Not sure why RAG triggers vector search for everyone. Retrieval Augmented Generation is as generic as it can get.

Most likely for the same reason that so many people seem to think they need a vector-specific database and a framework like langchain to build any type of GenAI-enabled application... the content marketing is working.

I think OpenAI also realized they're an AI major without a dance partner, when it comes to context.

Google (Android, Gmail, Maps, G Office), Apple (iPhone, Mail, Maps, Productivity), Microsoft (Office365, Windows, XBox).

In terms of moat and lock-in, that leaves OpenAI vulnerable to last mile customer hijacking.