The AI Research Lab You're Not Watching Is Inside Salesforce

You go looking for one thing and end up somewhere completely different, which is how I ended up reading Salesforce AI Research papers at midnight on the Fourth of July.

The rabbit hole started with a tweet from Caiming Xiong — who runs Salesforce AI Research — and the immediate cognitive response was something like: wait, Salesforce has a research wing? Like, a real one? The kind that publishes foundational work and not just white papers about how AI will transform your sales pipeline?

Yes. They do. It's been there.

BLIP and BLIP-2 came out of this lab — the vision-language pretraining work that basically everyone building multimodal systems has had to reckon with. CodeT5 and CodeT5+ came out of this lab. xGen, their long-context LLM series, came out of this lab. Moirai — a universal time series forecasting transformer that treats forecasting across domains as a single pretraining problem — came out of this lab, in March. They have been, without much fanfare, shipping research that matters.

The notable thing isn't any single paper. The notable thing is that Salesforce — the company whose name you associate with a cartoon cloud and enterprise SaaS renewals and Marc Benioff posting inspirational content — has been quietly running what is, by any honest measure, a legitimate AI research operation.

This is the part that should make you recalibrate something.

The standard model for AI research prestige runs through the obvious poles: Google DeepMind, OpenAI, Anthropic, Meta AI, maybe Microsoft Research if you're being generous. The enterprise software companies are not supposed to be on this list. They're supposed to be acquiring startups and announcing AI features at their annual conferences with the energy of someone who just discovered the word "copilot."

And yet here is Caiming Xiong, who was at Yahoo Research before this, with a team that has been doing the actual work — the kind that other researchers cite, the kind that ends up in other people's models, the kind that doesn't have a product launch attached to it.

The cynical read is that this is expensive window dressing, that Salesforce funds a research lab the same way banks fund art collections — for the prestige, not because anyone inside genuinely cares. Maybe. The papers don't read that way. Moirai is not a dressed-up product announcement. It's a serious attempt to solve a hard problem in time series forecasting by treating it the way the language modeling people treated language: stop building domain-specific models and just pretrain something large on everything.

Whether that bet pays off is a different question. But the bet is real, and the team making it is real, and most people writing about AI in 2024 have no idea it exists.

That's probably fine for Salesforce. They don't need the credit. But if you're trying to actually track where interesting AI work is happening — and you're only watching the obvious places — you've already missed some of it.